diff --git a/README.md b/README.md
index 99ea0df..086ef29 100644
--- a/README.md
+++ b/README.md
@@ -108,6 +108,7 @@ external data or GPU. It checks integration, not representative compression.
| --- | --- |
| [Actual FastText model](docs/fasttext_native_evidence_20261031/README.md) | 90,380 rows and word/sentence outputs byte-identical; whole bundle saves 0.46%, hot queries slower than dense; structural order has no added benefit here |
| [Structural contribution controls](docs/structural_native_evidence_20261030/README.md) | Trained SGNS/GRU references measure DATrie and approximate joint sharing separately; exact mode finds no identical rows to merge |
+| [FP16/INT8 + structural controls](docs/quantized_structural_evidence_20261031/README.md) | Same trained tables retain additional approximate savings after quantization; 4-order GRU INT8 regresses; quality failures and fixed-default audit are separate |
| [1.85M trained-row capacity](docs/cpu_capacity_evidence_20261030/README.md) / [sparse blocks](docs/sparse_blocks_evidence_20261030/README.md) | Smaller blocks lower native miss latency 16–27× relative to the previous native layout; raw mmap remains faster |
| [TinyEngram/Qwen trained integration](docs/tinyengram_qwen_evidence_20261030/README.md) / [CUDA controls](docs/qwen_cuda_performance_20261030/README.md) | Preserved trained outputs and measured CPU/device costs; native lookup remains slower than dense offload on these traces |
diff --git a/docs/index.md b/docs/index.md
index 67b7bf0..2a475ce 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -15,6 +15,7 @@
- [Actual legacy FastText model and matched layout controls](fasttext_native_evidence_20261031/README.md)
- [Original DATrie / mark-compact contribution ablations](structural_native_evidence_20261030/README.md)
+- [Same-trained-table FP16/INT8 and structural-sharing controls](quantized_structural_evidence_20261031/README.md)
- [1.85M trained-row CPU capacity](cpu_capacity_evidence_20261030/README.md)
- [Sparse-block native miss latency](sparse_blocks_evidence_20261030/README.md)
- [Native query/cache controls](native_query_evidence_20261030/README.md)
diff --git a/docs/ngram_memory_research_plan.md b/docs/ngram_memory_research_plan.md
index 5e97692..68a2a8b 100644
--- a/docs/ngram_memory_research_plan.md
+++ b/docs/ngram_memory_research_plan.md
@@ -125,6 +125,7 @@ work.
| Original-ID/byte-preserving storage and adapters | [Actual FastText outputs](fasttext_native_evidence_20261031/README.md), [trained TinyEngram/Qwen](tinyengram_qwen_evidence_20261030/README.md), Linux/macOS wheel CI | Frozen handles; reduction contracts differ; official Engram verification uses a demo, not trained weights |
| Sparse native queries | [1.85M trained rows](cpu_capacity_evidence_20261030/README.md) and [smaller blocks](sparse_blocks_evidence_20261030/README.md) | Misses improved 16–27× over the prior layout; raw mmap remains faster |
| DATrie / mark-compact | [Trained SGNS/GRU ablations](structural_native_evidence_20261030/README.md) | Exact key indexing helps these traces; no identical structural rows found; extra joint row savings are approximate |
+| Quantization + structural sharing | [18 trained tables, 486 controls](quantized_structural_evidence_20261031/README.md) | Additional savings remain after FP16/INT8 on these references; low-order INT8 can regress; validation-selected test failures are retained |
| End-to-end artifact cost | [FastText matched controls](fasttext_native_evidence_20261031/README.md) | Whole bundle saves only 0.46%; queries are slower; structural order adds no benefit on this model |
| Construction/runtime resources | [Implementation boundaries](native_ngram_memory.md) | Structural planning copies full rows/maps offline; cache budgets do not bound RSS; FastText retains an unused 200 MB hash allocation |
diff --git a/docs/quantized_structural_evidence_20261031/README.md b/docs/quantized_structural_evidence_20261031/README.md
new file mode 100644
index 0000000..a2b5d05
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/README.md
@@ -0,0 +1,118 @@
+# Quantization and original structural sharing on the same trained tables
+
+The original DATrie prefix/suffix planner and mark-compact retain an additional
+storage contribution after FP16 and per-row INT8 on these trained references.
+It is workload-dependent: INT8 compaction makes the 4-order GRU backend larger.
+Quantization and joint sharing are **approximate to original FP32 weights**;
+native packing preserves their encoded representative bytes and original IDs.
+
+## Matched storage results
+
+Three-seed **means**, fixed cosine and norm ratio 0.99 on decoded encoded rows.
+The denominator is the smallest same-encoding control among raw mmap, native
+identity, native lexical layout and exact encoded-byte dedup. Actual totals
+include the row table, forwarding map, codec metadata, INT8 scales and a common
+keys JSON. Positive values are extra space saved; negative values are regressions.
+
+| Reference / max order | Additional FP16 saving | Additional INT8 saving |
+| --- | ---: | ---: |
+| SGNS / 4 | +22.94% | +20.03% |
+| SGNS / 8 | +33.22% | +28.30% |
+| SGNS / 12 | +33.11% | +28.08% |
+| GRU / 4 | +5.26% | -4.56% |
+| GRU / 8 | +21.13% | +10.95% |
+| GRU / 12 | +25.42% | +15.10% |
+
+SGNS reports vector distortion, without a downstream task claim. The causal
+GRU has held-out character prediction metrics. The DATrie alternative replaces
+common keys JSON with an actual serialized exact-key index; its matched totals
+are also in [report.json](report.json) and [summary.json](summary.json). For INT8,
+those extra savings average 20.78/29.83/29.68% on SGNS and −4.90/12.14/16.90% on
+GRU at orders 4/8/12. Index lookup speed is outside this study.
+
+
+
+## Quality and selection boundaries
+
+The primary study selects the smallest backend within all validation budgets,
+separately by encoding: relative perplexity increase <= 0.5%, mean prediction
+KL <= 0.0001 nats, and top-1 changes <= 1% of positions. All choices are frozen
+before testing selected candidates; fixed raw FP16/INT8 test controls remain
+separate. Paired character-block intervals are reported without reselection.
+
+**20/27 primary validation choices pass all test budgets; seven fail mean KL.**
+The selected INT8 paths pass 8/9: all 4-order runs select raw INT8; 8-order seed 3
+fails KL; all three 12-order selections pass. The latter save an additional
+15.17–15.25% backend bytes over the best INT8 control, or 12.38–12.44% of the
+measured whole serving bundle. Their test perplexity increases 0.0051–0.0208%.
+These selected looser thresholds have maximum per-row relative L2 error of
+31–44%; small task changes do not establish small weight errors.
+
+The fixed default 0.99 bundles receive a **separate post-main-run test audit**
+in [fixed099-test-audit.json](fixed099-test-audit.json): **27/27 pass** the same
+budgets, with test perplexity changes -0.0050% to +0.0172% and maximum mean KL
+1.62e-05 nats. It does not replace any
+primary choices or failures and is not preregistered confirmatory evidence.
+Fixed thresholds were present in the original storage/validation sweep. Future
+quality confirmation needs independent data/checkpoints.
+
+The fixed 0.99 sweep reaches **14.31% maximum per-row relative L2 error** against
+original FP32, including INT8 rounding. “99% cosine” is not “1% vector error.”
+The query-weighted Frobenius curve below includes quantization's own error;
+additional error versus quantization-only is recorded separately. Relaxing the
+cosine guard on GRU adds little capacity benefit while increasing error tails.
+
+
+
+Warm CPU gather p50 is **1.04–1.21× raw encoded mmap** for the fixed joint paths,
+including Python/NumPy dequantization. This study supports extra storage savings,
+with a query cost; it does not show a numeric gather speedup. First query means
+empty application cache on warmed filesystem pages, not disk-cold I/O.
+
+## Reproduce and inspect
+
+The primary run used clean source commit
+[`3a54de2`](https://github.com/initial-d/ngram-store/commit/3a54de2d8a7b97ee1432980c41b4edb0f64566d3),
+Python 3.11.15, PyTorch 2.11.0+cu130, NumPy 2.2.6, Linux x86-64 and an AMD EPYC
+7742 shared host. One PyTorch/interop/BLAS worker at nice 19, **CPU only**;
+wall time 8m48s. No weights were trained, external text downloaded or GPU
+programs touched. The corpus/checkpoint/row/source/library hashes and exact
+command are in the raw report. Build flags, tests and timing are retained in
+[build.txt](build.txt), [tests.txt](tests.txt), [run.sh](run.sh), [run.log](run.log)
+and [environment-extra.json](environment-extra.json). An initial launcher failed
+because `/usr/bin/time` was unavailable; [its error](resources-launch-failure.txt)
+is retained. Bash timing was used for the successful run.
+
+This is **18 cached public-text trained models and 486 storage variants**:
+SGNS/GRU orders 4/8/12, seeds 3/7/11; FP32/FP16/INT8, four encoded controls and
+five joint cosine thresholds (0.99/0.98/0.97/0.95/0.90), norm ratio fixed at 0.99.
+The production Linux DATrie planner is used; the portable reference planner is a
+separate smoke-test path. Every encoded row and original-ID forwarding result
+was checked. Planning uses rounded/dequantized rows before physical compaction.
+INT8 is symmetric nearest-even [-127,127], with its float32 scale inside the row.
+
+With the original cached inputs available, build and run the command from
+`run.sh` inside the pinned checkout. Summarize from the current repository:
+
+```bash
+python -m tools.analyze_quantized_structural \
+ --report docs/quantized_structural_evidence_20261031/report.json \
+ --output-dir artifacts/quantized-analysis --plots
+```
+
+A later finite-range INT8 guard prevents decode overflow near max float32.
+[codec-revision-check.json](codec-revision-check.json) verifies identical encoded
+and decoded bytes for all **54 trained encoding inputs**, using
+[verify_codec_revision.py](verify_codec_revision.py). Primary timings remain
+at the pinned study source. [audit_fixed099.py](audit_fixed099.py) reproduces the
+separate default-threshold audit in the pinned remote checkout.
+
+The measured whole GRU bundle sums backend/codec, common keys JSON and an actual
+serialized unchanged backbone/registry; the separate keys JSON duplicates key
+information held inside that registry, so it is a conservative evaluation
+bundle, not a minimal facade. Runtime/framework, training checkpoint, logs and
+builder/RSS costs are excluded. CPU exclusivity and frequency were not pinned.
+These small English SGNS/GRU references are not an official trained DeepSeek
+Engram checkpoint, a paper-scale reproduction, all-language evidence or a
+comparison with every learned/vector quantizer. Raw text, keys, weights and
+serving bundles remain local.
diff --git a/docs/quantized_structural_evidence_20261031/audit_fixed099.py b/docs/quantized_structural_evidence_20261031/audit_fixed099.py
new file mode 100644
index 0000000..3bedacb
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/audit_fixed099.py
@@ -0,0 +1,82 @@
+"""Post-main-run fixed-default quality audit, separate from model selection.
+
+Uses the already declared 0.99 bundles; does not replace failed validation
+choices. This additional test audit is not preregistered confirmatory evidence.
+Run inside the pinned study checkout with the original locally cached inputs.
+"""
+from argparse import ArgumentParser, Namespace
+import hashlib
+import json
+from pathlib import Path
+import platform
+import shlex
+import subprocess
+import sys
+
+import torch
+
+from tools.benchmark_quantized_structural import score_bundle, file_sha
+from tools.benchmark_causal_lm_compression import text_splits, paired_ppl_interval
+from tools.causal_ngram_lm import CausalNgramLM, encode_segment, evaluate, metric_summary, prediction_delta
+
+
+def main():
+ p = ArgumentParser(description=__doc__)
+ p.add_argument("--study-root", type=Path, required=True)
+ p.add_argument("--output", type=Path, required=True)
+ args = p.parse_args()
+ torch.set_num_threads(1)
+ torch.set_num_interop_threads(1)
+ root = args.study_root
+ report_path = root / "results/report.json"
+ report = json.loads(report_path.read_text())
+ commit = subprocess.check_output(["git", "rev-parse", "HEAD"], text=True).strip()
+ if commit != report["commit"] or report["failures"] or len(report["runs"]) != 18:
+ raise ValueError("complete pinned primary study required")
+ corpus = Path("artifacts/native-ngram-20261018/corpus/tinyshakespeare.txt")
+ if file_sha(corpus) != report["corpus_sha256"]:
+ raise ValueError("corpus hash mismatch")
+ training = json.loads(Path("artifacts/causal-lm-compression-20261024/report.json").read_text())
+ text = text_splits(corpus.read_text(), training["settings"]["max_characters"])["test"]
+ cfg = Namespace(**report["settings"], native_library=Path("build/libmeft_rows.so"))
+ checks = []
+ for run in report["runs"]:
+ if run["category"] != "gru":
+ continue
+ checkpoint = Path("artifacts/causal-lm-compression-20261024") / f"order-{run['max_order']}-seed-{run['seed']}/reference.pt"
+ if file_sha(checkpoint) != run["checkpoint_sha256"]:
+ raise ValueError("checkpoint hash mismatch")
+ state = torch.load(checkpoint, map_location="cpu", weights_only=True)
+ model = CausalNgramLM(len(state["registry"]["characters"]) + 1, len(state["registry"]["grams"]), state["dim"])
+ model.load_state_dict(state["model_state"])
+ model.eval()
+ original = model.memory
+ segment = encode_segment(text, state["registry"])
+ baseline = evaluate(model, segment, cfg.burn_in, cfg.eval_chunk)
+ variants = {v["name"]: v for v in run["variants"]}
+ for encoding in ("fp32", "fp16", "int8"):
+ name = encoding + "-joint-0.99"
+ directory = root / "results" / f"gru-order-{run['max_order']}-seed-{run['seed']}" / name
+ if file_sha(directory / "table.men") != variants[name]["metadata"]["table_sha256"]:
+ raise ValueError("frozen bundle hash mismatch")
+ scores = score_bundle(model, original, directory, segment, cfg)
+ delta = prediction_delta(scores, baseline)
+ checks.append({"max_order": run["max_order"], "seed": run["seed"], "encoding": encoding,
+ "name": name, "checkpoint_sha256": run["checkpoint_sha256"],
+ "test": metric_summary(scores), "test_delta": delta,
+ "test_paired_block_interval": paired_ppl_interval(scores, baseline, seed=cfg.eval_seed),
+ "test_all_budgets_passed": delta["perplexity_relative_change"] <= cfg.max_ppl_increase
+ and delta["mean_prediction_kl_nats"] <= cfg.max_kl and delta["top1_changed_fraction"] <= cfg.max_top1_flip})
+ result = {"category": "post-main-run fixed-default test audit; not preregistered confirmatory evidence",
+ "primary_report_sha256": file_sha(report_path), "commit": commit,
+ "script_sha256": hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),
+ "command": shlex.join([sys.executable, *sys.argv]), "python": sys.version,
+ "pytorch": torch.__version__, "platform": platform.platform(), "device": "cpu", "gpu": "unused",
+ "threads": 1, "interop_threads": 1, "primary_validation_choices_unchanged": True,
+ "quality_budgets": {"ppl_relative": cfg.max_ppl_increase, "mean_kl_nats": cfg.max_kl, "top1_flip_fraction": cfg.max_top1_flip},
+ "checks": checks}
+ args.output.write_text(json.dumps(result, indent=2) + "\n")
+
+
+if __name__ == "__main__":
+ main()
diff --git a/docs/quantized_structural_evidence_20261031/build.txt b/docs/quantized_structural_evidence_20261031/build.txt
new file mode 100644
index 0000000..7cb4b6e
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/build.txt
@@ -0,0 +1,5 @@
+mkdir -p build
+cc -std=c11 -O2 -Wall -Wextra -Werror -fPIC -shared native/meft_rows.c -lz -o build/libmeft_rows.so
+cc -std=c11 -O2 -Wall -Wextra -Werror -Inative examples/native_gather.c native/meft_rows.c -lz -o build/native_gather
+mkdir -p build
+c++ -std=c++11 -O2 -Wall -Wextra -Werror -fPIC -shared -Isrc -Imylib/include native/meft_structural.cc -L./mylib/lib -ltrie -lpub -o build/libmeft_structural.so
diff --git a/docs/quantized_structural_evidence_20261031/codec-revision-check.json b/docs/quantized_structural_evidence_20261031/codec-revision-check.json
new file mode 100644
index 0000000..fd47559
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/codec-revision-check.json
@@ -0,0 +1,613 @@
+{
+ "study_commit": "3a54de2d8a7b97ee1432980c41b4edb0f64566d3",
+ "report_sha256": "eb73b44ef9ccd359160b38a371c2629f9c1f7ea8d0990450dc67ec125cfa08d0",
+ "revision_commit": "96fec051233c6c1c5ccb96fc31324b6f0a1ddd73",
+ "revision_codec_sha256": "83a4c206b171cc69788efa56023fd7df3138ee609e91ecbd084526d43e7e58c6",
+ "verification_script_sha256": "0dda2beaf2016aeb478d21c26799d285e3f951ff538def057c12948fe535a950",
+ "command": "/tmp/projects-push-20261028-venv/bin/python docs/quantized_structural_evidence_20261031/verify_codec_revision.py --report docs/quantized_structural_evidence_20261031/report.json --sgns-dir artifacts/native-ngram-20261018/trained-reference --gru-dir artifacts/causal-lm-compression-20261024 --output docs/quantized_structural_evidence_20261031/codec-revision-check.json",
+ "working_directory": "/Users/duyimin/Documents/Codex/2026-09-12/github/work/me_fasttext",
+ "python": "3.12.14 (main, Aug 25 2026, 13:50:33) [Clang 22.1.3 ]",
+ "pytorch": "2.14.1",
+ "numpy": "2.5.4",
+ "platform": "macOS-26.6.2-arm64-arm-64bit",
+ "device": "cpu",
+ "gpu": "unused",
+ "all_trained_encodings_unchanged": true,
+ "checks": [
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp32",
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "encoded_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp16",
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "encoded_sha256": "52145b2df4fa23885150f1e519fa52a4ec2a87a67d7d95a528f5872c4325106a",
+ "decoded_sha256": "296818b431b3b39a43a36929a876a94efefae21f382f9e84d4ec3286e3b60d62",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "int8",
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "encoded_sha256": "3b63b97be01c401d71fc1bf035e8bf67562f67ea700ec0b3e088fb0cf12870ee",
+ "decoded_sha256": "ff95d9bd81bb9e89be385208a5831f5c37605223682f339f292564a62faacc42",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp32",
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "encoded_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp16",
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "encoded_sha256": "4d5e89355e425d36118ebf293bc515da98012d2b8cfb74d7abef5c01f93e5d8b",
+ "decoded_sha256": "590fdfdcc048589926d117e6f06ff1ebd64a0fd2815cc57b8de70080aab734da",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "int8",
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "encoded_sha256": "1695a383af41a932d3e663426935548621422726475092855d2e0fca6971008f",
+ "decoded_sha256": "5208f73b4895fa6286c9b958779d6930895d8e6968f9ee970b34284151bd9200",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp32",
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "encoded_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp16",
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "encoded_sha256": "0ed0935ce9f5e17b24267182eca235c452de95389f0197f3762fa1b231326bce",
+ "decoded_sha256": "31d7dd18c6916ca11b7d50daa0b11adb6c6761085215b2407cd691841766649b",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "int8",
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "encoded_sha256": "39a5bfb46dcda9abae79229e8fb09380005212fafa61770793c05a633311fbe3",
+ "decoded_sha256": "49c40289aa5709af39591af7618b4d2796626664399a25a1da1c9f1231474746",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp32",
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "encoded_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp16",
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "encoded_sha256": "6246b2b9b270b1f329de11d92199f3fb5bb751123b26cb5a9f7b789465490a8b",
+ "decoded_sha256": "df8b79955d7468b7b2f0161394c5f6357e57ff0ab38d908a1ce9d750765b1361",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "int8",
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "encoded_sha256": "7c01f03bbb40873f368ce12e55598733cbd8271f035cb510761f8f5794d8d841",
+ "decoded_sha256": "9f777ab7bd00dedb4404ea8845e359fbaf82aaa095e98cfef4783fbcfeb33df4",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp32",
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "encoded_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp16",
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "encoded_sha256": "2b80ff97b7f3a0bdc95ac7eead482fad4fb6e9e7e25c67eb90f9c67d3d9aefef",
+ "decoded_sha256": "96287e438559898544cb2b53c3e3e1b1fc356fa0de75846da54b3e109092a525",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "int8",
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "encoded_sha256": "2f4c5b511ef141ff93c2e65b8dfe703922f9a8ceb888373c3404c3637462433c",
+ "decoded_sha256": "936ada1dd32ce053f2eae8e789655b9755eedb813052ae37096c04d212ea1676",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp32",
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "encoded_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp16",
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "encoded_sha256": "7181664941ce04822304ee6afc0b67f281d64784883f2bd11a965520e8851501",
+ "decoded_sha256": "745262c69f54fbc70dbe6d692411472e46846937961b94d94883afac06df3fa3",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "int8",
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "encoded_sha256": "145179f5873a93b683607b1475e5ff91182ff4d09cf914774cef2ef32c002ed1",
+ "decoded_sha256": "ed5fe2646195c36093dae73b4e27ecb710aaf4493c4643375246e406ec0f1dfd",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp32",
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "encoded_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp16",
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "encoded_sha256": "f398ae975042657c209fa7ae6041b7bf9b9ad537f36366d974ec5992fba17f9b",
+ "decoded_sha256": "965e67427fd2f63b0379f2c343356af421d0b17cd7971cffda2236625ac6d2b2",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "int8",
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "encoded_sha256": "4bdc5394076e8f3096248d2b1217897e6d6f7716cf114cd367d91f7100a27c2d",
+ "decoded_sha256": "d37754bf9356e1c82806b733d2f09d6fcd41d5b2725b70c407328b1a5c7fe149",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp32",
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "encoded_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp16",
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "encoded_sha256": "c09228c734adcadf26af9e437aa54e9dbbc0d275c3c2a829e871216873acebd6",
+ "decoded_sha256": "a6f992eab989830921c022ff3d6f88caceeb1eeebe107b7ccc88caeb09e9ca68",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "int8",
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "encoded_sha256": "0e2bdbfbc8dca09e94a85dd1f3349a0af8a377df57076f698fe6ceb081197dee",
+ "decoded_sha256": "f74165bb32ed3307f423ea4917e648ddb56eef15fa4521f69fd6f16fc3599f66",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp32",
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "encoded_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp16",
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "encoded_sha256": "891a04cf80ccfbc91b7e235af0b0a5c951005e27e177a089c05291a8bae10358",
+ "decoded_sha256": "58a8afde490d0e308f8add0008e5239d0de80c72e72d6d5ff0e13f47acc810a9",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "int8",
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "encoded_sha256": "c32e129ca5dd95b920e1532b4fa43658b2d93d5a6a71a0c7ba8b53ed4163e1c5",
+ "decoded_sha256": "1c37947f29f3da25c424fdd730690d3a4a1f55296f4034cf343a6892d5710e81",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp32",
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "encoded_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp16",
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "encoded_sha256": "080d6efb61408847148fbcc3c917499b8c29966c124b53c22e9aeca49ffbf968",
+ "decoded_sha256": "d87571fa9337131f33612c6b040cbda96d6836e255a7c1cac1b5b07b4f12bd42",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "int8",
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "encoded_sha256": "49e03d9e3f51e955fe94dc7e04708d1b097f5f5abbc78925511998693be3084e",
+ "decoded_sha256": "11d317c88faf40d248e5062c81409488c5bb22cad5013acb677fd2fa7b6038a7",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp32",
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "encoded_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp16",
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "encoded_sha256": "be3e6cb3faabcf8e7951e8696b5142f010d1678db8d4b7364a8077e60007e139",
+ "decoded_sha256": "811798cfb33aa60ba8fe68fa7bdc76afdb78f8ca17a89d0774d37291da0f6e18",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "int8",
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "encoded_sha256": "286e59cb1634e5510ae31256135a585ad96dc5281b739140ec86c3f850c402a8",
+ "decoded_sha256": "f6ffab41a644dd67e026935d7b50595b09e149dee71f6d058748771dac0f90df",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp32",
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "encoded_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp16",
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "encoded_sha256": "ae77e374b3687298bc905011989dd5cea95a297efdc59cae484cf38c92f8e8d1",
+ "decoded_sha256": "8f131fcc83bb844ad4e9d610693f55687d566532851a2724e31525d9369c4017",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "int8",
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "encoded_sha256": "33f22b0adf51608d72dafdd55c404069b6392555fde3ce9cbe2da3193605ee03",
+ "decoded_sha256": "c24044479e6740e64c7477e0129737cf3fc9833330c87891041560fc8979c289",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp32",
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "encoded_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp16",
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "encoded_sha256": "633749180e09d591673ccc3cf1af107d9c7a628d9e13375f718a588d76d8dfc8",
+ "decoded_sha256": "bac7d3501f88195092d63ba68dffadd151fe5e79cbc8b496139da8f07b095ef3",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "int8",
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "encoded_sha256": "23ed365ab52b2ccf14770d2c13290520d0466caddd4128ccac1db828a6843ff3",
+ "decoded_sha256": "7a21ec1590d5c8c461d39c7936eefcd8c2a50245cde875af4712b68897f74de2",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp32",
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "encoded_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp16",
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "encoded_sha256": "b9d730c7176cf58fede8bab1219fb42cb9ca785bdab964772b80efc4d4c0dcab",
+ "decoded_sha256": "5e1e9eb12dad3c914b817161ca8ba461f69a9a80c798a2f75f2a1c88f00a27a2",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "int8",
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "encoded_sha256": "7971ce6013cd10940ea0859b9f441072b9e4740e2e29e5921eacca0a1b7e229b",
+ "decoded_sha256": "825e9083d4fb773e7477d19b942e81d71aa7941df0d31375fb16b12950b6837f",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp32",
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "encoded_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp16",
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "encoded_sha256": "dc1bc47aaa923557f0743007570dfa78218b7b4f83c5f736f5d9ba4446f6a6c1",
+ "decoded_sha256": "add9a6b9c5875f02dc14dc3783c9953d7edb14426c4012649076815ddb6e1c71",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "int8",
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "encoded_sha256": "80c1610db9ec1aeaad5f260db8066dd9509b9548ce61e66c9563ff1efe6127ca",
+ "decoded_sha256": "fbf67b027d9c0471b313f44fdd647110f6a4ee32fdb01f42ab8203b4476c24a3",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp32",
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "encoded_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp16",
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "encoded_sha256": "90c4d98b998a5e102a57ac4c4bbec828fa37973f1ce190c7e5052edf80b98b29",
+ "decoded_sha256": "212e57b5fa9fcef9bbaf640d7bbb586eb6dec1a60e9e380099be530a0d30a4a2",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "int8",
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "encoded_sha256": "c6b56112cce3e1b0060ebf715a19f2aa46b58a9555a1fddb0277cd94bbec29c3",
+ "decoded_sha256": "2ac1cbfa580f67daaec9296dbc36834c852d35f0c2ac598a3062f05020c54a6d",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp32",
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "encoded_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp16",
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "encoded_sha256": "be124675b47f909e61b0388745ad00b3ed3a74bff02b3b5a333d38787d391165",
+ "decoded_sha256": "b9e46c364634283ab41e551138f1ada23921e31ee48a04dd3989a3da162f7b1f",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "int8",
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "encoded_sha256": "61f44b3bbe97af40b409ed69d9a05639fccb6f8c969bbfc0894eb7d256edc14c",
+ "decoded_sha256": "bb7617c7380056f8d86ab747df137ba8166ea9c736901a80e4911c8d837e3aa2",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp32",
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "encoded_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp16",
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "encoded_sha256": "0676720f8d94a19e6237b1d04171a8f31c1f922cb1452938976a65b4f39f5ef7",
+ "decoded_sha256": "bdea909d0738ce1a9112cfdf55e058286ffdf48616c95623b85c5865be33ab4e",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "int8",
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "encoded_sha256": "f84441c1beaa17ea1a7fb7964f5ef9138678520636b2201e52c2ff5a42fa6216",
+ "decoded_sha256": "126915e8d6c63c5a39ba052c391baeecb89382e4804054d3eecec1f79d5ff7a1",
+ "encoded_bytes_identical": true,
+ "decoded_bytes_identical": true
+ }
+ ],
+ "scope": "byte compatibility of a finite-range guard; benchmark timings remain tied to the original study commit"
+}
diff --git a/docs/quantized_structural_evidence_20261031/environment-extra.json b/docs/quantized_structural_evidence_20261031/environment-extra.json
new file mode 100644
index 0000000..7d6eb91
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/environment-extra.json
@@ -0,0 +1,11 @@
+{
+ "platform": "Linux-5.4.0-42-generic-x86_64-with-glibc2.35",
+ "cpu": "AMD EPYC 7742 64-Core Processor",
+ "logical_cpus": 255,
+ "host_MemTotal": "2113631756 kB",
+ "CUDA_VISIBLE_DEVICES": "",
+ "experiment_device": "cpu",
+ "experiment_GPU": "unused",
+ "worker_policy": "one PyTorch/interop/BLAS worker at nice 19",
+ "shared_host": "no exclusivity or CPU frequency pinning; no full-model latency/RSS claim"
+}
diff --git a/docs/quantized_structural_evidence_20261031/exit.txt b/docs/quantized_structural_evidence_20261031/exit.txt
new file mode 100644
index 0000000..573541a
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/exit.txt
@@ -0,0 +1 @@
+0
diff --git a/docs/quantized_structural_evidence_20261031/fixed099-test-audit.json b/docs/quantized_structural_evidence_20261031/fixed099-test-audit.json
new file mode 100644
index 0000000..7b7efd3
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/fixed099-test-audit.json
@@ -0,0 +1,994 @@
+{
+ "category": "post-main-run fixed-default test audit; not preregistered confirmatory evidence",
+ "primary_report_sha256": "eb73b44ef9ccd359160b38a371c2629f9c1f7ea8d0990450dc67ec125cfa08d0",
+ "commit": "3a54de2d8a7b97ee1432980c41b4edb0f64566d3",
+ "script_sha256": "dffa700970f77de375a1ee85b2ef8b9ee1e2b325400e5bf00fa75899da79fc3f",
+ "command": "/llm-align/duyimin/meft-cuda-20261030/compat/bin/python /llm-align/duyimin/ngram-quantized-20261031/audit_fixed099.py --study-root /llm-align/duyimin/ngram-quantized-20261031 --output /llm-align/duyimin/ngram-quantized-20261031/fixed099-test-audit.json",
+ "python": "3.11.15 (main, Mar 11 2026, 17:20:07) [GCC 14.3.0]",
+ "pytorch": "2.11.0+cu130",
+ "platform": "Linux-5.4.0-42-generic-x86_64-with-glibc2.35",
+ "device": "cpu",
+ "gpu": "unused",
+ "threads": 1,
+ "interop_threads": 1,
+ "primary_validation_choices_unchanged": true,
+ "quality_budgets": {
+ "ppl_relative": 0.005,
+ "mean_kl_nats": 0.0001,
+ "top1_flip_fraction": 0.01
+ },
+ "checks": [
+ {
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.99",
+ "checkpoint_sha256": "3576a59df8b4c48fb5cac466445e0472fd53af1450d4c7ffe30d612cf90a6214",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6103373426612577,
+ "character_perplexity": 5.004499174179104,
+ "next_character_accuracy": 0.5304761174576401,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -1.529360408315738e-05,
+ "mean_prediction_kl_nats": 1.7120831127003061e-06,
+ "prediction_kl_p99_nats": 2.61046410822912e-05,
+ "maximum_prediction_kl_nats": 0.006801602568104317,
+ "top1_changed_fraction": 0.00030667791152342254,
+ "accuracy_change": 0.00015333895576175038,
+ "maximum_absolute_logit_change": 0.23801040649414062,
+ "maximum_per_target_cross_entropy_increase": 0.04171037673950195
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -4.49829433719927e-05,
+ 1.578541096372571e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.99",
+ "checkpoint_sha256": "3576a59df8b4c48fb5cac466445e0472fd53af1450d4c7ffe30d612cf90a6214",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6103393421233734,
+ "character_perplexity": 5.004509180495614,
+ "next_character_accuracy": 0.5304761174576401,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -1.3294170547612814e-05,
+ "mean_prediction_kl_nats": 1.7180402708530163e-06,
+ "prediction_kl_p99_nats": 2.6131037333991233e-05,
+ "maximum_prediction_kl_nats": 0.006808221062134263,
+ "top1_changed_fraction": 0.00030667791152342254,
+ "accuracy_change": 0.00015333895576175038,
+ "maximum_absolute_logit_change": 0.2380363941192627,
+ "maximum_per_target_cross_entropy_increase": 0.0414736270904541
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -4.4146196455216834e-05,
+ 1.7773715874158032e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "int8",
+ "name": "int8-joint-0.99",
+ "checkpoint_sha256": "3576a59df8b4c48fb5cac466445e0472fd53af1450d4c7ffe30d612cf90a6214",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6103923585218516,
+ "character_perplexity": 5.004774508581822,
+ "next_character_accuracy": 0.5307827953691635,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 3.972292849695186e-05,
+ "mean_prediction_kl_nats": 4.650944368526283e-06,
+ "prediction_kl_p99_nats": 3.989562640500141e-05,
+ "maximum_prediction_kl_nats": 0.00023002610755576934,
+ "top1_changed_fraction": 0.001456720079736257,
+ "accuracy_change": 0.00046001686728514013,
+ "maximum_absolute_logit_change": 0.23500752449035645,
+ "maximum_per_target_cross_entropy_increase": 0.0469362735748291
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -1.9605531140350766e-05,
+ 9.964713218480278e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.99",
+ "checkpoint_sha256": "2a6ed05f8dd35bbd2851b5fa21fc34024df681e447fa8e426f5a8888bd13a567",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6155380906501848,
+ "character_perplexity": 5.030594110972346,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 1.926555862198874e-05,
+ "mean_prediction_kl_nats": 1.8032067723003644e-06,
+ "prediction_kl_p99_nats": 3.1012494730281556e-05,
+ "maximum_prediction_kl_nats": 0.007373327653580067,
+ "top1_changed_fraction": 7.666947788085564e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.27645206451416016,
+ "maximum_per_target_cross_entropy_increase": 0.05068206787109375
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -1.4909351380242913e-05,
+ 5.253368237548624e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.99",
+ "checkpoint_sha256": "2a6ed05f8dd35bbd2851b5fa21fc34024df681e447fa8e426f5a8888bd13a567",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6155369363808263,
+ "character_perplexity": 5.030588304315059,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 1.8111267692118655e-05,
+ "mean_prediction_kl_nats": 1.806716082822584e-06,
+ "prediction_kl_p99_nats": 3.0720688622257e-05,
+ "maximum_prediction_kl_nats": 0.007369757393799058,
+ "top1_changed_fraction": 7.666947788085564e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.2763652801513672,
+ "maximum_per_target_cross_entropy_increase": 0.050493717193603516
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -1.6793097959892104e-05,
+ 5.1660681147827815e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "int8",
+ "name": "int8-joint-0.99",
+ "checkpoint_sha256": "2a6ed05f8dd35bbd2851b5fa21fc34024df681e447fa8e426f5a8888bd13a567",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6155529379265168,
+ "character_perplexity": 5.030668802147704,
+ "next_character_accuracy": 0.5290960668557847,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 3.4113231218572145e-05,
+ "mean_prediction_kl_nats": 5.387190218382436e-06,
+ "prediction_kl_p99_nats": 4.374685590493577e-05,
+ "maximum_prediction_kl_nats": 0.007245252561058059,
+ "top1_changed_fraction": 0.0011500421682128345,
+ "accuracy_change": -0.00015333895576175038,
+ "maximum_absolute_logit_change": 0.2737741470336914,
+ "maximum_per_target_cross_entropy_increase": 0.04021501541137695
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -3.19065058511605e-05,
+ 9.508497865481893e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.99",
+ "checkpoint_sha256": "6f98acbd06c462d38cca0009605daa5986430f75993f62e281fc35ba0be9f1eb",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6054732126006208,
+ "character_perplexity": 4.98021574606221,
+ "next_character_accuracy": 0.5310128038028061,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -2.673351857648717e-05,
+ "mean_prediction_kl_nats": 4.141369138910643e-06,
+ "prediction_kl_p99_nats": 4.696098081167984e-05,
+ "maximum_prediction_kl_nats": 0.00394455171979604,
+ "top1_changed_fraction": 0.0002300084336425669,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.23698687553405762,
+ "maximum_per_target_cross_entropy_increase": 0.11077344417572021
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -8.15899644717028e-05,
+ 2.695571292072103e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.99",
+ "checkpoint_sha256": "6f98acbd06c462d38cca0009605daa5986430f75993f62e281fc35ba0be9f1eb",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6054735438230436,
+ "character_perplexity": 4.980217395621609,
+ "next_character_accuracy": 0.5310128038028061,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -2.6402304953656114e-05,
+ "mean_prediction_kl_nats": 4.123843285603775e-06,
+ "prediction_kl_p99_nats": 4.622206702970199e-05,
+ "maximum_prediction_kl_nats": 0.003955611389927352,
+ "top1_changed_fraction": 0.0002300084336425669,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.23705077171325684,
+ "maximum_per_target_cross_entropy_increase": 0.1109311580657959
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -8.179042015307988e-05,
+ 2.682603927912067e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "int8",
+ "name": "int8-joint-0.99",
+ "checkpoint_sha256": "6f98acbd06c462d38cca0009605daa5986430f75993f62e281fc35ba0be9f1eb",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6055010224164472,
+ "character_perplexity": 4.980354246870715,
+ "next_character_accuracy": 0.5312428122364486,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 1.0759404819319943e-06,
+ "mean_prediction_kl_nats": 7.821151145785465e-06,
+ "prediction_kl_p99_nats": 5.6903488780835576e-05,
+ "maximum_prediction_kl_nats": 0.003953598947138336,
+ "top1_changed_fraction": 0.0011500421682128345,
+ "accuracy_change": 0.00015333895576163936,
+ "maximum_absolute_logit_change": 0.24121397733688354,
+ "maximum_per_target_cross_entropy_increase": 0.11120235919952393
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -7.882844742669443e-05,
+ 7.906194596836203e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.99",
+ "checkpoint_sha256": "0e8078af3c09f2b3fc9f9326eeadf80476c38c986ac134e9d92181c35b724578",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7330032812638991,
+ "character_perplexity": 5.6576198426114725,
+ "next_character_accuracy": 0.5240358813156483,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 2.1604294667527183e-05,
+ "mean_prediction_kl_nats": 7.78168988858872e-06,
+ "prediction_kl_p99_nats": 0.00011058258695271181,
+ "maximum_prediction_kl_nats": 0.0022485554921807044,
+ "top1_changed_fraction": 0.0005366863451659894,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.24792087078094482,
+ "maximum_per_target_cross_entropy_increase": 0.08095335960388184
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -6.024743447126731e-05,
+ 0.00011064608135001141
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.99",
+ "checkpoint_sha256": "0e8078af3c09f2b3fc9f9326eeadf80476c38c986ac134e9d92181c35b724578",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7330043281755663,
+ "character_perplexity": 5.657625765642795,
+ "next_character_accuracy": 0.5240358813156483,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 2.2651229500529794e-05,
+ "mean_prediction_kl_nats": 7.79975451634245e-06,
+ "prediction_kl_p99_nats": 0.00010988026680244667,
+ "maximum_prediction_kl_nats": 0.002261641787947211,
+ "top1_changed_fraction": 0.0005366863451659894,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.24807631969451904,
+ "maximum_per_target_cross_entropy_increase": 0.08090543746948242
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -5.991744542410419e-05,
+ 0.00011220568407468823
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "int8",
+ "name": "int8-joint-0.99",
+ "checkpoint_sha256": "0e8078af3c09f2b3fc9f9326eeadf80476c38c986ac134e9d92181c35b724578",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.733011266983305,
+ "character_perplexity": 5.65766502295644,
+ "next_character_accuracy": 0.5241125507935291,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 2.959021848603195e-05,
+ "mean_prediction_kl_nats": 1.1626656487954728e-05,
+ "prediction_kl_p99_nats": 0.00013220916162419904,
+ "maximum_prediction_kl_nats": 0.002340190267592224,
+ "top1_changed_fraction": 0.0013033811239745456,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.2520638704299927,
+ "maximum_per_target_cross_entropy_increase": 0.0828244686126709
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -5.804055918160391e-05,
+ 0.00012844987515457383
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.99",
+ "checkpoint_sha256": "f773d277a5444631a7ca1865880469f92d01fd664f4da9363621026ccd307845",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7264698748348142,
+ "character_perplexity": 5.6207767991020114,
+ "next_character_accuracy": 0.5249559150502185,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -2.1575416673470984e-05,
+ "mean_prediction_kl_nats": 5.807643260194993e-06,
+ "prediction_kl_p99_nats": 0.00010141657576575024,
+ "maximum_prediction_kl_nats": 0.0007125883040050816,
+ "top1_changed_fraction": 0.0009967032124511231,
+ "accuracy_change": -0.00015333895576163936,
+ "maximum_absolute_logit_change": 0.25301599502563477,
+ "maximum_per_target_cross_entropy_increase": 0.05091524124145508
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -8.06561119497958e-05,
+ 4.467055448057322e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.99",
+ "checkpoint_sha256": "f773d277a5444631a7ca1865880469f92d01fd664f4da9363621026ccd307845",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.726467159977237,
+ "character_perplexity": 5.620761539514243,
+ "next_character_accuracy": 0.5249559150502185,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -2.42902119913508e-05,
+ "mean_prediction_kl_nats": 5.815575679108724e-06,
+ "prediction_kl_p99_nats": 0.00010166998443153778,
+ "maximum_prediction_kl_nats": 0.0007118815655093643,
+ "top1_changed_fraction": 0.0009967032124511231,
+ "accuracy_change": -0.00015333895576163936,
+ "maximum_absolute_logit_change": 0.2531658411026001,
+ "maximum_per_target_cross_entropy_increase": 0.0510706901550293
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -8.436666262374752e-05,
+ 4.227199612387919e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "int8",
+ "name": "int8-joint-0.99",
+ "checkpoint_sha256": "f773d277a5444631a7ca1865880469f92d01fd664f4da9363621026ccd307845",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7264412324560523,
+ "character_perplexity": 5.620615808989577,
+ "next_character_accuracy": 0.5250325845280993,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -5.021676728397291e-05,
+ "mean_prediction_kl_nats": 9.692142442536032e-06,
+ "prediction_kl_p99_nats": 0.000108525150499449,
+ "maximum_prediction_kl_nats": 0.0008209585404893964,
+ "top1_changed_fraction": 0.001456720079736257,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.25075554847717285,
+ "maximum_per_target_cross_entropy_increase": 0.050766944885253906
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00013128774527874365,
+ 3.87866733347212e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.99",
+ "checkpoint_sha256": "b8b0017a1ced25157ff108b6d38a5aaf5dfe56e59ce071a16755aa916bd279f5",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.72803269792906,
+ "character_perplexity": 5.629567946606118,
+ "next_character_accuracy": 0.5291727363336656,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -3.121919297631779e-05,
+ "mean_prediction_kl_nats": 1.0537167107346687e-05,
+ "prediction_kl_p99_nats": 0.00010991141652174422,
+ "maximum_prediction_kl_nats": 0.005400081972263237,
+ "top1_changed_fraction": 0.0005366863451659894,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.3041081428527832,
+ "maximum_per_target_cross_entropy_increase": 0.12840533256530762
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00011841685450632379,
+ 5.2997370424164765e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.99",
+ "checkpoint_sha256": "b8b0017a1ced25157ff108b6d38a5aaf5dfe56e59ce071a16755aa916bd279f5",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7280324823027406,
+ "character_perplexity": 5.629566732723234,
+ "next_character_accuracy": 0.5291727363336656,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -3.143481254055395e-05,
+ "mean_prediction_kl_nats": 1.0534875189048734e-05,
+ "prediction_kl_p99_nats": 0.00011052448372342929,
+ "maximum_prediction_kl_nats": 0.0053958865398814355,
+ "top1_changed_fraction": 0.0005366863451659894,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.30393123626708984,
+ "maximum_per_target_cross_entropy_increase": 0.12823724746704102
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00011842676496203541,
+ 5.1676833144090865e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "int8",
+ "name": "int8-joint-0.99",
+ "checkpoint_sha256": "b8b0017a1ced25157ff108b6d38a5aaf5dfe56e59ce071a16755aa916bd279f5",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7280487235241493,
+ "character_perplexity": 5.629658164505454,
+ "next_character_accuracy": 0.5293260752894273,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -1.5193969786458584e-05,
+ "mean_prediction_kl_nats": 1.4494837743019004e-05,
+ "prediction_kl_p99_nats": 0.00012026770353101802,
+ "maximum_prediction_kl_nats": 0.005769573978476713,
+ "top1_changed_fraction": 0.002070075902783102,
+ "accuracy_change": 7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.308438777923584,
+ "maximum_per_target_cross_entropy_increase": 0.1331770420074463
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0001269264060790776,
+ 8.718167565233331e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.99",
+ "checkpoint_sha256": "b1dc4a4b7a8ae2ed1ef9688aa0cccf047c766581b08fe1608fc5625cfc93d0d6",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.746792423275222,
+ "character_perplexity": 5.736173918029833,
+ "next_character_accuracy": 0.522579161235912,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00013713155786803245,
+ "mean_prediction_kl_nats": 8.387160403224065e-06,
+ "prediction_kl_p99_nats": 0.00010513767379755445,
+ "maximum_prediction_kl_nats": 0.016954457411952714,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 0.37196114659309387,
+ "maximum_per_target_cross_entropy_increase": 0.25717687606811523
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 4.1956194323943976e-05,
+ 0.0002322068940903893
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.99",
+ "checkpoint_sha256": "b1dc4a4b7a8ae2ed1ef9688aa0cccf047c766581b08fe1608fc5625cfc93d0d6",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467906956675798,
+ "character_perplexity": 5.736164008180495,
+ "next_character_accuracy": 0.522579161235912,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.0001354037148087213,
+ "mean_prediction_kl_nats": 8.39024307721582e-06,
+ "prediction_kl_p99_nats": 0.00010461888524421549,
+ "maximum_prediction_kl_nats": 0.0169355323084063,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 0.3721938729286194,
+ "maximum_per_target_cross_entropy_increase": 0.2576470375061035
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 4.106371111382881e-05,
+ 0.00023184921531607292
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "int8",
+ "name": "int8-joint-0.99",
+ "checkpoint_sha256": "b1dc4a4b7a8ae2ed1ef9688aa0cccf047c766581b08fe1608fc5625cfc93d0d6",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467712194244922,
+ "character_perplexity": 5.736052290343807,
+ "next_character_accuracy": 0.5227325001916737,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00011592502425189617,
+ "mean_prediction_kl_nats": 1.230479487328579e-05,
+ "prediction_kl_p99_nats": 0.00012420683382845062,
+ "maximum_prediction_kl_nats": 0.017946778591312353,
+ "top1_changed_fraction": 0.0019167369470213908,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.36850810050964355,
+ "maximum_per_target_cross_entropy_increase": 0.25458717346191406
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 1.0180332353287482e-05,
+ 0.00022089795485964647
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.99",
+ "checkpoint_sha256": "464242bc5b4367053b82ea7fa20f47536b2311fd0a3f4c4aaf6c0d4ed4a9e0c7",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467790163747046,
+ "character_perplexity": 5.736097014232285,
+ "next_character_accuracy": 0.5232691865368397,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -3.2684143610367924e-06,
+ "mean_prediction_kl_nats": 8.262291321689492e-06,
+ "prediction_kl_p99_nats": 0.00011778528468316299,
+ "maximum_prediction_kl_nats": 0.0065499537993800126,
+ "top1_changed_fraction": 0.0015333895576171127,
+ "accuracy_change": -0.00023000843364251455,
+ "maximum_absolute_logit_change": 0.3585476875305176,
+ "maximum_per_target_cross_entropy_increase": 0.06969976425170898
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -6.586554962312211e-05,
+ 5.9801644730695365e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.99",
+ "checkpoint_sha256": "464242bc5b4367053b82ea7fa20f47536b2311fd0a3f4c4aaf6c0d4ed4a9e0c7",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467771339496792,
+ "character_perplexity": 5.73608621646988,
+ "next_character_accuracy": 0.5232691865368397,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -5.150831462224922e-06,
+ "mean_prediction_kl_nats": 8.264931976770698e-06,
+ "prediction_kl_p99_nats": 0.00011714195461653951,
+ "maximum_prediction_kl_nats": 0.006543255282379136,
+ "top1_changed_fraction": 0.0015333895576171127,
+ "accuracy_change": -0.00023000843364251455,
+ "maximum_absolute_logit_change": 0.35859787464141846,
+ "maximum_per_target_cross_entropy_increase": 0.06992912292480469
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -6.79350738509818e-05,
+ 5.794176369511996e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "int8",
+ "name": "int8-joint-0.99",
+ "checkpoint_sha256": "464242bc5b4367053b82ea7fa20f47536b2311fd0a3f4c4aaf6c0d4ed4a9e0c7",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.746776289076794,
+ "character_perplexity": 5.736081370208216,
+ "next_character_accuracy": 0.5231925170589589,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -5.995699638609195e-06,
+ "mean_prediction_kl_nats": 1.1917872167478483e-05,
+ "prediction_kl_p99_nats": 0.00013011856410837118,
+ "maximum_prediction_kl_nats": 0.0066352149169198555,
+ "top1_changed_fraction": 0.0022234148585448133,
+ "accuracy_change": -0.00030667791152338975,
+ "maximum_absolute_logit_change": 0.3574690818786621,
+ "maximum_per_target_cross_entropy_increase": 0.06202554702758789
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -9.433998286496988e-05,
+ 9.024583100613763e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.99",
+ "checkpoint_sha256": "d2b8acfabccdc0f727220d8529bbe5b60c75d87bdefe10184232ad1310d868aa",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7495112993082373,
+ "character_perplexity": 5.7517910847663964,
+ "next_character_accuracy": 0.5253392624396228,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 7.424971586491312e-05,
+ "mean_prediction_kl_nats": 1.2781461306450973e-05,
+ "prediction_kl_p99_nats": 0.00013476159745037163,
+ "maximum_prediction_kl_nats": 0.004709698039633851,
+ "top1_changed_fraction": 0.0013033811239745456,
+ "accuracy_change": -0.00023000843364251455,
+ "maximum_absolute_logit_change": 0.38813066482543945,
+ "maximum_per_target_cross_entropy_increase": 0.19906091690063477
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -3.0042624596327334e-05,
+ 0.00018320990325721725
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.99",
+ "checkpoint_sha256": "d2b8acfabccdc0f727220d8529bbe5b60c75d87bdefe10184232ad1310d868aa",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7495114649184045,
+ "character_perplexity": 5.751792037321558,
+ "next_character_accuracy": 0.5253392624396228,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 7.441533834221121e-05,
+ "mean_prediction_kl_nats": 1.2761184812997986e-05,
+ "prediction_kl_p99_nats": 0.000135257071796772,
+ "maximum_prediction_kl_nats": 0.004700372572971191,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": -0.00023000843364251455,
+ "maximum_absolute_logit_change": 0.38828110694885254,
+ "maximum_per_target_cross_entropy_increase": 0.19883155822753906
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -2.8572421869474518e-05,
+ 0.00018264452205758513
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ {
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "int8",
+ "name": "int8-joint-0.99",
+ "checkpoint_sha256": "d2b8acfabccdc0f727220d8529bbe5b60c75d87bdefe10184232ad1310d868aa",
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7496088171948707,
+ "character_perplexity": 5.752352014627242,
+ "next_character_accuracy": 0.5254926013953845,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00017177959855028355,
+ "mean_prediction_kl_nats": 1.6229808150624952e-05,
+ "prediction_kl_p99_nats": 0.0001448249874305474,
+ "maximum_prediction_kl_nats": 0.004485428819249324,
+ "top1_changed_fraction": 0.0023767538143065247,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.3875594139099121,
+ "maximum_per_target_cross_entropy_increase": 0.2045726776123047
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 4.521165080624516e-05,
+ 0.00029340706023153957
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ }
+ ]
+}
diff --git a/docs/quantized_structural_evidence_20261031/rate-distortion.png b/docs/quantized_structural_evidence_20261031/rate-distortion.png
new file mode 100644
index 0000000..1e67b4f
Binary files /dev/null and b/docs/quantized_structural_evidence_20261031/rate-distortion.png differ
diff --git a/docs/quantized_structural_evidence_20261031/rate-distortion.svg b/docs/quantized_structural_evidence_20261031/rate-distortion.svg
new file mode 100644
index 0000000..0db38e0
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/rate-distortion.svg
@@ -0,0 +1,2103 @@
+
+
+
diff --git a/docs/quantized_structural_evidence_20261031/report.json b/docs/quantized_structural_evidence_20261031/report.json
new file mode 100644
index 0000000..161db67
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/report.json
@@ -0,0 +1,71221 @@
+{
+ "category": "cached public-text trained SGNS / causal GRU; CPU quantization + structural controls",
+ "commit": "3a54de2d8a7b97ee1432980c41b4edb0f64566d3",
+ "command": "/llm-align/duyimin/meft-cuda-20261030/compat/bin/python -m tools.benchmark_quantized_structural --sgns-dir artifacts/native-ngram-20261018/trained-reference --gru-dir artifacts/causal-lm-compression-20261024 --corpus artifacts/native-ngram-20261018/corpus/tinyshakespeare.txt --native-library build/libmeft_rows.so --structural-library build/libmeft_structural.so --planner native --output-dir /llm-align/duyimin/ngram-quantized-20261031/results --orders 4,8,12 --seeds 3,7,11 --thresholds .99,.98,.97,.95,.90 --norm-ratio .99 --repeat 40",
+ "working_directory": "/llm-align/duyimin/ngram-quantized-20261031/source",
+ "python": "3.11.15 (main, Mar 11 2026, 17:20:07) [GCC 14.3.0]",
+ "pytorch": "2.11.0+cu130",
+ "numpy": "2.2.6",
+ "platform": "Linux-5.4.0-42-generic-x86_64-with-glibc2.35",
+ "cpu": "AMD EPYC 7742 64-Core Processor",
+ "logical_cpus": 255,
+ "device": "cpu",
+ "gpu": "unused",
+ "threads": 1,
+ "interop_threads": 1,
+ "compiler": "cc (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0",
+ "source_sha256": {
+ "tools/quantized_memory.py": "393c84a8ee9d822c455ecbf077d584b9ee092a43dfce4842f5d0a9154ced779a",
+ "tools/benchmark_quantized_structural.py": "9613142ef7adbea588e97b2474a84009a37df412fac120ea7bbd27bdfd36d36a",
+ "tools/benchmark_cosine_merge.py": "0608a3312f7a7d97287b9a1f9d0d440bd4835ba9ba7437ab65116b269ca6a4c9",
+ "tools/structural_native.py": "acba60c392a5e672c903f96e2cf0427731791d82f9ce38b7334547f1dc498c3a",
+ "tools/native_row_store.py": "95d1d3b5e56c397d5e5f207702a1db018407eda62642887044003e16b2b3e8f8",
+ "tools/causal_ngram_lm.py": "16ec3fab33b5b2cddb1133ced7173fdf329e7a048def734a02c27457f4ddbd9e",
+ "native/meft_rows.c": "48c82769f3651635c091ef107254039de97b95a179520bd8e8ba874d08252258"
+ },
+ "working_tree_dirty": false,
+ "native_library_sha256": "49aa8e7e82c83e3ec0b103af62eeb3638d98a18a780a35bb5c80e435fc791c51",
+ "structural_library_sha256": "bd5a2b28a05e873426d943d01334d0beaf253944a3e9565d4514a807f21efb25",
+ "corpus_sha256": "86c4e6aa9db7c042ec79f339dcb96d42b0075e16b8fc2e86bf0ca57e2dc565ed",
+ "corpus_source": "https://raw.githubusercontent.com/karpathy/char-rnn/6f9487a6fe5b420b7ca9afb0d7c078e37c1d1b4e/data/tinyshakespeare/input.txt",
+ "training_report_sha256": {
+ "sgns": "d555aa72c2e761488e41b50b472cdfc88fd32fc29b73136e9effa522be050f02",
+ "gru": "0ab395574da8186552af5c342d8bc794dda3ef7c3c35a517397c431ea0ac3c2d"
+ },
+ "settings": {
+ "planner": "native",
+ "orders": "4,8,12",
+ "seeds": "3,7,11",
+ "thresholds": ".99,.98,.97,.95,.90",
+ "norm_ratio": 0.99,
+ "max_block_bytes": 8192,
+ "cache_bytes": 4194304,
+ "query_rows": 512,
+ "repeat": 40,
+ "burn_in": 64,
+ "eval_chunk": 256,
+ "eval_seed": 42,
+ "max_ppl_increase": 0.005,
+ "max_kl": 0.0001,
+ "max_top1_flip": 0.01,
+ "threshold_values": [
+ 0.99,
+ 0.98,
+ 0.97,
+ 0.95,
+ 0.9
+ ]
+ },
+ "selection_rule": "minimum complete backend + common key registry bytes within all validation budgets, separately by encoding; freeze before test; no test reselection",
+ "cost_assumptions": "existing cached weights/text; shared/local single CPU worker; no paid data or GPU; energy/currency cost unmeasured",
+ "limitations": [
+ "Small trained English SGNS/GRU references, not an official trained Engram checkpoint or paper-scale reproduction.",
+ "Reference planner is portable greedy prefix/suffix selection, not production DATrie; native planner is separately labeled.",
+ "Selection and primary storage totals use a common actual keys JSON; native runs also report a serialized DATrie alternative separately. No key lookup speed claim.",
+ "Full GRU bundle sums encoded backend, codec JSON, common keys JSON and actual serialized unchanged backbone/registry; excludes runtime/framework, logs and training checkpoint.",
+ "Row INT8 is symmetric nearest-even with float32 scale per row; no activation quantization or fine tuning.",
+ "Quantization occurs before planning; final error is measured against original float32 and quantization-only rows.",
+ "Filesystem pages are warm after row verification; first query has empty application cache, not disk-cold I/O.",
+ "Queries include Python/NumPy dequantization and allocations; no fused INT8 kernel, RSS, concurrent traffic or full-model latency claim.",
+ "One cached corpus, three seeds; SGNS reports vector distortion only, GRU reports character task quality.",
+ "Corpus license not independently established; raw text, keys, weights and serving bundles stay local."
+ ],
+ "runs": [
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 3,
+ "rows": 5333,
+ "dim": 128,
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "source_float32_bytes": 2730496,
+ "key_registry_bytes": 70934,
+ "unchanged_backbone_artifact_bytes": 0,
+ "alternative_datrie_key_index_bytes": 43056,
+ "query_rows": 512,
+ "query_ids_sha256": "87677d489e681e6c0a39b8460231cf990090e3318abb13c912a002c527931d01",
+ "frequency_trace_sha256": "0e7a36c623f2fc5cc1e4fd51bc27b7b98d60f0c1b31a56fb66253e951084f56a",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 512,
+ "storage": {
+ "artifact_bytes": 2730496,
+ "unique_rows": 5333,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "table_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.01689741713926196,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2731299,
+ "backend_with_key_registry_bytes": 2802233,
+ "backend_with_datrie_key_index_bytes": 2774355,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.398305,
+ "warm_p50_ms": 0.062508,
+ "warm_p95_ms": 0.0663244,
+ "warm_p99_ms": 0.06718761,
+ "samples_ms": [
+ 0.067187,
+ 0.065301,
+ 0.067188,
+ 0.065791,
+ 0.064812,
+ 0.06579,
+ 0.066279,
+ 0.064813,
+ 0.064813,
+ 0.06579,
+ 0.065302,
+ 0.063905,
+ 0.062927,
+ 0.062927,
+ 0.062997,
+ 0.061949,
+ 0.062438,
+ 0.063766,
+ 0.062857,
+ 0.062368,
+ 0.062019,
+ 0.062439,
+ 0.062508,
+ 0.062508,
+ 0.060552,
+ 0.062438,
+ 0.062508,
+ 0.062508,
+ 0.062508,
+ 0.062438,
+ 0.061461,
+ 0.060064,
+ 0.06202,
+ 0.062019,
+ 0.061949,
+ 0.061949,
+ 0.060483,
+ 0.062019,
+ 0.060972,
+ 0.060972
+ ],
+ "warm_rows_per_second": 8190951.558200551
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 2607844,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 334,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 334,
+ "zlib_blocks": 301,
+ "xor_blocks": 33,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "table_sha256": "6f073b2460802333d25edda9ad3a7793a83c167a107f7fb2adba18242a6ac9df",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25159990903921425,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2609048,
+ "backend_with_key_registry_bytes": 2679982,
+ "backend_with_datrie_key_index_bytes": 2652104,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.514488,
+ "warm_p50_ms": 0.070051,
+ "warm_p95_ms": 0.07391629999999999,
+ "warm_p99_ms": 0.07656017,
+ "samples_ms": [
+ 0.077105,
+ 0.075708,
+ 0.073333,
+ 0.073822,
+ 0.071937,
+ 0.071937,
+ 0.069631,
+ 0.070121,
+ 0.070958,
+ 0.069562,
+ 0.071028,
+ 0.070051,
+ 0.070051,
+ 0.07054,
+ 0.069981,
+ 0.069981,
+ 0.069632,
+ 0.07054,
+ 0.070051,
+ 0.069911,
+ 0.070051,
+ 0.07005,
+ 0.069562,
+ 0.069981,
+ 0.068655,
+ 0.073334,
+ 0.069632,
+ 0.069632,
+ 0.069981,
+ 0.069562,
+ 0.069632,
+ 0.069562,
+ 0.07047,
+ 0.069981,
+ 0.069981,
+ 0.07012,
+ 0.072425,
+ 0.071937,
+ 0.071937,
+ 0.070051
+ ],
+ "warm_rows_per_second": 7308960.614409502
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 2607844,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 334,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 334,
+ "zlib_blocks": 301,
+ "xor_blocks": 33,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "table_sha256": "6f073b2460802333d25edda9ad3a7793a83c167a107f7fb2adba18242a6ac9df",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.30454140016809106,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2609047,
+ "backend_with_key_registry_bytes": 2679981,
+ "backend_with_datrie_key_index_bytes": 2652103,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.504011,
+ "warm_p50_ms": 0.06994600000000001,
+ "warm_p95_ms": 0.07247485000000001,
+ "warm_p99_ms": 0.07455345999999999,
+ "samples_ms": [
+ 0.075289,
+ 0.071937,
+ 0.071937,
+ 0.071517,
+ 0.071937,
+ 0.071937,
+ 0.072006,
+ 0.071936,
+ 0.071937,
+ 0.071937,
+ 0.072425,
+ 0.072425,
+ 0.072356,
+ 0.072426,
+ 0.071378,
+ 0.069562,
+ 0.069073,
+ 0.069003,
+ 0.069143,
+ 0.073403,
+ 0.071936,
+ 0.069632,
+ 0.070051,
+ 0.069492,
+ 0.069632,
+ 0.069073,
+ 0.069562,
+ 0.069143,
+ 0.069562,
+ 0.069073,
+ 0.069562,
+ 0.069143,
+ 0.069562,
+ 0.069911,
+ 0.069073,
+ 0.069632,
+ 0.069982,
+ 0.069492,
+ 0.068654,
+ 0.069981
+ ],
+ "warm_rows_per_second": 7319932.519372086
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 2607844,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 334,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 334,
+ "zlib_blocks": 301,
+ "xor_blocks": 33,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "table_sha256": "6f073b2460802333d25edda9ad3a7793a83c167a107f7fb2adba18242a6ac9df",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2597450870089233,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2609047,
+ "backend_with_key_registry_bytes": 2679981,
+ "backend_with_datrie_key_index_bytes": 2652103,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.520145,
+ "warm_p50_ms": 0.070051,
+ "warm_p95_ms": 0.07524694999999999,
+ "warm_p99_ms": 0.08178527999999999,
+ "samples_ms": [
+ 0.075778,
+ 0.075219,
+ 0.073333,
+ 0.073333,
+ 0.073822,
+ 0.072425,
+ 0.073333,
+ 0.073822,
+ 0.072844,
+ 0.073263,
+ 0.073333,
+ 0.069981,
+ 0.069493,
+ 0.069562,
+ 0.069632,
+ 0.069493,
+ 0.069981,
+ 0.070959,
+ 0.069492,
+ 0.068584,
+ 0.070051,
+ 0.070051,
+ 0.069073,
+ 0.0704,
+ 0.0704,
+ 0.069631,
+ 0.070051,
+ 0.069004,
+ 0.069981,
+ 0.070051,
+ 0.085626,
+ 0.0704,
+ 0.069492,
+ 0.069073,
+ 0.068654,
+ 0.070889,
+ 0.068026,
+ 0.069143,
+ 0.069073,
+ 0.069073
+ ],
+ "warm_rows_per_second": 7308960.614409502
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1967696,
+ "unique_rows": 3997,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 250,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 250,
+ "zlib_blocks": 250,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3997
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "eeafd58493e15990ba17872aaeb6f4c349db6fb94a51dc3e78ef1310f3cb4daf",
+ "table_sha256": "bd06f338aeef379b593f0b2e56db41e22889fdc16c0df3679bea537fc0785997",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.18648262601345778,
+ "representatives_sha256": "ef4aaba8cd352e6da1938c6dd5a8b0c9b84e3d822af421eab1e84b6232912067",
+ "structurally_merged_rows": 1336,
+ "backend_artifact_bytes": 1968898,
+ "backend_with_key_registry_bytes": 2039832,
+ "backend_with_datrie_key_index_bytes": 2011954,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900040679606642,
+ "mean_cosine": 0.9991888464230989,
+ "p05_cosine": 0.9948229470039848,
+ "relative_l2_mean": 0.01884328678926973,
+ "relative_l2_p95": 0.10175910746440275,
+ "relative_l2_max": 0.14212765851757603,
+ "weighted_relative_l2_mean": 0.007670254799828044,
+ "weighted_relative_frobenius": 0.00684819310974004,
+ "weighted_mean_cosine": 0.9997023595827568,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900040679606642,
+ "mean_cosine": 0.9991888464230989,
+ "p05_cosine": 0.9948229470039848,
+ "relative_l2_mean": 0.01884328678926973,
+ "relative_l2_p95": 0.10175910746440275,
+ "relative_l2_max": 0.14212765851757603,
+ "weighted_relative_l2_mean": 0.007670254799828044,
+ "weighted_relative_frobenius": 0.00684819310974004,
+ "weighted_mean_cosine": 0.9997023595827568,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.429572,
+ "warm_p50_ms": 0.070958,
+ "warm_p95_ms": 0.07438085,
+ "warm_p99_ms": 0.08375999999999999,
+ "samples_ms": [
+ 0.075708,
+ 0.073823,
+ 0.073823,
+ 0.071517,
+ 0.071029,
+ 0.070959,
+ 0.071447,
+ 0.070959,
+ 0.070889,
+ 0.070958,
+ 0.069073,
+ 0.071029,
+ 0.069073,
+ 0.069492,
+ 0.073333,
+ 0.070958,
+ 0.069073,
+ 0.069073,
+ 0.069982,
+ 0.068095,
+ 0.069073,
+ 0.069073,
+ 0.088908,
+ 0.074311,
+ 0.070959,
+ 0.071448,
+ 0.071378,
+ 0.070959,
+ 0.068654,
+ 0.068584,
+ 0.07047,
+ 0.069073,
+ 0.069632,
+ 0.068584,
+ 0.069493,
+ 0.070959,
+ 0.069143,
+ 0.069562,
+ 0.071378,
+ 0.070959
+ ],
+ "warm_rows_per_second": 7215535.950844162
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1918572,
+ "unique_rows": 3895,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 244,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 244,
+ "zlib_blocks": 244,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3895
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "4a137b5bf285721c37e05981b0724a2710a24167455eeee315ce3daf2501efef",
+ "table_sha256": "2f0226e13568f5c5118c63140f4d44afc222f4488c2a2e045734b9bcc733053d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.17670658300630748,
+ "representatives_sha256": "bac02f935f7a7d9440de1468a50c989f3fdc61b31b8e915bbb3b0e68aad2309c",
+ "structurally_merged_rows": 1438,
+ "backend_artifact_bytes": 1919774,
+ "backend_with_key_registry_bytes": 1990708,
+ "backend_with_datrie_key_index_bytes": 1962830,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800525166930897,
+ "mean_cosine": 0.9989374160530267,
+ "p05_cosine": 0.9936510900109885,
+ "relative_l2_mean": 0.021932775249007413,
+ "relative_l2_p95": 0.11307263013339221,
+ "relative_l2_max": 0.19900629357720578,
+ "weighted_relative_l2_mean": 0.00870488587784729,
+ "weighted_relative_frobenius": 0.008087597127804755,
+ "weighted_mean_cosine": 0.999617368148953,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800525166930897,
+ "mean_cosine": 0.9989374160530267,
+ "p05_cosine": 0.9936510900109885,
+ "relative_l2_mean": 0.021932775249007413,
+ "relative_l2_p95": 0.11307263013339221,
+ "relative_l2_max": 0.19900629357720578,
+ "weighted_relative_l2_mean": 0.00870488587784729,
+ "weighted_relative_frobenius": 0.008087597127804755,
+ "weighted_mean_cosine": 0.999617368148953,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.459324,
+ "warm_p50_ms": 0.069143,
+ "warm_p95_ms": 0.073822,
+ "warm_p99_ms": 0.07527075,
+ "samples_ms": [
+ 0.076197,
+ 0.073263,
+ 0.073822,
+ 0.073333,
+ 0.073333,
+ 0.072844,
+ 0.071448,
+ 0.073822,
+ 0.071378,
+ 0.071378,
+ 0.069073,
+ 0.069143,
+ 0.068095,
+ 0.069074,
+ 0.069143,
+ 0.069003,
+ 0.069074,
+ 0.069073,
+ 0.069003,
+ 0.068165,
+ 0.069562,
+ 0.069003,
+ 0.068165,
+ 0.069562,
+ 0.069492,
+ 0.069143,
+ 0.069143,
+ 0.069004,
+ 0.070121,
+ 0.069073,
+ 0.069073,
+ 0.069143,
+ 0.069003,
+ 0.069003,
+ 0.069631,
+ 0.068165,
+ 0.069003,
+ 0.069562,
+ 0.068096,
+ 0.069003
+ ],
+ "warm_rows_per_second": 7404943.37821616
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1905402,
+ "unique_rows": 3868,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 242,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 242,
+ "zlib_blocks": 242,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3868
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "f509e53e8e15a5029ceb28d12a424c8a0de2172555baf317b58ddcad79813d6d",
+ "table_sha256": "492c98fd156ff1255b72982a41821d49f58278aa95ab2334341bd43c58738de0",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1765840109437704,
+ "representatives_sha256": "3d2722b4790e7f3828089970b17116fd5bb2830718d9d18f091dce19a7c23da6",
+ "structurally_merged_rows": 1465,
+ "backend_artifact_bytes": 1906604,
+ "backend_with_key_registry_bytes": 1977538,
+ "backend_with_datrie_key_index_bytes": 1949660,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9708564945641197,
+ "mean_cosine": 0.998818066779941,
+ "p05_cosine": 0.9932313242577301,
+ "relative_l2_mean": 0.02302971706719548,
+ "relative_l2_p95": 0.11659229645192355,
+ "relative_l2_max": 0.24115678654601577,
+ "weighted_relative_l2_mean": 0.009130428528990344,
+ "weighted_relative_frobenius": 0.009068352203125815,
+ "weighted_mean_cosine": 0.9995700085264972,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9708564945641197,
+ "mean_cosine": 0.998818066779941,
+ "p05_cosine": 0.9932313242577301,
+ "relative_l2_mean": 0.02302971706719548,
+ "relative_l2_p95": 0.11659229645192355,
+ "relative_l2_max": 0.24115678654601577,
+ "weighted_relative_l2_mean": 0.009130428528990344,
+ "weighted_relative_frobenius": 0.009068352203125815,
+ "weighted_mean_cosine": 0.9995700085264972,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.257902,
+ "warm_p50_ms": 0.069073,
+ "warm_p95_ms": 0.08381335,
+ "warm_p99_ms": 0.10016053999999999,
+ "samples_ms": [
+ 0.085207,
+ 0.076616,
+ 0.073822,
+ 0.072425,
+ 0.109721,
+ 0.073334,
+ 0.08374,
+ 0.081365,
+ 0.072356,
+ 0.072844,
+ 0.07047,
+ 0.069632,
+ 0.069073,
+ 0.068654,
+ 0.069073,
+ 0.067607,
+ 0.068724,
+ 0.069492,
+ 0.068514,
+ 0.068095,
+ 0.069981,
+ 0.068584,
+ 0.069003,
+ 0.069493,
+ 0.068654,
+ 0.068584,
+ 0.068584,
+ 0.068654,
+ 0.068095,
+ 0.068584,
+ 0.067676,
+ 0.069073,
+ 0.069562,
+ 0.068654,
+ 0.067606,
+ 0.067676,
+ 0.068585,
+ 0.069073,
+ 0.067676,
+ 0.069073
+ ],
+ "warm_rows_per_second": 7412447.700259147
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1902965,
+ "unique_rows": 3863,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 242,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 242,
+ "zlib_blocks": 242,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3863
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "ea0037236596c4044283efecdbd91c5c4ef1891882a2a3426bdd8d00cc461cc6",
+ "table_sha256": "5e9afb06f27da24c68c8af32ce619639990966c6905967e0885482dda50aac4a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.17827019095420837,
+ "representatives_sha256": "2d91381efe39349a4f7459fd91e99d9596260a0a212cd952ba52df34a1785bac",
+ "structurally_merged_rows": 1470,
+ "backend_artifact_bytes": 1904167,
+ "backend_with_key_registry_bytes": 1975101,
+ "backend_with_datrie_key_index_bytes": 1947223,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9504355998657713,
+ "mean_cosine": 0.9987848342638137,
+ "p05_cosine": 0.9930668648311156,
+ "relative_l2_mean": 0.023277964105222548,
+ "relative_l2_p95": 0.11772121655625935,
+ "relative_l2_max": 0.31347274120328616,
+ "weighted_relative_l2_mean": 0.00920108182463547,
+ "weighted_relative_frobenius": 0.0091938901341558,
+ "weighted_mean_cosine": 0.9995605502594191,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9504355998657713,
+ "mean_cosine": 0.9987848342638137,
+ "p05_cosine": 0.9930668648311156,
+ "relative_l2_mean": 0.023277964105222548,
+ "relative_l2_p95": 0.11772121655625935,
+ "relative_l2_max": 0.31347274120328616,
+ "weighted_relative_l2_mean": 0.00920108182463547,
+ "weighted_relative_frobenius": 0.0091938901341558,
+ "weighted_mean_cosine": 0.9995605502594191,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.17081,
+ "warm_p50_ms": 0.0692825,
+ "warm_p95_ms": 0.07312784999999998,
+ "warm_p99_ms": 0.0865967,
+ "samples_ms": [
+ 0.078502,
+ 0.091772,
+ 0.072845,
+ 0.070958,
+ 0.069632,
+ 0.070958,
+ 0.072425,
+ 0.070051,
+ 0.069981,
+ 0.069073,
+ 0.069632,
+ 0.069073,
+ 0.068584,
+ 0.069073,
+ 0.068584,
+ 0.069562,
+ 0.068165,
+ 0.070959,
+ 0.069492,
+ 0.069492,
+ 0.067187,
+ 0.067607,
+ 0.069562,
+ 0.069073,
+ 0.068165,
+ 0.068165,
+ 0.069004,
+ 0.069073,
+ 0.068654,
+ 0.067607,
+ 0.067187,
+ 0.069003,
+ 0.069143,
+ 0.069562,
+ 0.069422,
+ 0.069562,
+ 0.067187,
+ 0.069562,
+ 0.068165,
+ 0.069493
+ ],
+ "warm_rows_per_second": 7390033.558257857
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1898125,
+ "unique_rows": 3853,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 241,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 241,
+ "zlib_blocks": 241,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3853
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "ceef06043a75f8efe99c23f3abace42679238d890bc5c8c4e58117d380ceab5f",
+ "table_sha256": "22adc45ddfc7350dcc495198ebed5054bef3d9537cc51569d93401d662de0ed3",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.17404765309765935,
+ "representatives_sha256": "a560f77b4cdb526521bda23eaf1036640ed787744b9adb92198b323f829440b6",
+ "structurally_merged_rows": 1480,
+ "backend_artifact_bytes": 1899327,
+ "backend_with_key_registry_bytes": 1970261,
+ "backend_with_datrie_key_index_bytes": 1942383,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.905778580067121,
+ "mean_cosine": 0.9986470538910384,
+ "p05_cosine": 0.9928871918850045,
+ "relative_l2_mean": 0.023994103980148295,
+ "relative_l2_p95": 0.11947499907741611,
+ "relative_l2_max": 0.4327155490851253,
+ "weighted_relative_l2_mean": 0.009653463963460706,
+ "weighted_relative_frobenius": 0.01186738501987831,
+ "weighted_mean_cosine": 0.999472486628229,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.905778580067121,
+ "mean_cosine": 0.9986470538910384,
+ "p05_cosine": 0.9928871918850045,
+ "relative_l2_mean": 0.023994103980148295,
+ "relative_l2_p95": 0.11947499907741611,
+ "relative_l2_max": 0.4327155490851253,
+ "weighted_relative_l2_mean": 0.009653463963460706,
+ "weighted_relative_frobenius": 0.01186738501987831,
+ "weighted_mean_cosine": 0.999472486628229,
+ "max_relative_norm_change": 0.010098903037758608
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.169832,
+ "warm_p50_ms": 0.069492,
+ "warm_p95_ms": 0.07568785,
+ "warm_p99_ms": 0.09195322999999998,
+ "samples_ms": [
+ 0.075639,
+ 0.076616,
+ 0.074311,
+ 0.073403,
+ 0.072356,
+ 0.070959,
+ 0.072915,
+ 0.071937,
+ 0.071867,
+ 0.071936,
+ 0.071448,
+ 0.070959,
+ 0.070539,
+ 0.069492,
+ 0.069492,
+ 0.069562,
+ 0.068095,
+ 0.069492,
+ 0.069073,
+ 0.067677,
+ 0.069492,
+ 0.069562,
+ 0.067676,
+ 0.067676,
+ 0.067188,
+ 0.067746,
+ 0.068515,
+ 0.068584,
+ 0.068654,
+ 0.066698,
+ 0.068165,
+ 0.069492,
+ 0.068235,
+ 0.101759,
+ 0.068584,
+ 0.067676,
+ 0.069004,
+ 0.069632,
+ 0.068584,
+ 0.069492
+ ],
+ "warm_rows_per_second": 7367754.561676163
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 1365248,
+ "unique_rows": 5333,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "296818b431b3b39a43a36929a876a94efefae21f382f9e84d4ec3286e3b60d62",
+ "table_sha256": "52145b2df4fa23885150f1e519fa52a4ec2a87a67d7d95a528f5872c4325106a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.017670071916654706,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1366052,
+ "backend_with_key_registry_bytes": 1436986,
+ "backend_with_datrie_key_index_bytes": 1409108,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999597685811,
+ "mean_cosine": 0.9999999790028546,
+ "p05_cosine": 0.9999999730846733,
+ "relative_l2_mean": 0.00020681644286260717,
+ "relative_l2_p95": 0.00023407261360449034,
+ "relative_l2_max": 0.00028366458574764547,
+ "weighted_relative_l2_mean": 0.0002068006001193933,
+ "weighted_relative_frobenius": 0.00020696152811249227,
+ "weighted_mean_cosine": 0.9999999790474928,
+ "max_relative_norm_change": 0.00012388940698008832
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.471709,
+ "warm_p50_ms": 0.2202445,
+ "warm_p95_ms": 0.2232411,
+ "warm_p99_ms": 0.24625502,
+ "samples_ms": [
+ 0.221398,
+ 0.221397,
+ 0.220489,
+ 0.220489,
+ 0.220001,
+ 0.221886,
+ 0.219511,
+ 0.220489,
+ 0.220908,
+ 0.220001,
+ 0.220001,
+ 0.219581,
+ 0.222864,
+ 0.256388,
+ 0.22049,
+ 0.220489,
+ 0.22007,
+ 0.220559,
+ 0.219581,
+ 0.220489,
+ 0.220489,
+ 0.219093,
+ 0.219931,
+ 0.219512,
+ 0.219022,
+ 0.219512,
+ 0.219093,
+ 0.22,
+ 0.220489,
+ 0.22,
+ 0.230406,
+ 0.221956,
+ 0.220419,
+ 0.220489,
+ 0.219023,
+ 0.219512,
+ 0.219512,
+ 0.221886,
+ 0.219022,
+ 0.219581
+ ],
+ "warm_rows_per_second": 2324689.152283031
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1329270,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 167,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 167,
+ "zlib_blocks": 146,
+ "xor_blocks": 21,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "296818b431b3b39a43a36929a876a94efefae21f382f9e84d4ec3286e3b60d62",
+ "table_sha256": "f75dbfb59c763f1db9c4e95650a7059e44bb50c6bed40b09ce7fb914583d1ac9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.12936148000881076,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1330475,
+ "backend_with_key_registry_bytes": 1401409,
+ "backend_with_datrie_key_index_bytes": 1373531,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999597685811,
+ "mean_cosine": 0.9999999790028546,
+ "p05_cosine": 0.9999999730846733,
+ "relative_l2_mean": 0.00020681644286260717,
+ "relative_l2_p95": 0.00023407261360449034,
+ "relative_l2_max": 0.00028366458574764547,
+ "weighted_relative_l2_mean": 0.0002068006001193933,
+ "weighted_relative_frobenius": 0.00020696152811249227,
+ "weighted_mean_cosine": 0.9999999790474928,
+ "max_relative_norm_change": 0.00012388940698008832
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.122351,
+ "warm_p50_ms": 0.22901,
+ "warm_p95_ms": 0.24519529999999998,
+ "warm_p99_ms": 0.26012393,
+ "samples_ms": [
+ 0.229917,
+ 0.229499,
+ 0.263093,
+ 0.232781,
+ 0.22901,
+ 0.22901,
+ 0.229918,
+ 0.229987,
+ 0.229429,
+ 0.229429,
+ 0.228521,
+ 0.22901,
+ 0.229498,
+ 0.227543,
+ 0.229009,
+ 0.227543,
+ 0.228032,
+ 0.22894,
+ 0.229499,
+ 0.244654,
+ 0.230895,
+ 0.228521,
+ 0.229079,
+ 0.229499,
+ 0.229988,
+ 0.22901,
+ 0.228032,
+ 0.228102,
+ 0.228521,
+ 0.228521,
+ 0.228521,
+ 0.227613,
+ 0.22859,
+ 0.227124,
+ 0.22894,
+ 0.228102,
+ 0.25548,
+ 0.231804,
+ 0.231315,
+ 0.229429
+ ],
+ "warm_rows_per_second": 2235710.2309942795
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1329270,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 167,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 167,
+ "zlib_blocks": 146,
+ "xor_blocks": 21,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "296818b431b3b39a43a36929a876a94efefae21f382f9e84d4ec3286e3b60d62",
+ "table_sha256": "f75dbfb59c763f1db9c4e95650a7059e44bb50c6bed40b09ce7fb914583d1ac9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14403626509010792,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1330474,
+ "backend_with_key_registry_bytes": 1401408,
+ "backend_with_datrie_key_index_bytes": 1373530,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999597685811,
+ "mean_cosine": 0.9999999790028546,
+ "p05_cosine": 0.9999999730846733,
+ "relative_l2_mean": 0.00020681644286260717,
+ "relative_l2_p95": 0.00023407261360449034,
+ "relative_l2_max": 0.00028366458574764547,
+ "weighted_relative_l2_mean": 0.0002068006001193933,
+ "weighted_relative_frobenius": 0.00020696152811249227,
+ "weighted_mean_cosine": 0.9999999790474928,
+ "max_relative_norm_change": 0.00012388940698008832
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.118929,
+ "warm_p50_ms": 0.228102,
+ "warm_p95_ms": 0.23328039999999997,
+ "warm_p99_ms": 0.24966992999999998,
+ "samples_ms": [
+ 0.232781,
+ 0.22901,
+ 0.22901,
+ 0.229009,
+ 0.228032,
+ 0.229428,
+ 0.228521,
+ 0.228032,
+ 0.22901,
+ 0.229917,
+ 0.227962,
+ 0.228451,
+ 0.227054,
+ 0.227613,
+ 0.228101,
+ 0.242769,
+ 0.22901,
+ 0.229079,
+ 0.227613,
+ 0.22901,
+ 0.228032,
+ 0.227543,
+ 0.227054,
+ 0.228032,
+ 0.227124,
+ 0.227613,
+ 0.22859,
+ 0.228521,
+ 0.228102,
+ 0.227613,
+ 0.229079,
+ 0.228102,
+ 0.227543,
+ 0.254082,
+ 0.230477,
+ 0.227194,
+ 0.228521,
+ 0.227613,
+ 0.227543,
+ 0.227543
+ ],
+ "warm_rows_per_second": 2244609.86751541
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1329270,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 167,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 167,
+ "zlib_blocks": 146,
+ "xor_blocks": 21,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "296818b431b3b39a43a36929a876a94efefae21f382f9e84d4ec3286e3b60d62",
+ "table_sha256": "f75dbfb59c763f1db9c4e95650a7059e44bb50c6bed40b09ce7fb914583d1ac9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.12945192400366068,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1330474,
+ "backend_with_key_registry_bytes": 1401408,
+ "backend_with_datrie_key_index_bytes": 1373530,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999597685811,
+ "mean_cosine": 0.9999999790028546,
+ "p05_cosine": 0.9999999730846733,
+ "relative_l2_mean": 0.00020681644286260717,
+ "relative_l2_p95": 0.00023407261360449034,
+ "relative_l2_max": 0.00028366458574764547,
+ "weighted_relative_l2_mean": 0.0002068006001193933,
+ "weighted_relative_frobenius": 0.00020696152811249227,
+ "weighted_mean_cosine": 0.9999999790474928,
+ "max_relative_norm_change": 0.00012388940698008832
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.092878,
+ "warm_p50_ms": 0.227613,
+ "warm_p95_ms": 0.23241074999999997,
+ "warm_p99_ms": 0.25195856,
+ "samples_ms": [
+ 0.230896,
+ 0.229988,
+ 0.227613,
+ 0.228521,
+ 0.228032,
+ 0.227125,
+ 0.228032,
+ 0.228102,
+ 0.227612,
+ 0.227962,
+ 0.227054,
+ 0.227543,
+ 0.227613,
+ 0.257924,
+ 0.231873,
+ 0.227613,
+ 0.227054,
+ 0.226566,
+ 0.227543,
+ 0.228451,
+ 0.227124,
+ 0.226635,
+ 0.227613,
+ 0.227054,
+ 0.228452,
+ 0.226984,
+ 0.228102,
+ 0.227124,
+ 0.227613,
+ 0.227054,
+ 0.227543,
+ 0.242628,
+ 0.228102,
+ 0.228032,
+ 0.228102,
+ 0.226705,
+ 0.228101,
+ 0.227543,
+ 0.226076,
+ 0.226565
+ ],
+ "warm_rows_per_second": 2249432.150184743
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1009649,
+ "unique_rows": 3997,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 125,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 125,
+ "zlib_blocks": 125,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3997
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "56577f0054e9bfe36a676c73b2a50a564a8e54def7869e4135431b1876315339",
+ "table_sha256": "3a77aa8f9a9551522beca76ca19c5a0ad41d6a6c8cbb46d51f64e3dee3a45007",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.12075786292552948,
+ "representatives_sha256": "621c15475c9148850fa0f5e6b5803a1b164ef21e957b62c9921bb49c75625178",
+ "structurally_merged_rows": 1336,
+ "backend_artifact_bytes": 1010851,
+ "backend_with_key_registry_bytes": 1081785,
+ "backend_with_datrie_key_index_bytes": 1053907,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900084933818281,
+ "mean_cosine": 0.9991890700905423,
+ "p05_cosine": 0.9948229505519887,
+ "relative_l2_mean": 0.018995755979649764,
+ "relative_l2_p95": 0.1017615287758719,
+ "relative_l2_max": 0.14214298103013276,
+ "weighted_relative_l2_mean": 0.007851847095531779,
+ "weighted_relative_frobenius": 0.0068512795300496,
+ "weighted_mean_cosine": 0.999702408364441,
+ "max_relative_norm_change": 0.01001484036538288
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900086049689072,
+ "mean_cosine": 0.9991890832133987,
+ "p05_cosine": 0.9948215430114717,
+ "relative_l2_mean": 0.018840930858532067,
+ "relative_l2_p95": 0.10175836805142259,
+ "relative_l2_max": 0.14215957344489455,
+ "weighted_relative_l2_mean": 0.007669569439410469,
+ "weighted_relative_frobenius": 0.006848195096817759,
+ "weighted_mean_cosine": 0.9997024267715556,
+ "max_relative_norm_change": 0.009992405462840135
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.25981,
+ "warm_p50_ms": 0.228521,
+ "warm_p95_ms": 0.24101915000000002,
+ "warm_p99_ms": 0.28488896999999996,
+ "samples_ms": [
+ 0.231803,
+ 0.311283,
+ 0.240813,
+ 0.235575,
+ 0.227543,
+ 0.228521,
+ 0.228521,
+ 0.227054,
+ 0.227124,
+ 0.233689,
+ 0.231385,
+ 0.229918,
+ 0.230406,
+ 0.228521,
+ 0.230336,
+ 0.22894,
+ 0.228451,
+ 0.227124,
+ 0.243606,
+ 0.229499,
+ 0.228032,
+ 0.227194,
+ 0.228101,
+ 0.227054,
+ 0.228521,
+ 0.228521,
+ 0.227473,
+ 0.228451,
+ 0.228521,
+ 0.227473,
+ 0.228452,
+ 0.228521,
+ 0.228521,
+ 0.22894,
+ 0.226985,
+ 0.240883,
+ 0.22901,
+ 0.227963,
+ 0.228032,
+ 0.227054
+ ],
+ "warm_rows_per_second": 2240494.309056936
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 985005,
+ "unique_rows": 3896,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 122,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 122,
+ "zlib_blocks": 122,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3896
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "f5d33231e24c8f132f83b9c5008bb3677ec1e68973b17764be083a3cbbf8c302",
+ "table_sha256": "f9284dd9668c818554f06daf9c604550bc371fbcc5510513090a8419659f4547",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.11457613506354392,
+ "representatives_sha256": "19ec1b57421ee9ee3cd9ca87aea2ac4337628b44f7866c1d7f41de2cd9b7d0e2",
+ "structurally_merged_rows": 1437,
+ "backend_artifact_bytes": 986206,
+ "backend_with_key_registry_bytes": 1057140,
+ "backend_with_datrie_key_index_bytes": 1029262,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800489128113278,
+ "mean_cosine": 0.9989397183095924,
+ "p05_cosine": 0.993663425740816,
+ "relative_l2_mean": 0.02205349566786507,
+ "relative_l2_p95": 0.11278343686234345,
+ "relative_l2_max": 0.19902522582008436,
+ "weighted_relative_l2_mean": 0.008877263586252356,
+ "weighted_relative_frobenius": 0.00808873698855015,
+ "weighted_mean_cosine": 0.9996180084835264,
+ "max_relative_norm_change": 0.01001484036538288
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800460873989189,
+ "mean_cosine": 0.9989397316916933,
+ "p05_cosine": 0.9936630700639634,
+ "relative_l2_mean": 0.02190255174227501,
+ "relative_l2_p95": 0.11276128827773571,
+ "relative_l2_max": 0.19903711226739598,
+ "weighted_relative_l2_mean": 0.008696273342306349,
+ "weighted_relative_frobenius": 0.00808615720538015,
+ "weighted_mean_cosine": 0.999618026372571,
+ "max_relative_norm_change": 0.009992405462840135
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.993994,
+ "warm_p50_ms": 0.228486,
+ "warm_p95_ms": 0.23786174999999998,
+ "warm_p99_ms": 0.26060998999999996,
+ "samples_ms": [
+ 0.231873,
+ 0.233689,
+ 0.233759,
+ 0.230407,
+ 0.228101,
+ 0.228591,
+ 0.271124,
+ 0.23753,
+ 0.230825,
+ 0.229429,
+ 0.229359,
+ 0.228032,
+ 0.227613,
+ 0.228032,
+ 0.227054,
+ 0.228102,
+ 0.22901,
+ 0.227683,
+ 0.227613,
+ 0.227124,
+ 0.228451,
+ 0.228032,
+ 0.232292,
+ 0.244165,
+ 0.232222,
+ 0.230406,
+ 0.228032,
+ 0.228521,
+ 0.226635,
+ 0.227124,
+ 0.228521,
+ 0.229499,
+ 0.230895,
+ 0.228939,
+ 0.227054,
+ 0.227054,
+ 0.226705,
+ 0.227124,
+ 0.228102,
+ 0.226565
+ ],
+ "warm_rows_per_second": 2240837.5130204917
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 978338,
+ "unique_rows": 3869,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 121,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 121,
+ "zlib_blocks": 121,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3869
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "8f0202a7fb5ec31e3b8888cd7c7c4fdbe10c3ddcf39c785ccac9558634f3a14e",
+ "table_sha256": "97d3e9c3d1f460ab967627667c11aa1c241988f4bb31da7f4c2dcb861932f020",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10647167498245835,
+ "representatives_sha256": "073ba4d3a1b4ac31faaa46a5ca0e2db48458c5f20b358f8c8591f17855fa1c88",
+ "structurally_merged_rows": 1464,
+ "backend_artifact_bytes": 979539,
+ "backend_with_key_registry_bytes": 1050473,
+ "backend_with_datrie_key_index_bytes": 1022595,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9708490360808352,
+ "mean_cosine": 0.9988203685915727,
+ "p05_cosine": 0.9932592470627866,
+ "relative_l2_mean": 0.02314940399431931,
+ "relative_l2_p95": 0.11618531979442444,
+ "relative_l2_max": 0.2411893391599019,
+ "weighted_relative_l2_mean": 0.009302414627571623,
+ "weighted_relative_frobenius": 0.009069396417274616,
+ "weighted_mean_cosine": 0.9995706486644645,
+ "max_relative_norm_change": 0.01001484036538288
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.970842172012824,
+ "mean_cosine": 0.9988203783840274,
+ "p05_cosine": 0.9932570907999037,
+ "relative_l2_mean": 0.022999513106467594,
+ "relative_l2_p95": 0.11618595519489394,
+ "relative_l2_max": 0.24122132562691045,
+ "weighted_relative_l2_mean": 0.009121825212211722,
+ "weighted_relative_frobenius": 0.009067214030164977,
+ "weighted_mean_cosine": 0.9995706644779443,
+ "max_relative_norm_change": 0.009992405462840135
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.23376,
+ "warm_p50_ms": 0.2301625,
+ "warm_p95_ms": 0.23755879999999996,
+ "warm_p99_ms": 0.2555154,
+ "samples_ms": [
+ 0.237042,
+ 0.230896,
+ 0.228521,
+ 0.230895,
+ 0.230337,
+ 0.229918,
+ 0.229987,
+ 0.227613,
+ 0.229498,
+ 0.22908,
+ 0.228102,
+ 0.228591,
+ 0.229009,
+ 0.22908,
+ 0.228102,
+ 0.22908,
+ 0.247378,
+ 0.236483,
+ 0.23327,
+ 0.230407,
+ 0.231315,
+ 0.22901,
+ 0.229429,
+ 0.227962,
+ 0.227543,
+ 0.230895,
+ 0.229917,
+ 0.229988,
+ 0.22901,
+ 0.236622,
+ 0.230477,
+ 0.231804,
+ 0.231874,
+ 0.260718,
+ 0.232781,
+ 0.230337,
+ 0.2332,
+ 0.230896,
+ 0.230896,
+ 0.229987
+ ],
+ "warm_rows_per_second": 2224515.28811166
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 977146,
+ "unique_rows": 3864,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 121,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 121,
+ "zlib_blocks": 121,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3864
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "3396955427481df83fa6aa3f893710aa3a49e20860d1219929761de5f9a4e75a",
+ "table_sha256": "39362b7ec3317f68b01ca1a14ade56e072e51dd6445e5d2702775e4432571e14",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.12416584091261029,
+ "representatives_sha256": "25bf0bd565e884f16ce8cbf05f9fff8a630df54e9140a7ee68b4d5f5517d239b",
+ "structurally_merged_rows": 1469,
+ "backend_artifact_bytes": 978347,
+ "backend_with_key_registry_bytes": 1049281,
+ "backend_with_datrie_key_index_bytes": 1021403,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.950434353491954,
+ "mean_cosine": 0.9987871357522384,
+ "p05_cosine": 0.9931252291914241,
+ "relative_l2_mean": 0.02339745145168532,
+ "relative_l2_p95": 0.11726686621678044,
+ "relative_l2_max": 0.3134766420480686,
+ "weighted_relative_l2_mean": 0.009373011120801047,
+ "weighted_relative_frobenius": 0.009194923000522387,
+ "weighted_mean_cosine": 0.999561190305399,
+ "max_relative_norm_change": 0.01001484036538288
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9504358435378185,
+ "mean_cosine": 0.9987871427722179,
+ "p05_cosine": 0.9931260463402736,
+ "relative_l2_mean": 0.023247770374172798,
+ "relative_l2_p95": 0.11729238713488532,
+ "relative_l2_max": 0.31347037234704306,
+ "weighted_relative_l2_mean": 0.009192481419313435,
+ "weighted_relative_frobenius": 0.009192769743581646,
+ "weighted_mean_cosine": 0.9995612053298076,
+ "max_relative_norm_change": 0.009992405462840135
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.125436,
+ "warm_p50_ms": 0.228765,
+ "warm_p95_ms": 0.23528159999999995,
+ "warm_p99_ms": 0.25599066000000004,
+ "samples_ms": [
+ 0.23362,
+ 0.229499,
+ 0.229429,
+ 0.228451,
+ 0.22894,
+ 0.229918,
+ 0.228521,
+ 0.229987,
+ 0.228032,
+ 0.228451,
+ 0.229499,
+ 0.229079,
+ 0.261765,
+ 0.234667,
+ 0.230407,
+ 0.229499,
+ 0.22852,
+ 0.22901,
+ 0.228102,
+ 0.227543,
+ 0.229429,
+ 0.22901,
+ 0.228521,
+ 0.228032,
+ 0.228451,
+ 0.228521,
+ 0.228521,
+ 0.22852,
+ 0.229918,
+ 0.227055,
+ 0.246959,
+ 0.22901,
+ 0.22908,
+ 0.22859,
+ 0.227613,
+ 0.228102,
+ 0.229009,
+ 0.228102,
+ 0.228032,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2238104.605162503
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 974750,
+ "unique_rows": 3854,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 121,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 121,
+ "zlib_blocks": 121,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3854
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "13e42c39bd70190e08245f5ac4a7783b4b009c202e190736a4525352897caabc",
+ "table_sha256": "739154e18f71b39e7ed249adea90d7312072f700f1d253c3f2e77f735021dce8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10525943897664547,
+ "representatives_sha256": "3d8c5e73c20f030e595ce9bf416febfd1467bafe0a85b0aaf71a2d9d8dba784b",
+ "structurally_merged_rows": 1479,
+ "backend_artifact_bytes": 975951,
+ "backend_with_key_registry_bytes": 1046885,
+ "backend_with_datrie_key_index_bytes": 1019007,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9057759604519998,
+ "mean_cosine": 0.9986493500187473,
+ "p05_cosine": 0.9929141331823574,
+ "relative_l2_mean": 0.024113234275497934,
+ "relative_l2_p95": 0.11913783578599228,
+ "relative_l2_max": 0.4327194967056403,
+ "weighted_relative_l2_mean": 0.009825166011066526,
+ "weighted_relative_frobenius": 0.011868171826604816,
+ "weighted_mean_cosine": 0.9994731241847539,
+ "max_relative_norm_change": 0.01001484036538288
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9057849138351362,
+ "mean_cosine": 0.998649356363167,
+ "p05_cosine": 0.9929129395400632,
+ "relative_l2_mean": 0.023963917739104442,
+ "relative_l2_p95": 0.1191314173908099,
+ "relative_l2_max": 0.4326941991008237,
+ "weighted_relative_l2_mean": 0.00964486284567928,
+ "weighted_relative_frobenius": 0.011866546570363908,
+ "weighted_mean_cosine": 0.9994731399294087,
+ "max_relative_norm_change": 0.009992405462840135
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.143944,
+ "warm_p50_ms": 0.22845149999999997,
+ "warm_p95_ms": 0.23314149999999997,
+ "warm_p99_ms": 0.24275075,
+ "samples_ms": [
+ 0.231803,
+ 0.229498,
+ 0.22901,
+ 0.229429,
+ 0.227473,
+ 0.228033,
+ 0.243677,
+ 0.229918,
+ 0.228032,
+ 0.228521,
+ 0.22894,
+ 0.227124,
+ 0.227543,
+ 0.227613,
+ 0.228521,
+ 0.228591,
+ 0.22901,
+ 0.227613,
+ 0.228102,
+ 0.226565,
+ 0.227962,
+ 0.227124,
+ 0.226635,
+ 0.228102,
+ 0.241302,
+ 0.228521,
+ 0.22887,
+ 0.228451,
+ 0.227962,
+ 0.232363,
+ 0.230896,
+ 0.232712,
+ 0.22901,
+ 0.227543,
+ 0.228452,
+ 0.22894,
+ 0.227054,
+ 0.228451,
+ 0.227962,
+ 0.228032
+ ],
+ "warm_rows_per_second": 2241175.9169889456
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 132,
+ "storage": {
+ "artifact_bytes": 703956,
+ "unique_rows": 5333,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "ff95d9bd81bb9e89be385208a5831f5c37605223682f339f292564a62faacc42",
+ "table_sha256": "3b63b97be01c401d71fc1bf035e8bf67562f67ea700ec0b3e088fb0cf12870ee",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.016697739949449897,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 704803,
+ "backend_with_key_registry_bytes": 775737,
+ "backend_with_datrie_key_index_bytes": 747859,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999472196271418,
+ "mean_cosine": 0.9999785442031286,
+ "p05_cosine": 0.9999672497233414,
+ "relative_l2_mean": 0.006517774800990935,
+ "relative_l2_p95": 0.008115772137778587,
+ "relative_l2_max": 0.010321824347442474,
+ "weighted_relative_l2_mean": 0.0064498560165530084,
+ "weighted_relative_frobenius": 0.006476714793365053,
+ "weighted_mean_cosine": 0.9999790489880547,
+ "max_relative_norm_change": 0.002208400347090758
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.324622,
+ "warm_p50_ms": 0.12253649999999999,
+ "warm_p95_ms": 0.12406324999999999,
+ "warm_p99_ms": 0.14719789999999996,
+ "samples_ms": [
+ 0.123479,
+ 0.122152,
+ 0.120686,
+ 0.122083,
+ 0.122501,
+ 0.123409,
+ 0.122571,
+ 0.122572,
+ 0.122641,
+ 0.122571,
+ 0.123969,
+ 0.120546,
+ 0.12299,
+ 0.160844,
+ 0.125854,
+ 0.122082,
+ 0.122991,
+ 0.122991,
+ 0.122502,
+ 0.121664,
+ 0.122991,
+ 0.120686,
+ 0.121105,
+ 0.121105,
+ 0.122502,
+ 0.121593,
+ 0.121105,
+ 0.121104,
+ 0.12348,
+ 0.122502,
+ 0.121594,
+ 0.122571,
+ 0.122572,
+ 0.122083,
+ 0.123479,
+ 0.122501,
+ 0.121524,
+ 0.122991,
+ 0.122571,
+ 0.12306
+ ],
+ "warm_rows_per_second": 4178346.859915209
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 710740,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 87,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 86,
+ "zlib_blocks": 39,
+ "xor_blocks": 47,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "ff95d9bd81bb9e89be385208a5831f5c37605223682f339f292564a62faacc42",
+ "table_sha256": "c912a7850306da43eeb6fc979dda6b777254023fc7c02b40c915fbcbe5f843ef",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.09135067299939692,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 711984,
+ "backend_with_key_registry_bytes": 782918,
+ "backend_with_datrie_key_index_bytes": 755040,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999472196271418,
+ "mean_cosine": 0.9999785442031286,
+ "p05_cosine": 0.9999672497233414,
+ "relative_l2_mean": 0.006517774800990935,
+ "relative_l2_p95": 0.008115772137778587,
+ "relative_l2_max": 0.010321824347442474,
+ "weighted_relative_l2_mean": 0.0064498560165530084,
+ "weighted_relative_frobenius": 0.006476714793365053,
+ "weighted_mean_cosine": 0.9999790489880547,
+ "max_relative_norm_change": 0.002208400347090758
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.465923,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.13570454999999998,
+ "warm_p99_ms": 0.15485767999999997,
+ "samples_ms": [
+ 0.135701,
+ 0.132489,
+ 0.133396,
+ 0.133467,
+ 0.132,
+ 0.132489,
+ 0.132489,
+ 0.132,
+ 0.132978,
+ 0.132489,
+ 0.132489,
+ 0.130114,
+ 0.130533,
+ 0.16706,
+ 0.135772,
+ 0.132978,
+ 0.131581,
+ 0.132489,
+ 0.132,
+ 0.132978,
+ 0.130114,
+ 0.129625,
+ 0.131511,
+ 0.132,
+ 0.131022,
+ 0.133466,
+ 0.132489,
+ 0.132489,
+ 0.132,
+ 0.131512,
+ 0.130115,
+ 0.132489,
+ 0.131092,
+ 0.131022,
+ 0.130114,
+ 0.132489,
+ 0.130044,
+ 0.131092,
+ 0.131092,
+ 0.129626
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 710740,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 87,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 86,
+ "zlib_blocks": 39,
+ "xor_blocks": 47,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "ff95d9bd81bb9e89be385208a5831f5c37605223682f339f292564a62faacc42",
+ "table_sha256": "c912a7850306da43eeb6fc979dda6b777254023fc7c02b40c915fbcbe5f843ef",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.0982857090421021,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 711983,
+ "backend_with_key_registry_bytes": 782917,
+ "backend_with_datrie_key_index_bytes": 755039,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999472196271418,
+ "mean_cosine": 0.9999785442031286,
+ "p05_cosine": 0.9999672497233414,
+ "relative_l2_mean": 0.006517774800990935,
+ "relative_l2_p95": 0.008115772137778587,
+ "relative_l2_max": 0.010321824347442474,
+ "weighted_relative_l2_mean": 0.0064498560165530084,
+ "weighted_relative_frobenius": 0.006476714793365053,
+ "weighted_mean_cosine": 0.9999790489880547,
+ "max_relative_norm_change": 0.002208400347090758
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.461244,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.1349094,
+ "warm_p99_ms": 0.14126749,
+ "samples_ms": [
+ 0.134793,
+ 0.134305,
+ 0.132489,
+ 0.132,
+ 0.132489,
+ 0.132419,
+ 0.132489,
+ 0.135772,
+ 0.134864,
+ 0.132419,
+ 0.131512,
+ 0.130604,
+ 0.131511,
+ 0.144781,
+ 0.134235,
+ 0.132978,
+ 0.134305,
+ 0.132001,
+ 0.131023,
+ 0.131581,
+ 0.132,
+ 0.132908,
+ 0.132,
+ 0.132,
+ 0.132,
+ 0.131023,
+ 0.129626,
+ 0.130603,
+ 0.132001,
+ 0.132001,
+ 0.131022,
+ 0.131022,
+ 0.130114,
+ 0.129625,
+ 0.132908,
+ 0.133466,
+ 0.131511,
+ 0.131022,
+ 0.131092,
+ 0.130114
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 710740,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 87,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 86,
+ "zlib_blocks": 39,
+ "xor_blocks": 47,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "ff95d9bd81bb9e89be385208a5831f5c37605223682f339f292564a62faacc42",
+ "table_sha256": "c912a7850306da43eeb6fc979dda6b777254023fc7c02b40c915fbcbe5f843ef",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.0779950509313494,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 711983,
+ "backend_with_key_registry_bytes": 782917,
+ "backend_with_datrie_key_index_bytes": 755039,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999472196271418,
+ "mean_cosine": 0.9999785442031286,
+ "p05_cosine": 0.9999672497233414,
+ "relative_l2_mean": 0.006517774800990935,
+ "relative_l2_p95": 0.008115772137778587,
+ "relative_l2_max": 0.010321824347442474,
+ "weighted_relative_l2_mean": 0.0064498560165530084,
+ "weighted_relative_frobenius": 0.006476714793365053,
+ "weighted_mean_cosine": 0.9999790489880547,
+ "max_relative_norm_change": 0.002208400347090758
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.493372,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.13299895,
+ "warm_p99_ms": 0.14238657,
+ "samples_ms": [
+ 0.132978,
+ 0.132,
+ 0.131511,
+ 0.132978,
+ 0.131512,
+ 0.130603,
+ 0.130114,
+ 0.130184,
+ 0.129695,
+ 0.132,
+ 0.131022,
+ 0.130534,
+ 0.128648,
+ 0.132,
+ 0.131092,
+ 0.132908,
+ 0.131511,
+ 0.132489,
+ 0.132,
+ 0.132978,
+ 0.132,
+ 0.131511,
+ 0.132,
+ 0.131023,
+ 0.148134,
+ 0.131093,
+ 0.130534,
+ 0.133397,
+ 0.132,
+ 0.131022,
+ 0.131512,
+ 0.132489,
+ 0.131511,
+ 0.131511,
+ 0.130114,
+ 0.131092,
+ 0.132,
+ 0.131581,
+ 0.130534,
+ 0.130603
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 548987,
+ "unique_rows": 4008,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 65,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 65,
+ "zlib_blocks": 65,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 4008
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "a62056b83a73807b7f61d0d279d775f45e8ddf42470fec39f10ca2260297ab7a",
+ "table_sha256": "ed8808852ba87ad1d586536fdc3595716af0f270819ebd31949162460d023d0c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08318845299072564,
+ "representatives_sha256": "c45646621d3662f6bdd6db39b96d8782a760803e9976afc20c6d1e9619e7d6dc",
+ "structurally_merged_rows": 1325,
+ "backend_artifact_bytes": 550228,
+ "backend_with_key_registry_bytes": 621162,
+ "backend_with_datrie_key_index_bytes": 593284,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9899460254280686,
+ "mean_cosine": 0.9991811235787342,
+ "p05_cosine": 0.9948715931227619,
+ "relative_l2_mean": 0.023595053063924704,
+ "relative_l2_p95": 0.10137669486868771,
+ "relative_l2_max": 0.14188875915362323,
+ "weighted_relative_l2_mean": 0.013356548138067683,
+ "weighted_relative_frobenius": 0.009423937975726222,
+ "weighted_mean_cosine": 0.999684497516677,
+ "max_relative_norm_change": 0.011193257051772623
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900452817001179,
+ "mean_cosine": 0.9991922911278868,
+ "p05_cosine": 0.9948825321900114,
+ "relative_l2_mean": 0.018784309769533047,
+ "relative_l2_p95": 0.10116678712301178,
+ "relative_l2_max": 0.14080572574517558,
+ "weighted_relative_l2_mean": 0.007722098091802112,
+ "weighted_relative_frobenius": 0.006974880559954164,
+ "weighted_mean_cosine": 0.9997005177842563,
+ "max_relative_norm_change": 0.01006708671648114
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.365434,
+ "warm_p50_ms": 0.1308125,
+ "warm_p95_ms": 0.13349129999999998,
+ "warm_p99_ms": 0.14111948,
+ "samples_ms": [
+ 0.135283,
+ 0.132978,
+ 0.132908,
+ 0.131511,
+ 0.132978,
+ 0.132419,
+ 0.131511,
+ 0.129625,
+ 0.132489,
+ 0.130044,
+ 0.130603,
+ 0.130603,
+ 0.130603,
+ 0.131511,
+ 0.131092,
+ 0.130603,
+ 0.130115,
+ 0.130115,
+ 0.131581,
+ 0.130045,
+ 0.131022,
+ 0.130533,
+ 0.132,
+ 0.130603,
+ 0.130603,
+ 0.144851,
+ 0.130603,
+ 0.131022,
+ 0.130603,
+ 0.131022,
+ 0.131512,
+ 0.133397,
+ 0.131511,
+ 0.130114,
+ 0.130044,
+ 0.130044,
+ 0.132,
+ 0.130603,
+ 0.129206,
+ 0.129206
+ ],
+ "warm_rows_per_second": 3913999.0444338275
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 535945,
+ "unique_rows": 3905,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3905
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "fe9a13a24d7badd950819bc72cd81e968580858204fc855ec74d3b0643be5042",
+ "table_sha256": "e9f5c4701824eb9c555e8e1a617353fc3d02e59e8ebd1761a89695834c1a4012",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06120644696056843,
+ "representatives_sha256": "e34ed59fb3228c3c30a0bddf1d6d5bcaf09407ee494b8ba3e8d0fe58994e2db9",
+ "structurally_merged_rows": 1428,
+ "backend_artifact_bytes": 537186,
+ "backend_with_key_registry_bytes": 608120,
+ "backend_with_datrie_key_index_bytes": 580242,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9802774097437427,
+ "mean_cosine": 0.998929785427547,
+ "p05_cosine": 0.9937296122853513,
+ "relative_l2_mean": 0.026576276458257365,
+ "relative_l2_p95": 0.11231761955071509,
+ "relative_l2_max": 0.1978354117400544,
+ "weighted_relative_l2_mean": 0.014354728414426603,
+ "weighted_relative_frobenius": 0.010362495928818662,
+ "weighted_mean_cosine": 0.9995995000114486,
+ "max_relative_norm_change": 0.011193257051772623
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800731413504202,
+ "mean_cosine": 0.998940075547549,
+ "p05_cosine": 0.9937107115725136,
+ "relative_l2_mean": 0.021892984914000747,
+ "relative_l2_p95": 0.11232107990369113,
+ "relative_l2_max": 0.19888994022366926,
+ "weighted_relative_l2_mean": 0.008762672231207715,
+ "weighted_relative_frobenius": 0.008200859341694282,
+ "weighted_mean_cosine": 0.9996152353969843,
+ "max_relative_norm_change": 0.010083280214336047
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.281764,
+ "warm_p50_ms": 0.1315465,
+ "warm_p95_ms": 0.13599199999999997,
+ "warm_p99_ms": 0.16018965,
+ "samples_ms": [
+ 0.134864,
+ 0.132,
+ 0.131511,
+ 0.130603,
+ 0.148133,
+ 0.132,
+ 0.132,
+ 0.132,
+ 0.132001,
+ 0.132908,
+ 0.131022,
+ 0.130533,
+ 0.130044,
+ 0.130603,
+ 0.130673,
+ 0.132978,
+ 0.133466,
+ 0.132,
+ 0.132908,
+ 0.132908,
+ 0.130184,
+ 0.131092,
+ 0.130603,
+ 0.135353,
+ 0.132978,
+ 0.130603,
+ 0.132489,
+ 0.131093,
+ 0.130534,
+ 0.131511,
+ 0.132489,
+ 0.131511,
+ 0.130603,
+ 0.131092,
+ 0.167898,
+ 0.131511,
+ 0.131512,
+ 0.131581,
+ 0.131023,
+ 0.132
+ ],
+ "warm_rows_per_second": 3892159.8066083095
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 532479,
+ "unique_rows": 3877,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3877
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "279d44baa6b1e24848dbbae5502d1b2a3e190389e64a41ad0f8db09f357f8b85",
+ "table_sha256": "b86ef26f935a4f709ff8c2d5ced04e0822f3631bdd5a6c9c3b17e5fb2fbdd601",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06654365290887654,
+ "representatives_sha256": "7e670f5f6be473ee772c7b5444a1f4414f57b9366d57a9d047dd783aac94e016",
+ "structurally_merged_rows": 1456,
+ "backend_artifact_bytes": 533720,
+ "backend_with_key_registry_bytes": 604654,
+ "backend_with_datrie_key_index_bytes": 576776,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9708558855923936,
+ "mean_cosine": 0.9988065929090025,
+ "p05_cosine": 0.9932626568383942,
+ "relative_l2_mean": 0.027676963177214083,
+ "relative_l2_p95": 0.11607854092455877,
+ "relative_l2_max": 0.2411916582064447,
+ "weighted_relative_l2_mean": 0.014778498075646365,
+ "weighted_relative_frobenius": 0.011146495163277887,
+ "weighted_mean_cosine": 0.9995510349733676,
+ "max_relative_norm_change": 0.011193257051772623
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9707988177143576,
+ "mean_cosine": 0.9988167892682104,
+ "p05_cosine": 0.9932631200825917,
+ "relative_l2_mean": 0.02302821874678023,
+ "relative_l2_p95": 0.11619886723268123,
+ "relative_l2_max": 0.24136072450226767,
+ "weighted_relative_l2_mean": 0.009199261212029994,
+ "weighted_relative_frobenius": 0.009171906564283882,
+ "weighted_mean_cosine": 0.9995667272156469,
+ "max_relative_norm_change": 0.010083280214336047
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.068748,
+ "warm_p50_ms": 0.131023,
+ "warm_p95_ms": 0.13754865,
+ "warm_p99_ms": 0.16122598999999999,
+ "samples_ms": [
+ 0.133886,
+ 0.144781,
+ 0.133816,
+ 0.132419,
+ 0.130604,
+ 0.132489,
+ 0.132,
+ 0.131022,
+ 0.133397,
+ 0.130114,
+ 0.131022,
+ 0.129626,
+ 0.130604,
+ 0.131023,
+ 0.131023,
+ 0.131511,
+ 0.130603,
+ 0.132,
+ 0.132908,
+ 0.131023,
+ 0.131022,
+ 0.131092,
+ 0.130603,
+ 0.130184,
+ 0.130115,
+ 0.131511,
+ 0.131022,
+ 0.132489,
+ 0.135771,
+ 0.137168,
+ 0.13626,
+ 0.17174,
+ 0.133397,
+ 0.132419,
+ 0.130603,
+ 0.131022,
+ 0.129625,
+ 0.130114,
+ 0.131022,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3907710.8599253567
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 531853,
+ "unique_rows": 3872,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3872
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "63fba38a625fe4a0ddda6f97aff2446fd5aa274436d4a55d9249067c23b8dffd",
+ "table_sha256": "641fbc9933bb3432d857d5498e5d5e464fd2d90461f22178ba392b429427b7d1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.07724418584257364,
+ "representatives_sha256": "06ecfc12beba03cd91a9249a99a4a38025424fd923a12c9a7f29649ce2ab4d0b",
+ "structurally_merged_rows": 1461,
+ "backend_artifact_bytes": 533094,
+ "backend_with_key_registry_bytes": 604028,
+ "backend_with_datrie_key_index_bytes": 576150,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9645468363051497,
+ "mean_cosine": 0.9987760377805035,
+ "p05_cosine": 0.9931920311536553,
+ "relative_l2_mean": 0.027910005840439022,
+ "relative_l2_p95": 0.11689623701885093,
+ "relative_l2_max": 0.26511495929110795,
+ "weighted_relative_l2_mean": 0.014844824072176337,
+ "weighted_relative_frobenius": 0.011151053798317613,
+ "weighted_mean_cosine": 0.9995423387144133,
+ "max_relative_norm_change": 0.011193257051772623
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9645029199086544,
+ "mean_cosine": 0.9987861966537901,
+ "p05_cosine": 0.9932018647006915,
+ "relative_l2_mean": 0.0232673906702478,
+ "relative_l2_p95": 0.11682543464816218,
+ "relative_l2_max": 0.26537271472670454,
+ "weighted_relative_l2_mean": 0.009267331650062473,
+ "weighted_relative_frobenius": 0.009177450374893622,
+ "weighted_mean_cosine": 0.9995580202878689,
+ "max_relative_norm_change": 0.010083280214336047
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.081598,
+ "warm_p50_ms": 0.133956,
+ "warm_p95_ms": 0.15709494999999998,
+ "warm_p99_ms": 0.16217901999999998,
+ "samples_ms": [
+ 0.140521,
+ 0.135771,
+ 0.146667,
+ 0.138635,
+ 0.135771,
+ 0.134444,
+ 0.134864,
+ 0.153721,
+ 0.157493,
+ 0.135353,
+ 0.134305,
+ 0.132489,
+ 0.133397,
+ 0.134445,
+ 0.132419,
+ 0.132489,
+ 0.133816,
+ 0.134445,
+ 0.157074,
+ 0.13633,
+ 0.133956,
+ 0.138146,
+ 0.131023,
+ 0.13249,
+ 0.131511,
+ 0.133956,
+ 0.131023,
+ 0.132489,
+ 0.130044,
+ 0.137587,
+ 0.131511,
+ 0.131092,
+ 0.131581,
+ 0.131022,
+ 0.131092,
+ 0.132001,
+ 0.165175,
+ 0.13633,
+ 0.132489,
+ 0.132978
+ ],
+ "warm_rows_per_second": 3822150.5568992808
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 530590,
+ "unique_rows": 3862,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3862
+ },
+ "source_float32_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3",
+ "decoded_float32_sha256": "1dfd1c63f4e34ea36b814ccaaebc1d91a590a9ea1cd59b1186b8c098c20a2b5b",
+ "table_sha256": "983ab5dbd887857491f325d5904539399a4e1036df5c694c132f6d5f4e6dad90",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06423777109012008,
+ "representatives_sha256": "d58b671fc633ea2423d8b8262426ceb2cd1c7cdd442e0870782cfe892ae3c292",
+ "structurally_merged_rows": 1471,
+ "backend_artifact_bytes": 531831,
+ "backend_with_key_registry_bytes": 602765,
+ "backend_with_datrie_key_index_bytes": 574887,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9056773953279651,
+ "mean_cosine": 0.9986380550398255,
+ "p05_cosine": 0.992963940004652,
+ "relative_l2_mean": 0.02861425529997481,
+ "relative_l2_p95": 0.11860347312203266,
+ "relative_l2_max": 0.43300731849080465,
+ "weighted_relative_l2_mean": 0.015290162033414687,
+ "weighted_relative_frobenius": 0.013444069472239167,
+ "weighted_mean_cosine": 0.9994541553710345,
+ "max_relative_norm_change": 0.011193257051772623
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9059670556494595,
+ "mean_cosine": 0.998648271260613,
+ "p05_cosine": 0.9929359343583033,
+ "relative_l2_mean": 0.023984002914871804,
+ "relative_l2_p95": 0.11877170468980819,
+ "relative_l2_max": 0.4324839597442224,
+ "weighted_relative_l2_mean": 0.009719963455562435,
+ "weighted_relative_frobenius": 0.011857271754713355,
+ "weighted_mean_cosine": 0.9994698880712519,
+ "max_relative_norm_change": 0.010083280214336047
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.1909,
+ "warm_p50_ms": 0.13297750000000003,
+ "warm_p95_ms": 0.15240084999999998,
+ "warm_p99_ms": 0.17796869999999998,
+ "samples_ms": [
+ 0.137238,
+ 0.133397,
+ 0.139054,
+ 0.135842,
+ 0.134445,
+ 0.133816,
+ 0.132978,
+ 0.132978,
+ 0.132909,
+ 0.132978,
+ 0.151905,
+ 0.132908,
+ 0.188292,
+ 0.161822,
+ 0.132977,
+ 0.131022,
+ 0.132,
+ 0.132489,
+ 0.131511,
+ 0.134375,
+ 0.134305,
+ 0.133397,
+ 0.151416,
+ 0.136191,
+ 0.130044,
+ 0.132,
+ 0.131093,
+ 0.131023,
+ 0.131093,
+ 0.130045,
+ 0.131512,
+ 0.130603,
+ 0.133886,
+ 0.134375,
+ 0.133467,
+ 0.132489,
+ 0.130184,
+ 0.130604,
+ 0.130603,
+ 0.132489
+ ],
+ "warm_rows_per_second": 3850275.4225338865
+ }
+ }
+ ],
+ "baseline_quality": {},
+ "training_row_sha256": "90831a04809658c42b56b37046b1ad7b6035c1d55bb5c3ab2f777c5ca66145b3"
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 7,
+ "rows": 5333,
+ "dim": 128,
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "source_float32_bytes": 2730496,
+ "key_registry_bytes": 70934,
+ "unchanged_backbone_artifact_bytes": 0,
+ "alternative_datrie_key_index_bytes": 43056,
+ "query_rows": 512,
+ "query_ids_sha256": "87677d489e681e6c0a39b8460231cf990090e3318abb13c912a002c527931d01",
+ "frequency_trace_sha256": "0e7a36c623f2fc5cc1e4fd51bc27b7b98d60f0c1b31a56fb66253e951084f56a",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 512,
+ "storage": {
+ "artifact_bytes": 2730496,
+ "unique_rows": 5333,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "table_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.013780047884210944,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2731299,
+ "backend_with_key_registry_bytes": 2802233,
+ "backend_with_datrie_key_index_bytes": 2774355,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.40927,
+ "warm_p50_ms": 0.0619845,
+ "warm_p95_ms": 0.0725193,
+ "warm_p99_ms": 0.07776177,
+ "samples_ms": [
+ 0.064324,
+ 0.062019,
+ 0.062927,
+ 0.06153,
+ 0.06146,
+ 0.062019,
+ 0.062019,
+ 0.06195,
+ 0.060553,
+ 0.060483,
+ 0.061531,
+ 0.062438,
+ 0.060063,
+ 0.06146,
+ 0.06153,
+ 0.061949,
+ 0.062857,
+ 0.062438,
+ 0.062019,
+ 0.062508,
+ 0.061041,
+ 0.061879,
+ 0.06146,
+ 0.062508,
+ 0.06202,
+ 0.062019,
+ 0.060972,
+ 0.060622,
+ 0.062019,
+ 0.074311,
+ 0.079968,
+ 0.072425,
+ 0.062508,
+ 0.062019,
+ 0.060971,
+ 0.061879,
+ 0.060622,
+ 0.060971,
+ 0.06188,
+ 0.062019
+ ],
+ "warm_rows_per_second": 8260129.548516161
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 2607447,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 334,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 334,
+ "zlib_blocks": 303,
+ "xor_blocks": 31,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "table_sha256": "cad6c1e0a0b642f6988be6946fa27bddec2ec19f3bd843aea6e233b9f92e51c6",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.23365640896372497,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2608651,
+ "backend_with_key_registry_bytes": 2679585,
+ "backend_with_datrie_key_index_bytes": 2651707,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.506316,
+ "warm_p50_ms": 0.0704,
+ "warm_p95_ms": 0.07869759999999996,
+ "warm_p99_ms": 0.09331459,
+ "samples_ms": [
+ 0.075289,
+ 0.078083,
+ 0.075708,
+ 0.072844,
+ 0.090375,
+ 0.074312,
+ 0.073403,
+ 0.073823,
+ 0.095194,
+ 0.07473,
+ 0.072844,
+ 0.073822,
+ 0.068026,
+ 0.07005,
+ 0.068654,
+ 0.073822,
+ 0.070959,
+ 0.07047,
+ 0.0704,
+ 0.0704,
+ 0.068724,
+ 0.069074,
+ 0.07047,
+ 0.070539,
+ 0.07054,
+ 0.069073,
+ 0.069074,
+ 0.069632,
+ 0.070051,
+ 0.069912,
+ 0.069981,
+ 0.070051,
+ 0.069562,
+ 0.068514,
+ 0.069632,
+ 0.069493,
+ 0.069981,
+ 0.07005,
+ 0.068654,
+ 0.07047
+ ],
+ "warm_rows_per_second": 7272727.2727272725
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 2607447,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 334,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 334,
+ "zlib_blocks": 303,
+ "xor_blocks": 31,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "table_sha256": "cad6c1e0a0b642f6988be6946fa27bddec2ec19f3bd843aea6e233b9f92e51c6",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2612808980047703,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2608650,
+ "backend_with_key_registry_bytes": 2679584,
+ "backend_with_datrie_key_index_bytes": 2651706,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.512043,
+ "warm_p50_ms": 0.0705395,
+ "warm_p95_ms": 0.08014344999999996,
+ "warm_p99_ms": 0.0982451,
+ "samples_ms": [
+ 0.078991,
+ 0.07948,
+ 0.076616,
+ 0.092749,
+ 0.076615,
+ 0.074311,
+ 0.073822,
+ 0.073822,
+ 0.07473,
+ 0.073333,
+ 0.073333,
+ 0.069562,
+ 0.069073,
+ 0.070121,
+ 0.070051,
+ 0.070889,
+ 0.101759,
+ 0.071029,
+ 0.071028,
+ 0.07061,
+ 0.070469,
+ 0.069981,
+ 0.069143,
+ 0.071029,
+ 0.069562,
+ 0.069492,
+ 0.070889,
+ 0.069562,
+ 0.069073,
+ 0.070958,
+ 0.069981,
+ 0.070401,
+ 0.069073,
+ 0.069562,
+ 0.071378,
+ 0.069981,
+ 0.069562,
+ 0.070051,
+ 0.069981,
+ 0.069493
+ ],
+ "warm_rows_per_second": 7258344.615428234
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 2607447,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 334,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 334,
+ "zlib_blocks": 303,
+ "xor_blocks": 31,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "table_sha256": "cad6c1e0a0b642f6988be6946fa27bddec2ec19f3bd843aea6e233b9f92e51c6",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2417168691754341,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2608650,
+ "backend_with_key_registry_bytes": 2679584,
+ "backend_with_datrie_key_index_bytes": 2651706,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.496957,
+ "warm_p50_ms": 0.0702605,
+ "warm_p95_ms": 0.07680025,
+ "warm_p99_ms": 0.09978062999999998,
+ "samples_ms": [
+ 0.07899,
+ 0.076685,
+ 0.072496,
+ 0.071447,
+ 0.072356,
+ 0.072914,
+ 0.072426,
+ 0.071028,
+ 0.113073,
+ 0.07054,
+ 0.071936,
+ 0.071937,
+ 0.071029,
+ 0.069981,
+ 0.070889,
+ 0.071378,
+ 0.070889,
+ 0.069562,
+ 0.069632,
+ 0.069562,
+ 0.07047,
+ 0.069074,
+ 0.070889,
+ 0.070958,
+ 0.069143,
+ 0.069981,
+ 0.069073,
+ 0.069143,
+ 0.069073,
+ 0.069981,
+ 0.069981,
+ 0.07012,
+ 0.070121,
+ 0.069981,
+ 0.070051,
+ 0.069143,
+ 0.069074,
+ 0.0704,
+ 0.069981,
+ 0.070051
+ ],
+ "warm_rows_per_second": 7287167.042648429
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1956893,
+ "unique_rows": 3975,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 249,
+ "zlib_blocks": 249,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3975
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "bcbff57cdbbcae9108a7f5ce95eacb8f52057ba4a29da04bbeaac79c1304fc66",
+ "table_sha256": "6590a2eea087ca111e3a502e1d26ea0d6b2928cfdd2bea9973afe361499d8407",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.18209400493651628,
+ "representatives_sha256": "6a868016e073c8843b96d23a54b5040cd43bd7de8fd10a4ef0fd77fb6d5356f3",
+ "structurally_merged_rows": 1358,
+ "backend_artifact_bytes": 1958095,
+ "backend_with_key_registry_bytes": 2029029,
+ "backend_with_datrie_key_index_bytes": 2001151,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900498274896388,
+ "mean_cosine": 0.9991649326466401,
+ "p05_cosine": 0.9946380657924886,
+ "relative_l2_mean": 0.01921774489139315,
+ "relative_l2_p95": 0.10370633425993789,
+ "relative_l2_max": 0.14087181586808722,
+ "weighted_relative_l2_mean": 0.007669195302930095,
+ "weighted_relative_frobenius": 0.006743679010149183,
+ "weighted_mean_cosine": 0.9997009272174809,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900498274896388,
+ "mean_cosine": 0.9991649326466401,
+ "p05_cosine": 0.9946380657924886,
+ "relative_l2_mean": 0.01921774489139315,
+ "relative_l2_p95": 0.10370633425993789,
+ "relative_l2_max": 0.14087181586808722,
+ "weighted_relative_l2_mean": 0.007669195302930095,
+ "weighted_relative_frobenius": 0.006743679010149183,
+ "weighted_mean_cosine": 0.9997009272174809,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.219559,
+ "warm_p50_ms": 0.070051,
+ "warm_p95_ms": 0.0744507,
+ "warm_p99_ms": 0.08089676,
+ "samples_ms": [
+ 0.077105,
+ 0.074242,
+ 0.074311,
+ 0.074311,
+ 0.072845,
+ 0.073334,
+ 0.071867,
+ 0.071867,
+ 0.072355,
+ 0.070051,
+ 0.07005,
+ 0.069562,
+ 0.069492,
+ 0.083321,
+ 0.070958,
+ 0.069981,
+ 0.070958,
+ 0.069911,
+ 0.069562,
+ 0.070121,
+ 0.069981,
+ 0.074311,
+ 0.069632,
+ 0.069003,
+ 0.070051,
+ 0.070051,
+ 0.07047,
+ 0.069492,
+ 0.069073,
+ 0.069562,
+ 0.069492,
+ 0.069073,
+ 0.070609,
+ 0.069632,
+ 0.069003,
+ 0.069142,
+ 0.069562,
+ 0.07047,
+ 0.069073,
+ 0.071518
+ ],
+ "warm_rows_per_second": 7308960.614409502
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1915441,
+ "unique_rows": 3889,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 244,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 243,
+ "zlib_blocks": 243,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3889
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "99a85de2ff15a843787c443e90b9c1dd25d61a0acd43c2878fa12520da63e324",
+ "table_sha256": "c479ac14cfa94eedb8c7789490d50097e48391f3f7556aedfe8b721a8677549b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1829489329829812,
+ "representatives_sha256": "ffe6744fb8805a4bf8e5b3fd5ab7e84ef8c0d556dc0063e801ba1ec429934754",
+ "structurally_merged_rows": 1444,
+ "backend_artifact_bytes": 1916643,
+ "backend_with_key_registry_bytes": 1987577,
+ "backend_with_datrie_key_index_bytes": 1959699,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800443686538259,
+ "mean_cosine": 0.9989493697934625,
+ "p05_cosine": 0.9936464642843283,
+ "relative_l2_mean": 0.021843835841248167,
+ "relative_l2_p95": 0.11269904523450791,
+ "relative_l2_max": 0.19941118699057925,
+ "weighted_relative_l2_mean": 0.008671316985251821,
+ "weighted_relative_frobenius": 0.009987027162445301,
+ "weighted_mean_cosine": 0.9996186797240005,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800443686538259,
+ "mean_cosine": 0.9989493697934625,
+ "p05_cosine": 0.9936464642843283,
+ "relative_l2_mean": 0.021843835841248167,
+ "relative_l2_p95": 0.11269904523450791,
+ "relative_l2_max": 0.19941118699057925,
+ "weighted_relative_l2_mean": 0.008671316985251821,
+ "weighted_relative_frobenius": 0.009987027162445301,
+ "weighted_mean_cosine": 0.9996186797240005,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.255946,
+ "warm_p50_ms": 0.07043550000000001,
+ "warm_p95_ms": 0.07389535,
+ "warm_p99_ms": 0.09339562999999998,
+ "samples_ms": [
+ 0.075289,
+ 0.073822,
+ 0.073822,
+ 0.072774,
+ 0.072356,
+ 0.072915,
+ 0.070959,
+ 0.073333,
+ 0.070121,
+ 0.069981,
+ 0.070959,
+ 0.104972,
+ 0.071867,
+ 0.070959,
+ 0.07054,
+ 0.069981,
+ 0.07054,
+ 0.069073,
+ 0.070889,
+ 0.07047,
+ 0.069981,
+ 0.069562,
+ 0.07054,
+ 0.069981,
+ 0.07047,
+ 0.069143,
+ 0.069981,
+ 0.07047,
+ 0.070051,
+ 0.069632,
+ 0.069073,
+ 0.070401,
+ 0.069492,
+ 0.069632,
+ 0.069562,
+ 0.070889,
+ 0.069492,
+ 0.070051,
+ 0.069562,
+ 0.069492
+ ],
+ "warm_rows_per_second": 7269061.765728928
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1909992,
+ "unique_rows": 3878,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 243,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 243,
+ "zlib_blocks": 243,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3878
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "565c2bc329ce53a0df8e37745f7a91eee8511cd55df1cd2ce4dcfe98fa366bb8",
+ "table_sha256": "ed3305384acbdfe75476c9295724ebd61bae347ac82710479133bcde6b51108f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1924193510785699,
+ "representatives_sha256": "7d436e33228c3b99bbd1f3685fbf66aa455c3e4abf4ca1c4abd4453aabfcd630",
+ "structurally_merged_rows": 1455,
+ "backend_artifact_bytes": 1911194,
+ "backend_with_key_registry_bytes": 1982128,
+ "backend_with_datrie_key_index_bytes": 1954250,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.970372584563533,
+ "mean_cosine": 0.9988918051069324,
+ "p05_cosine": 0.9934823025035637,
+ "relative_l2_mean": 0.02234575232524709,
+ "relative_l2_p95": 0.11406973290377129,
+ "relative_l2_max": 0.24459732928838826,
+ "weighted_relative_l2_mean": 0.00889438856026394,
+ "weighted_relative_frobenius": 0.010374481889143803,
+ "weighted_mean_cosine": 0.9995919364799863,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.970372584563533,
+ "mean_cosine": 0.9988918051069324,
+ "p05_cosine": 0.9934823025035637,
+ "relative_l2_mean": 0.02234575232524709,
+ "relative_l2_p95": 0.11406973290377129,
+ "relative_l2_max": 0.24459732928838826,
+ "weighted_relative_l2_mean": 0.00889438856026394,
+ "weighted_relative_frobenius": 0.010374481889143803,
+ "weighted_mean_cosine": 0.9995919364799863,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.123876,
+ "warm_p50_ms": 0.069073,
+ "warm_p95_ms": 0.07396264999999999,
+ "warm_p99_ms": 0.07985388,
+ "samples_ms": [
+ 0.076616,
+ 0.073823,
+ 0.072845,
+ 0.073753,
+ 0.071867,
+ 0.071937,
+ 0.073822,
+ 0.073333,
+ 0.071936,
+ 0.072844,
+ 0.069003,
+ 0.068165,
+ 0.069073,
+ 0.068584,
+ 0.071029,
+ 0.069562,
+ 0.067606,
+ 0.069143,
+ 0.069004,
+ 0.069492,
+ 0.069073,
+ 0.069562,
+ 0.069003,
+ 0.069073,
+ 0.068654,
+ 0.069003,
+ 0.069073,
+ 0.069073,
+ 0.081924,
+ 0.070051,
+ 0.069003,
+ 0.069142,
+ 0.067676,
+ 0.068095,
+ 0.068654,
+ 0.068514,
+ 0.069631,
+ 0.068095,
+ 0.069003,
+ 0.068165
+ ],
+ "warm_rows_per_second": 7412447.700259147
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1907108,
+ "unique_rows": 3872,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 242,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 242,
+ "zlib_blocks": 242,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3872
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "5f3897021cb5b4d01c49b4f6087e8c83906d531045def14546ddcb40820c166e",
+ "table_sha256": "f5d2240c19cd8b101915061ddee216f591a8dc046e02251627f8be9d4d5dc5ec",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.178886191919446,
+ "representatives_sha256": "991fd3955e7ce0366e0b95960da44d05f5d2571d45f270fef67728c8bf6c01aa",
+ "structurally_merged_rows": 1461,
+ "backend_artifact_bytes": 1908310,
+ "backend_with_key_registry_bytes": 1979244,
+ "backend_with_datrie_key_index_bytes": 1951366,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9539393104479431,
+ "mean_cosine": 0.9988509765773617,
+ "p05_cosine": 0.9933682216614086,
+ "relative_l2_mean": 0.022647303085639288,
+ "relative_l2_p95": 0.1151708920097435,
+ "relative_l2_max": 0.3049388897587284,
+ "weighted_relative_l2_mean": 0.00901257678779398,
+ "weighted_relative_frobenius": 0.010616035362168956,
+ "weighted_mean_cosine": 0.9995754077438983,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9539393104479431,
+ "mean_cosine": 0.9988509765773617,
+ "p05_cosine": 0.9933682216614086,
+ "relative_l2_mean": 0.022647303085639288,
+ "relative_l2_p95": 0.1151708920097435,
+ "relative_l2_max": 0.3049388897587284,
+ "weighted_relative_l2_mean": 0.00901257678779398,
+ "weighted_relative_frobenius": 0.010616035362168956,
+ "weighted_mean_cosine": 0.9995754077438983,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.111654,
+ "warm_p50_ms": 0.07050500000000001,
+ "warm_p95_ms": 0.07761844999999999,
+ "warm_p99_ms": 0.08153316,
+ "samples_ms": [
+ 0.078083,
+ 0.073823,
+ 0.072845,
+ 0.072845,
+ 0.071937,
+ 0.072355,
+ 0.073333,
+ 0.073822,
+ 0.072844,
+ 0.073333,
+ 0.073823,
+ 0.072426,
+ 0.077594,
+ 0.076197,
+ 0.075219,
+ 0.070051,
+ 0.069004,
+ 0.07047,
+ 0.071028,
+ 0.069073,
+ 0.069981,
+ 0.07047,
+ 0.069562,
+ 0.07012,
+ 0.07054,
+ 0.07047,
+ 0.069003,
+ 0.069143,
+ 0.069003,
+ 0.069981,
+ 0.069562,
+ 0.07005,
+ 0.069492,
+ 0.083739,
+ 0.07054,
+ 0.070958,
+ 0.069003,
+ 0.069073,
+ 0.069631,
+ 0.068514
+ ],
+ "warm_rows_per_second": 7261896.319409969
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1903248,
+ "unique_rows": 3864,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 242,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 242,
+ "zlib_blocks": 242,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3864
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "538675e4df80dda66f752b18b57af79ef1484be9de004fc46fc284e5907d7574",
+ "table_sha256": "a0b617de45f4577f4f307ecc63f901bebc97d16c01fd11ec7493f733d878e4a0",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1804608351085335,
+ "representatives_sha256": "dfad668d84d3a9dde10fed2c9e5964b59d2c5eaccc988a384af8a2c7bc4af52c",
+ "structurally_merged_rows": 1469,
+ "backend_artifact_bytes": 1904450,
+ "backend_with_key_registry_bytes": 1975384,
+ "backend_with_datrie_key_index_bytes": 1947506,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9037877545708831,
+ "mean_cosine": 0.9987401100496023,
+ "p05_cosine": 0.9932851026858404,
+ "relative_l2_mean": 0.02322164997411739,
+ "relative_l2_p95": 0.1161852959435186,
+ "relative_l2_max": 0.4386936628207428,
+ "weighted_relative_l2_mean": 0.009396651320714072,
+ "weighted_relative_frobenius": 0.014093234342090312,
+ "weighted_mean_cosine": 0.9995019785254639,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9037877545708831,
+ "mean_cosine": 0.9987401100496023,
+ "p05_cosine": 0.9932851026858404,
+ "relative_l2_mean": 0.02322164997411739,
+ "relative_l2_p95": 0.1161852959435186,
+ "relative_l2_max": 0.4386936628207428,
+ "weighted_relative_l2_mean": 0.009396651320714072,
+ "weighted_relative_frobenius": 0.014093234342090312,
+ "weighted_mean_cosine": 0.9995019785254639,
+ "max_relative_norm_change": 0.010048403769123795
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.248753,
+ "warm_p50_ms": 0.0700505,
+ "warm_p95_ms": 0.07893404999999998,
+ "warm_p99_ms": 0.09836647999999999,
+ "samples_ms": [
+ 0.078501,
+ 0.073822,
+ 0.10553,
+ 0.075638,
+ 0.072845,
+ 0.072425,
+ 0.071936,
+ 0.087162,
+ 0.0748,
+ 0.073752,
+ 0.070401,
+ 0.074311,
+ 0.073333,
+ 0.071937,
+ 0.070121,
+ 0.069981,
+ 0.070051,
+ 0.070051,
+ 0.07012,
+ 0.069562,
+ 0.069982,
+ 0.069981,
+ 0.070051,
+ 0.069562,
+ 0.069492,
+ 0.069562,
+ 0.069981,
+ 0.069562,
+ 0.069073,
+ 0.069632,
+ 0.068165,
+ 0.069562,
+ 0.07005,
+ 0.068654,
+ 0.069981,
+ 0.068095,
+ 0.069632,
+ 0.069981,
+ 0.070051,
+ 0.069632
+ ],
+ "warm_rows_per_second": 7309012.783634664
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 1365248,
+ "unique_rows": 5333,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "590fdfdcc048589926d117e6f06ff1ebd64a0fd2815cc57b8de70080aab734da",
+ "table_sha256": "4d5e89355e425d36118ebf293bc515da98012d2b8cfb74d7abef5c01f93e5d8b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.01722392486408353,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1366052,
+ "backend_with_key_registry_bytes": 1436986,
+ "backend_with_datrie_key_index_bytes": 1409108,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999643352301,
+ "mean_cosine": 0.9999999789923787,
+ "p05_cosine": 0.9999999731100682,
+ "relative_l2_mean": 0.00020673286727305166,
+ "relative_l2_p95": 0.00023477543729639956,
+ "relative_l2_max": 0.0002705172201136542,
+ "weighted_relative_l2_mean": 0.00020763766850666958,
+ "weighted_relative_frobenius": 0.0002090819556359097,
+ "weighted_mean_cosine": 0.9999999788037498,
+ "max_relative_norm_change": 0.00012555925984645075
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.45327,
+ "warm_p50_ms": 0.22,
+ "warm_p95_ms": 0.22364954999999995,
+ "warm_p99_ms": 0.24744582,
+ "samples_ms": [
+ 0.222794,
+ 0.220978,
+ 0.219511,
+ 0.220489,
+ 0.220978,
+ 0.219512,
+ 0.220909,
+ 0.220978,
+ 0.22,
+ 0.220908,
+ 0.219442,
+ 0.239905,
+ 0.222375,
+ 0.220978,
+ 0.221467,
+ 0.220908,
+ 0.218603,
+ 0.219162,
+ 0.219581,
+ 0.219581,
+ 0.219441,
+ 0.218533,
+ 0.22,
+ 0.22,
+ 0.220489,
+ 0.219023,
+ 0.22049,
+ 0.22,
+ 0.220489,
+ 0.252267,
+ 0.222794,
+ 0.219581,
+ 0.218604,
+ 0.219581,
+ 0.219581,
+ 0.22,
+ 0.219441,
+ 0.21993,
+ 0.220489,
+ 0.220489
+ ],
+ "warm_rows_per_second": 2327272.727272727
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1329267,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 167,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 167,
+ "zlib_blocks": 145,
+ "xor_blocks": 22,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "590fdfdcc048589926d117e6f06ff1ebd64a0fd2815cc57b8de70080aab734da",
+ "table_sha256": "c27d497e5ec13867fc6a5cb4eab31b628eed9707d4617ebfe091170fdca264b0",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1538615459576249,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1330472,
+ "backend_with_key_registry_bytes": 1401406,
+ "backend_with_datrie_key_index_bytes": 1373528,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999643352301,
+ "mean_cosine": 0.9999999789923787,
+ "p05_cosine": 0.9999999731100682,
+ "relative_l2_mean": 0.00020673286727305166,
+ "relative_l2_p95": 0.00023477543729639956,
+ "relative_l2_max": 0.0002705172201136542,
+ "weighted_relative_l2_mean": 0.00020763766850666958,
+ "weighted_relative_frobenius": 0.0002090819556359097,
+ "weighted_mean_cosine": 0.9999999788037498,
+ "max_relative_norm_change": 0.00012555925984645075
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.113342,
+ "warm_p50_ms": 0.22901,
+ "warm_p95_ms": 0.23362624999999995,
+ "warm_p99_ms": 0.26109463,
+ "samples_ms": [
+ 0.231315,
+ 0.230896,
+ 0.229918,
+ 0.229917,
+ 0.229848,
+ 0.22894,
+ 0.229499,
+ 0.269239,
+ 0.231804,
+ 0.228521,
+ 0.229499,
+ 0.22901,
+ 0.232851,
+ 0.230895,
+ 0.229498,
+ 0.22901,
+ 0.22894,
+ 0.229499,
+ 0.229009,
+ 0.227962,
+ 0.228102,
+ 0.227543,
+ 0.227124,
+ 0.227054,
+ 0.248356,
+ 0.229918,
+ 0.228032,
+ 0.226635,
+ 0.231804,
+ 0.231385,
+ 0.229988,
+ 0.229987,
+ 0.227194,
+ 0.227124,
+ 0.228521,
+ 0.228521,
+ 0.227054,
+ 0.228032,
+ 0.227543,
+ 0.227543
+ ],
+ "warm_rows_per_second": 2235710.2309942795
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1329267,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 167,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 167,
+ "zlib_blocks": 145,
+ "xor_blocks": 22,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "590fdfdcc048589926d117e6f06ff1ebd64a0fd2815cc57b8de70080aab734da",
+ "table_sha256": "c27d497e5ec13867fc6a5cb4eab31b628eed9707d4617ebfe091170fdca264b0",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14749816106632352,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1330471,
+ "backend_with_key_registry_bytes": 1401405,
+ "backend_with_datrie_key_index_bytes": 1373527,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999643352301,
+ "mean_cosine": 0.9999999789923787,
+ "p05_cosine": 0.9999999731100682,
+ "relative_l2_mean": 0.00020673286727305166,
+ "relative_l2_p95": 0.00023477543729639956,
+ "relative_l2_max": 0.0002705172201136542,
+ "weighted_relative_l2_mean": 0.00020763766850666958,
+ "weighted_relative_frobenius": 0.0002090819556359097,
+ "weighted_mean_cosine": 0.9999999788037498,
+ "max_relative_norm_change": 0.00012555925984645075
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.123748,
+ "warm_p50_ms": 0.22859049999999997,
+ "warm_p95_ms": 0.23499894999999996,
+ "warm_p99_ms": 0.25888589,
+ "samples_ms": [
+ 0.232362,
+ 0.231803,
+ 0.230826,
+ 0.230895,
+ 0.228591,
+ 0.22901,
+ 0.22901,
+ 0.229499,
+ 0.22894,
+ 0.247937,
+ 0.230477,
+ 0.22859,
+ 0.22901,
+ 0.229498,
+ 0.227543,
+ 0.227124,
+ 0.226636,
+ 0.228451,
+ 0.227962,
+ 0.227543,
+ 0.22901,
+ 0.227124,
+ 0.228521,
+ 0.228032,
+ 0.228521,
+ 0.228032,
+ 0.265886,
+ 0.230896,
+ 0.228032,
+ 0.234318,
+ 0.227543,
+ 0.227962,
+ 0.227473,
+ 0.227613,
+ 0.227544,
+ 0.227962,
+ 0.230336,
+ 0.230407,
+ 0.228591,
+ 0.228102
+ ],
+ "warm_rows_per_second": 2239813.115593168
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1329267,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 167,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 167,
+ "zlib_blocks": 145,
+ "xor_blocks": 22,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "590fdfdcc048589926d117e6f06ff1ebd64a0fd2815cc57b8de70080aab734da",
+ "table_sha256": "c27d497e5ec13867fc6a5cb4eab31b628eed9707d4617ebfe091170fdca264b0",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13302542688325047,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1330471,
+ "backend_with_key_registry_bytes": 1401405,
+ "backend_with_datrie_key_index_bytes": 1373527,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999643352301,
+ "mean_cosine": 0.9999999789923787,
+ "p05_cosine": 0.9999999731100682,
+ "relative_l2_mean": 0.00020673286727305166,
+ "relative_l2_p95": 0.00023477543729639956,
+ "relative_l2_max": 0.0002705172201136542,
+ "weighted_relative_l2_mean": 0.00020763766850666958,
+ "weighted_relative_frobenius": 0.0002090819556359097,
+ "weighted_mean_cosine": 0.9999999788037498,
+ "max_relative_norm_change": 0.00012555925984645075
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.130802,
+ "warm_p50_ms": 0.2289745,
+ "warm_p95_ms": 0.23567629999999998,
+ "warm_p99_ms": 0.24956665,
+ "samples_ms": [
+ 0.235156,
+ 0.229499,
+ 0.228032,
+ 0.227543,
+ 0.228381,
+ 0.230896,
+ 0.235086,
+ 0.22894,
+ 0.245562,
+ 0.22908,
+ 0.234178,
+ 0.228451,
+ 0.227473,
+ 0.227963,
+ 0.228032,
+ 0.227613,
+ 0.229499,
+ 0.229429,
+ 0.22901,
+ 0.233759,
+ 0.228521,
+ 0.228032,
+ 0.229009,
+ 0.233759,
+ 0.22894,
+ 0.252127,
+ 0.231315,
+ 0.228591,
+ 0.229009,
+ 0.22894,
+ 0.229009,
+ 0.228591,
+ 0.229079,
+ 0.226705,
+ 0.228521,
+ 0.22852,
+ 0.227543,
+ 0.227963,
+ 0.229848,
+ 0.229009
+ ],
+ "warm_rows_per_second": 2236056.8534924197
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1003858,
+ "unique_rows": 3974,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 125,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 125,
+ "zlib_blocks": 125,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3974
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "0af80157f7cec6abae5c8ce14fab160a50fc8c5c8f53170ca34bd5ea8a5abfea",
+ "table_sha256": "93f2ac65cdc2b3db24833dae2c21c216a6da481f1034c89cf04c84a27848c36b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.11185197299346328,
+ "representatives_sha256": "9205e2842ffd9234a133ac92f8a1de72edcec80d72ef19103846162fb17ba50f",
+ "structurally_merged_rows": 1359,
+ "backend_artifact_bytes": 1005060,
+ "backend_with_key_registry_bytes": 1075994,
+ "backend_with_datrie_key_index_bytes": 1048116,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900522844601786,
+ "mean_cosine": 0.9991640375629814,
+ "p05_cosine": 0.994637472144233,
+ "relative_l2_mean": 0.019390212010629005,
+ "relative_l2_p95": 0.10369440961747017,
+ "relative_l2_max": 0.1408746714806218,
+ "weighted_relative_l2_mean": 0.007857378964608506,
+ "weighted_relative_frobenius": 0.006748195908759052,
+ "weighted_mean_cosine": 0.9997006623119686,
+ "max_relative_norm_change": 0.010017310244631894
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900483856073877,
+ "mean_cosine": 0.9991640300187963,
+ "p05_cosine": 0.9946371997686602,
+ "relative_l2_mean": 0.01923620903668173,
+ "relative_l2_p95": 0.10369892786180715,
+ "relative_l2_max": 0.14089291500468373,
+ "weighted_relative_l2_mean": 0.0076744640102752925,
+ "weighted_relative_frobenius": 0.0067452159501694255,
+ "weighted_mean_cosine": 0.9997006691776221,
+ "max_relative_norm_change": 0.01002151577688962
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.949505,
+ "warm_p50_ms": 0.2286955,
+ "warm_p95_ms": 0.23725484999999996,
+ "warm_p99_ms": 0.25157525999999997,
+ "samples_ms": [
+ 0.236972,
+ 0.235575,
+ 0.232293,
+ 0.229429,
+ 0.230406,
+ 0.229429,
+ 0.228451,
+ 0.22894,
+ 0.22852,
+ 0.228521,
+ 0.230406,
+ 0.22887,
+ 0.229429,
+ 0.228521,
+ 0.227054,
+ 0.242629,
+ 0.228521,
+ 0.228451,
+ 0.22901,
+ 0.227124,
+ 0.228102,
+ 0.228521,
+ 0.228521,
+ 0.227124,
+ 0.227194,
+ 0.229498,
+ 0.229918,
+ 0.234737,
+ 0.231314,
+ 0.231384,
+ 0.229918,
+ 0.228451,
+ 0.257295,
+ 0.231803,
+ 0.228102,
+ 0.228032,
+ 0.227543,
+ 0.228381,
+ 0.227543,
+ 0.227543
+ ],
+ "warm_rows_per_second": 2238784.759647654
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 982914,
+ "unique_rows": 3888,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 122,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 122,
+ "zlib_blocks": 122,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3888
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "57bc51a9a2d3fcd6169ebf714210f12d448cd80c00b891e9a028edf2c76553ef",
+ "table_sha256": "489d1abf7dd67e6d5e892f3b492148f767fa16b42cc9fa234b5e6d130fa1f2bc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10486064408905804,
+ "representatives_sha256": "277f019233b6f3bcb6249b67edee96960c8bd860780e6d0f441e0b2f195a85e8",
+ "structurally_merged_rows": 1445,
+ "backend_artifact_bytes": 984115,
+ "backend_with_key_registry_bytes": 1055049,
+ "backend_with_datrie_key_index_bytes": 1027171,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800424596233465,
+ "mean_cosine": 0.9989484689878441,
+ "p05_cosine": 0.9936446007266101,
+ "relative_l2_mean": 0.022013031924056462,
+ "relative_l2_p95": 0.11268591633659679,
+ "relative_l2_max": 0.1994210431944021,
+ "weighted_relative_l2_mean": 0.008858269765568291,
+ "weighted_relative_frobenius": 0.0099899658176846,
+ "weighted_mean_cosine": 0.9996184151248343,
+ "max_relative_norm_change": 0.010017310244631894
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800416084647343,
+ "mean_cosine": 0.9989484649095248,
+ "p05_cosine": 0.9936429216293082,
+ "relative_l2_mean": 0.0218623178175955,
+ "relative_l2_p95": 0.11267980114319776,
+ "relative_l2_max": 0.19942483799749702,
+ "weighted_relative_l2_mean": 0.008676578811208108,
+ "weighted_relative_frobenius": 0.009987917577146297,
+ "weighted_mean_cosine": 0.9996184239003195,
+ "max_relative_norm_change": 0.01002151577688962
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.938051,
+ "warm_p50_ms": 0.22900949999999998,
+ "warm_p95_ms": 0.23153759999999995,
+ "warm_p99_ms": 0.25725179,
+ "samples_ms": [
+ 0.230895,
+ 0.22859,
+ 0.22901,
+ 0.229009,
+ 0.230895,
+ 0.229918,
+ 0.229918,
+ 0.229918,
+ 0.22901,
+ 0.229987,
+ 0.22901,
+ 0.229498,
+ 0.228521,
+ 0.229499,
+ 0.227613,
+ 0.243747,
+ 0.229918,
+ 0.229009,
+ 0.226635,
+ 0.227194,
+ 0.22852,
+ 0.229499,
+ 0.229009,
+ 0.228102,
+ 0.227543,
+ 0.227613,
+ 0.228032,
+ 0.22901,
+ 0.226984,
+ 0.228451,
+ 0.229429,
+ 0.229009,
+ 0.227962,
+ 0.265886,
+ 0.230406,
+ 0.229498,
+ 0.22894,
+ 0.229429,
+ 0.228032,
+ 0.22852
+ ],
+ "warm_rows_per_second": 2235715.1122551686
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 980262,
+ "unique_rows": 3877,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 122,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 122,
+ "zlib_blocks": 122,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3877
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "612c9ba3bc97d5c29f8e2f370f231068dc671aa45b7a3395fb771afe2fdee4f4",
+ "table_sha256": "8eeec69636b33454a26b28c61ed60da3875531986a2afca64bf707960cd85d00",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10655394801869988,
+ "representatives_sha256": "4bbd7d01ad58d7ae302ef3ed98e470624383f7b5a9f3f268f8ddeb9609451463",
+ "structurally_merged_rows": 1456,
+ "backend_artifact_bytes": 981463,
+ "backend_with_key_registry_bytes": 1052397,
+ "backend_with_datrie_key_index_bytes": 1024519,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.970374756828758,
+ "mean_cosine": 0.9988909074575562,
+ "p05_cosine": 0.9934819196205635,
+ "relative_l2_mean": 0.02251451345037744,
+ "relative_l2_p95": 0.11408063305368717,
+ "relative_l2_max": 0.24458470025994816,
+ "weighted_relative_l2_mean": 0.00908119703018639,
+ "weighted_relative_frobenius": 0.010377290834367692,
+ "weighted_mean_cosine": 0.9995916730687094,
+ "max_relative_norm_change": 0.010017310244631894
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9703788044938143,
+ "mean_cosine": 0.9988909009032235,
+ "p05_cosine": 0.9934813324528675,
+ "relative_l2_mean": 0.022364235080523234,
+ "relative_l2_p95": 0.1140800857828706,
+ "relative_l2_max": 0.2445684164990833,
+ "weighted_relative_l2_mean": 0.008899661323857288,
+ "weighted_relative_frobenius": 0.010375380765568555,
+ "weighted_mean_cosine": 0.9995916793899848,
+ "max_relative_norm_change": 0.01002151577688962
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.284953,
+ "warm_p50_ms": 0.22897499999999998,
+ "warm_p95_ms": 0.23607435,
+ "warm_p99_ms": 0.25046152,
+ "samples_ms": [
+ 0.231315,
+ 0.22894,
+ 0.229918,
+ 0.228521,
+ 0.229918,
+ 0.235575,
+ 0.233759,
+ 0.230895,
+ 0.22901,
+ 0.233759,
+ 0.245562,
+ 0.230895,
+ 0.229429,
+ 0.228451,
+ 0.22901,
+ 0.232782,
+ 0.23327,
+ 0.231803,
+ 0.230477,
+ 0.228591,
+ 0.22901,
+ 0.228521,
+ 0.228102,
+ 0.229079,
+ 0.226705,
+ 0.227613,
+ 0.227124,
+ 0.253594,
+ 0.233201,
+ 0.228521,
+ 0.228032,
+ 0.227613,
+ 0.227613,
+ 0.228102,
+ 0.228032,
+ 0.226635,
+ 0.22859,
+ 0.227613,
+ 0.227054,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2236051.970739164
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 978672,
+ "unique_rows": 3871,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 121,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 121,
+ "zlib_blocks": 121,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3871
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "e1c4e492180d3f8aadf7964d5d73b2ccab7d0f5b43af055766c9334f87af0542",
+ "table_sha256": "636ed25eab773bd9a5f38c6b1f00298c8181cc2fa02545b77ed4ada9328cf84f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10646545910276473,
+ "representatives_sha256": "212263313d6ae3a91d10ddbf617f814b1c08438e03dc93b0fe181bfa7a5fa9f2",
+ "structurally_merged_rows": 1462,
+ "backend_artifact_bytes": 979873,
+ "backend_with_key_registry_bytes": 1050807,
+ "backend_with_datrie_key_index_bytes": 1022929,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9539257055513554,
+ "mean_cosine": 0.9988500772716368,
+ "p05_cosine": 0.9933672096666714,
+ "relative_l2_mean": 0.022815834585993885,
+ "relative_l2_p95": 0.11517078666926303,
+ "relative_l2_max": 0.30498064308212897,
+ "weighted_relative_l2_mean": 0.00919929728101851,
+ "weighted_relative_frobenius": 0.010618802479643075,
+ "weighted_mean_cosine": 0.9995751438592744,
+ "max_relative_norm_change": 0.010017310244631894
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9539276501271642,
+ "mean_cosine": 0.9988500722968754,
+ "p05_cosine": 0.9933665941569161,
+ "relative_l2_mean": 0.022665786508838028,
+ "relative_l2_p95": 0.11516667455361981,
+ "relative_l2_max": 0.3049780712056362,
+ "weighted_relative_l2_mean": 0.009017849454424659,
+ "weighted_relative_frobenius": 0.010616933019063798,
+ "weighted_mean_cosine": 0.9995751505526811,
+ "max_relative_norm_change": 0.01002151577688962
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.188852,
+ "warm_p50_ms": 0.228521,
+ "warm_p95_ms": 0.2387707,
+ "warm_p99_ms": 0.25491474000000003,
+ "samples_ms": [
+ 0.23327,
+ 0.231314,
+ 0.230895,
+ 0.229499,
+ 0.227613,
+ 0.261207,
+ 0.238439,
+ 0.233759,
+ 0.229918,
+ 0.230406,
+ 0.229917,
+ 0.232362,
+ 0.231873,
+ 0.229918,
+ 0.227124,
+ 0.227543,
+ 0.227543,
+ 0.228032,
+ 0.228521,
+ 0.226636,
+ 0.228451,
+ 0.227543,
+ 0.245073,
+ 0.232712,
+ 0.22901,
+ 0.233689,
+ 0.228451,
+ 0.227962,
+ 0.227124,
+ 0.228032,
+ 0.226286,
+ 0.227124,
+ 0.228521,
+ 0.228032,
+ 0.23313,
+ 0.228521,
+ 0.227613,
+ 0.227194,
+ 0.227124,
+ 0.228032
+ ],
+ "warm_rows_per_second": 2240494.309056936
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 976769,
+ "unique_rows": 3863,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 121,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 121,
+ "zlib_blocks": 121,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3863
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "10807aca492adfffea37a8a846e93364723fe3ff514ff35f03072b2e27fbacc8",
+ "table_sha256": "afc91dc5e1502b2be927ab8476e155f4bea0ba5d4b6af74bc58c0779b103093c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10740412608720362,
+ "representatives_sha256": "f5f18875305939bf17ea579b32984c74a8ca2c9d708c8fd969ad4f13c7dd6ea7",
+ "structurally_merged_rows": 1470,
+ "backend_artifact_bytes": 977970,
+ "backend_with_key_registry_bytes": 1048904,
+ "backend_with_datrie_key_index_bytes": 1021026,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9037792527690864,
+ "mean_cosine": 0.9987392065989559,
+ "p05_cosine": 0.9932839992700851,
+ "relative_l2_mean": 0.023389868840546283,
+ "relative_l2_p95": 0.11619517829097764,
+ "relative_l2_max": 0.43871066183303176,
+ "weighted_relative_l2_mean": 0.00958315800202989,
+ "weighted_relative_frobenius": 0.01409531774117527,
+ "weighted_mean_cosine": 0.9995017141288418,
+ "max_relative_norm_change": 0.010017310244631894
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9037807344037438,
+ "mean_cosine": 0.9987391954638807,
+ "p05_cosine": 0.9932847885012078,
+ "relative_l2_mean": 0.023240150456866986,
+ "relative_l2_p95": 0.11619874549726088,
+ "relative_l2_max": 0.43869563375108495,
+ "weighted_relative_l2_mean": 0.009401932052507061,
+ "weighted_relative_frobenius": 0.014094520590897984,
+ "weighted_mean_cosine": 0.9995017151988554,
+ "max_relative_norm_change": 0.01002151577688962
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.251709,
+ "warm_p50_ms": 0.229918,
+ "warm_p95_ms": 0.23778939999999998,
+ "warm_p99_ms": 0.25599151,
+ "samples_ms": [
+ 0.23327,
+ 0.233759,
+ 0.23327,
+ 0.230407,
+ 0.231385,
+ 0.26449,
+ 0.232781,
+ 0.230896,
+ 0.229499,
+ 0.22901,
+ 0.228032,
+ 0.228521,
+ 0.228032,
+ 0.22894,
+ 0.22901,
+ 0.231873,
+ 0.234108,
+ 0.232782,
+ 0.229918,
+ 0.228101,
+ 0.228032,
+ 0.228032,
+ 0.242699,
+ 0.232711,
+ 0.229987,
+ 0.232292,
+ 0.231314,
+ 0.229918,
+ 0.229499,
+ 0.232293,
+ 0.227963,
+ 0.227543,
+ 0.228451,
+ 0.227962,
+ 0.228032,
+ 0.227613,
+ 0.227613,
+ 0.228521,
+ 0.233689,
+ 0.237531
+ ],
+ "warm_rows_per_second": 2226880.887968754
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 132,
+ "storage": {
+ "artifact_bytes": 703956,
+ "unique_rows": 5333,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "5208f73b4895fa6286c9b958779d6930895d8e6968f9ee970b34284151bd9200",
+ "table_sha256": "1695a383af41a932d3e663426935548621422726475092855d2e0fca6971008f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.013848282862454653,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 704803,
+ "backend_with_key_registry_bytes": 775737,
+ "backend_with_datrie_key_index_bytes": 747859,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999401787751158,
+ "mean_cosine": 0.999979209003611,
+ "p05_cosine": 0.9999685946479178,
+ "relative_l2_mean": 0.006417865941609127,
+ "relative_l2_p95": 0.007940894253651317,
+ "relative_l2_max": 0.01093801745716858,
+ "weighted_relative_l2_mean": 0.006443848531946682,
+ "weighted_relative_frobenius": 0.006579211484067163,
+ "weighted_mean_cosine": 0.9999790416837705,
+ "max_relative_norm_change": 0.0023069212797726902
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.317498,
+ "warm_p50_ms": 0.122502,
+ "warm_p95_ms": 0.1254769,
+ "warm_p99_ms": 0.13296165,
+ "samples_ms": [
+ 0.125435,
+ 0.123479,
+ 0.123479,
+ 0.122502,
+ 0.122921,
+ 0.124458,
+ 0.123061,
+ 0.122082,
+ 0.121594,
+ 0.122572,
+ 0.123061,
+ 0.121174,
+ 0.121594,
+ 0.12299,
+ 0.12299,
+ 0.122991,
+ 0.121105,
+ 0.122501,
+ 0.121594,
+ 0.120616,
+ 0.122013,
+ 0.122572,
+ 0.122572,
+ 0.122571,
+ 0.122572,
+ 0.121105,
+ 0.122082,
+ 0.126273,
+ 0.121663,
+ 0.122991,
+ 0.137238,
+ 0.122991,
+ 0.122502,
+ 0.121036,
+ 0.121035,
+ 0.121524,
+ 0.121524,
+ 0.122432,
+ 0.121593,
+ 0.121175
+ ],
+ "warm_rows_per_second": 4179523.5996147003
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 712514,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 87,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 86,
+ "zlib_blocks": 39,
+ "xor_blocks": 47,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "5208f73b4895fa6286c9b958779d6930895d8e6968f9ee970b34284151bd9200",
+ "table_sha256": "ed7fa4a02a98023f8dbae82b832b3d270653cea1709997b9c0432c8c36bb5de8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.07655254704877734,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 713758,
+ "backend_with_key_registry_bytes": 784692,
+ "backend_with_datrie_key_index_bytes": 756814,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999401787751158,
+ "mean_cosine": 0.999979209003611,
+ "p05_cosine": 0.9999685946479178,
+ "relative_l2_mean": 0.006417865941609127,
+ "relative_l2_p95": 0.007940894253651317,
+ "relative_l2_max": 0.01093801745716858,
+ "weighted_relative_l2_mean": 0.006443848531946682,
+ "weighted_relative_frobenius": 0.006579211484067163,
+ "weighted_mean_cosine": 0.9999790416837705,
+ "max_relative_norm_change": 0.0023069212797726902
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.440431,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.13729709999999995,
+ "warm_p99_ms": 0.15884004,
+ "samples_ms": [
+ 0.134445,
+ 0.132001,
+ 0.131092,
+ 0.131022,
+ 0.165105,
+ 0.136331,
+ 0.132489,
+ 0.131511,
+ 0.131511,
+ 0.134374,
+ 0.132908,
+ 0.132,
+ 0.133816,
+ 0.132978,
+ 0.132978,
+ 0.132489,
+ 0.136679,
+ 0.132,
+ 0.131092,
+ 0.130673,
+ 0.131022,
+ 0.132489,
+ 0.132978,
+ 0.132,
+ 0.131092,
+ 0.132,
+ 0.132419,
+ 0.132489,
+ 0.131023,
+ 0.131023,
+ 0.131023,
+ 0.132,
+ 0.132,
+ 0.131092,
+ 0.149041,
+ 0.132489,
+ 0.131511,
+ 0.130604,
+ 0.136261,
+ 0.130603
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 712514,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 87,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 86,
+ "zlib_blocks": 39,
+ "xor_blocks": 47,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "5208f73b4895fa6286c9b958779d6930895d8e6968f9ee970b34284151bd9200",
+ "table_sha256": "ed7fa4a02a98023f8dbae82b832b3d270653cea1709997b9c0432c8c36bb5de8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1030453268904239,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 713757,
+ "backend_with_key_registry_bytes": 784691,
+ "backend_with_datrie_key_index_bytes": 756813,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999401787751158,
+ "mean_cosine": 0.999979209003611,
+ "p05_cosine": 0.9999685946479178,
+ "relative_l2_mean": 0.006417865941609127,
+ "relative_l2_p95": 0.007940894253651317,
+ "relative_l2_max": 0.01093801745716858,
+ "weighted_relative_l2_mean": 0.006443848531946682,
+ "weighted_relative_frobenius": 0.006579211484067163,
+ "weighted_mean_cosine": 0.9999790416837705,
+ "max_relative_norm_change": 0.0023069212797726902
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.502381,
+ "warm_p50_ms": 0.13130199999999997,
+ "warm_p95_ms": 0.13998995,
+ "warm_p99_ms": 0.14282131,
+ "samples_ms": [
+ 0.139962,
+ 0.135353,
+ 0.13626,
+ 0.134374,
+ 0.132,
+ 0.132489,
+ 0.132,
+ 0.130115,
+ 0.130115,
+ 0.131512,
+ 0.131092,
+ 0.132,
+ 0.130603,
+ 0.131092,
+ 0.131511,
+ 0.130044,
+ 0.130114,
+ 0.131092,
+ 0.131511,
+ 0.131512,
+ 0.131023,
+ 0.132419,
+ 0.131511,
+ 0.131092,
+ 0.144292,
+ 0.131023,
+ 0.129626,
+ 0.131512,
+ 0.129207,
+ 0.130604,
+ 0.131511,
+ 0.129207,
+ 0.128717,
+ 0.131511,
+ 0.130534,
+ 0.131093,
+ 0.140521,
+ 0.129625,
+ 0.130115,
+ 0.131511
+ ],
+ "warm_rows_per_second": 3899407.472848853
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 712514,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 87,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 86,
+ "zlib_blocks": 39,
+ "xor_blocks": 47,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "5208f73b4895fa6286c9b958779d6930895d8e6968f9ee970b34284151bd9200",
+ "table_sha256": "ed7fa4a02a98023f8dbae82b832b3d270653cea1709997b9c0432c8c36bb5de8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.07710743695497513,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 713757,
+ "backend_with_key_registry_bytes": 784691,
+ "backend_with_datrie_key_index_bytes": 756813,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999401787751158,
+ "mean_cosine": 0.999979209003611,
+ "p05_cosine": 0.9999685946479178,
+ "relative_l2_mean": 0.006417865941609127,
+ "relative_l2_p95": 0.007940894253651317,
+ "relative_l2_max": 0.01093801745716858,
+ "weighted_relative_l2_mean": 0.006443848531946682,
+ "weighted_relative_frobenius": 0.006579211484067163,
+ "weighted_mean_cosine": 0.9999790416837705,
+ "max_relative_norm_change": 0.0023069212797726902
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.437987,
+ "warm_p50_ms": 0.1305685,
+ "warm_p95_ms": 0.13727029999999998,
+ "warm_p99_ms": 0.16159836,
+ "samples_ms": [
+ 0.135771,
+ 0.132419,
+ 0.132489,
+ 0.130603,
+ 0.130603,
+ 0.131511,
+ 0.132,
+ 0.170832,
+ 0.13675,
+ 0.130115,
+ 0.131581,
+ 0.129695,
+ 0.130114,
+ 0.130114,
+ 0.130114,
+ 0.130603,
+ 0.129136,
+ 0.130044,
+ 0.129555,
+ 0.129067,
+ 0.131023,
+ 0.130604,
+ 0.130185,
+ 0.130185,
+ 0.130115,
+ 0.131023,
+ 0.132001,
+ 0.129137,
+ 0.130534,
+ 0.130534,
+ 0.130534,
+ 0.129626,
+ 0.130603,
+ 0.128299,
+ 0.129695,
+ 0.130114,
+ 0.130603,
+ 0.147156,
+ 0.131023,
+ 0.130604
+ ],
+ "warm_rows_per_second": 3921313.3336141566
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 546329,
+ "unique_rows": 3977,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 65,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 64,
+ "zlib_blocks": 64,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3977
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "92880d90fc03d5cf3eead03b8520a9862a554add3465aaba0f4a40c16f0900cd",
+ "table_sha256": "8fdf4025fddce61416df8d32ea46611f332df585b3f46a74511e95488f428e70",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06986642582342029,
+ "representatives_sha256": "645a1bae78604bd202f048c41638ba6c7eb3b361380b0dad44050b671e577b19",
+ "structurally_merged_rows": 1356,
+ "backend_artifact_bytes": 547570,
+ "backend_with_key_registry_bytes": 618504,
+ "backend_with_datrie_key_index_bytes": 590626,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900137733965927,
+ "mean_cosine": 0.9991445037125608,
+ "p05_cosine": 0.994620323071824,
+ "relative_l2_mean": 0.024068633513318827,
+ "relative_l2_p95": 0.10367830120510269,
+ "relative_l2_max": 0.1411979477508241,
+ "weighted_relative_l2_mean": 0.013385180034982664,
+ "weighted_relative_frobenius": 0.009414340895132815,
+ "weighted_mean_cosine": 0.999680359251853,
+ "max_relative_norm_change": 0.011010627250046198
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900061938975968,
+ "mean_cosine": 0.9991546302983773,
+ "p05_cosine": 0.9945684798321575,
+ "relative_l2_mean": 0.019365973926123135,
+ "relative_l2_p95": 0.10420154679809014,
+ "relative_l2_max": 0.14183076097570288,
+ "weighted_relative_l2_mean": 0.0077497218899546345,
+ "weighted_relative_frobenius": 0.006861641269675814,
+ "weighted_mean_cosine": 0.9996964030636359,
+ "max_relative_norm_change": 0.010000496728499994
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.135586,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.14236145,
+ "warm_p99_ms": 0.15935577999999997,
+ "samples_ms": [
+ 0.136191,
+ 0.133397,
+ 0.130114,
+ 0.131092,
+ 0.132489,
+ 0.132001,
+ 0.142826,
+ 0.132,
+ 0.142337,
+ 0.133467,
+ 0.131092,
+ 0.131511,
+ 0.130114,
+ 0.130603,
+ 0.131022,
+ 0.130534,
+ 0.131023,
+ 0.131023,
+ 0.132489,
+ 0.132,
+ 0.130184,
+ 0.169924,
+ 0.135283,
+ 0.131022,
+ 0.130604,
+ 0.130185,
+ 0.131512,
+ 0.132978,
+ 0.132,
+ 0.129695,
+ 0.131511,
+ 0.129625,
+ 0.130603,
+ 0.132419,
+ 0.131511,
+ 0.131022,
+ 0.132,
+ 0.130603,
+ 0.132489,
+ 0.141498
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 535034,
+ "unique_rows": 3888,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3888
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "46de879590309ce9fba0421c568f33bf654acaa684151f0cacba63d878cddb71",
+ "table_sha256": "f5d885199c8c2656804be6296f41356c9129e24120bbeb8ca88263487b2fb53e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06275161588564515,
+ "representatives_sha256": "48ad575a29acd9ac1a08b0d40b2a7e55fde5b214c6f072df671c11c296d5cef4",
+ "structurally_merged_rows": 1445,
+ "backend_artifact_bytes": 536275,
+ "backend_with_key_registry_bytes": 607209,
+ "backend_with_datrie_key_index_bytes": 579331,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9806577959612811,
+ "mean_cosine": 0.9989227975167413,
+ "p05_cosine": 0.9935783568809586,
+ "relative_l2_mean": 0.02667343768566283,
+ "relative_l2_p95": 0.11329985959849637,
+ "relative_l2_max": 0.19763775510265796,
+ "weighted_relative_l2_mean": 0.014358287795764325,
+ "weighted_relative_frobenius": 0.011897454390846881,
+ "weighted_mean_cosine": 0.9995975749788588,
+ "max_relative_norm_change": 0.011010627250046198
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9805934384917934,
+ "mean_cosine": 0.9989324308298284,
+ "p05_cosine": 0.9935625597806309,
+ "relative_l2_mean": 0.022079363647086014,
+ "relative_l2_p95": 0.11344617562647018,
+ "relative_l2_max": 0.19807229127563497,
+ "weighted_relative_l2_mean": 0.008766229228731699,
+ "weighted_relative_frobenius": 0.01001933521861662,
+ "weighted_mean_cosine": 0.99961331448753,
+ "max_relative_norm_change": 0.010054287587219024
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.077338,
+ "warm_p50_ms": 0.132419,
+ "warm_p95_ms": 0.13852409999999998,
+ "warm_p99_ms": 0.16038694999999997,
+ "samples_ms": [
+ 0.133886,
+ 0.132,
+ 0.132489,
+ 0.133886,
+ 0.133397,
+ 0.169784,
+ 0.132,
+ 0.138147,
+ 0.133397,
+ 0.132978,
+ 0.131093,
+ 0.133397,
+ 0.132908,
+ 0.131092,
+ 0.131512,
+ 0.132,
+ 0.132489,
+ 0.131581,
+ 0.130533,
+ 0.132978,
+ 0.132,
+ 0.132419,
+ 0.132489,
+ 0.132909,
+ 0.132419,
+ 0.132,
+ 0.130114,
+ 0.132489,
+ 0.131023,
+ 0.132,
+ 0.131092,
+ 0.132908,
+ 0.132,
+ 0.130534,
+ 0.145689,
+ 0.133397,
+ 0.131023,
+ 0.132489,
+ 0.131022,
+ 0.130603
+ ],
+ "warm_rows_per_second": 3866514.624034315
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 533505,
+ "unique_rows": 3876,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3876
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "2df46e39c77b01a7cd6794a6017d58b9c4fd0b0c8a6cfa0b40063ed84298037d",
+ "table_sha256": "e724bfcd1107587aeda1ca6e0eab4b71b0adb40bcfd5c86e60d2ea5610a4a5da",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06969936494715512,
+ "representatives_sha256": "5af76ac7da327441b5d21cb174a802de2b99b43f5c565afa8201077eaea031f8",
+ "structurally_merged_rows": 1457,
+ "backend_artifact_bytes": 534746,
+ "backend_with_key_registry_bytes": 605680,
+ "backend_with_datrie_key_index_bytes": 577802,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9703153181619626,
+ "mean_cosine": 0.9988612658052478,
+ "p05_cosine": 0.9933448473360086,
+ "relative_l2_mean": 0.02719928555694859,
+ "relative_l2_p95": 0.11534418300944277,
+ "relative_l2_max": 0.24441432359402698,
+ "weighted_relative_l2_mean": 0.014598146855964915,
+ "weighted_relative_frobenius": 0.01226492094403948,
+ "weighted_mean_cosine": 0.9995686228195856,
+ "max_relative_norm_change": 0.011010627250046198
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9702101180641064,
+ "mean_cosine": 0.9988708674464274,
+ "p05_cosine": 0.9933316178451478,
+ "relative_l2_mean": 0.02261960249918222,
+ "relative_l2_p95": 0.11551730729681066,
+ "relative_l2_max": 0.2437288874566517,
+ "weighted_relative_l2_mean": 0.0090110590653126,
+ "weighted_relative_frobenius": 0.010455888060020459,
+ "weighted_mean_cosine": 0.9995842673801404,
+ "max_relative_norm_change": 0.010054287587219024
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.971319,
+ "warm_p50_ms": 0.131581,
+ "warm_p95_ms": 0.1366627,
+ "warm_p99_ms": 0.16107101999999998,
+ "samples_ms": [
+ 0.135771,
+ 0.134375,
+ 0.132978,
+ 0.132,
+ 0.132489,
+ 0.132908,
+ 0.132,
+ 0.132908,
+ 0.132,
+ 0.132,
+ 0.132,
+ 0.132489,
+ 0.132,
+ 0.131512,
+ 0.131023,
+ 0.131022,
+ 0.142965,
+ 0.130603,
+ 0.131511,
+ 0.131092,
+ 0.130044,
+ 0.172647,
+ 0.130114,
+ 0.129207,
+ 0.136191,
+ 0.131511,
+ 0.131581,
+ 0.131511,
+ 0.132001,
+ 0.136331,
+ 0.130604,
+ 0.131581,
+ 0.130533,
+ 0.130604,
+ 0.130603,
+ 0.128787,
+ 0.132,
+ 0.131581,
+ 0.131022,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3891139.298226948
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 532751,
+ "unique_rows": 3870,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3870
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "1326558a36adab45eb74657d89612cf5d4e966215e2b6767ec780ece8ec58a18",
+ "table_sha256": "ffe88508bd76be26454466a622516d4c6528496b960d89628909a825fbe396b5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.07675173482857645,
+ "representatives_sha256": "9283c93386c03bb302c9f7b181f223f010bb6252536ccb514cdd28e60d378107",
+ "structurally_merged_rows": 1463,
+ "backend_artifact_bytes": 533992,
+ "backend_with_key_registry_bytes": 604926,
+ "backend_with_datrie_key_index_bytes": 577048,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9536840917457966,
+ "mean_cosine": 0.9988204158980151,
+ "p05_cosine": 0.9933078276560026,
+ "relative_l2_mean": 0.0274940284608687,
+ "relative_l2_p95": 0.11572949516287494,
+ "relative_l2_max": 0.3058572569757398,
+ "weighted_relative_l2_mean": 0.014713795536892137,
+ "weighted_relative_frobenius": 0.012470527446288186,
+ "weighted_mean_cosine": 0.9995520841447151,
+ "max_relative_norm_change": 0.011010627250046198
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9535673050738439,
+ "mean_cosine": 0.9988300177098813,
+ "p05_cosine": 0.9932640311753331,
+ "relative_l2_mean": 0.022921253553751512,
+ "relative_l2_p95": 0.11644333695897566,
+ "relative_l2_max": 0.3063228619560601,
+ "weighted_relative_l2_mean": 0.00912933391785955,
+ "weighted_relative_frobenius": 0.01069657460148448,
+ "weighted_mean_cosine": 0.9995677220810844,
+ "max_relative_norm_change": 0.010054287587219024
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.135516,
+ "warm_p50_ms": 0.13221,
+ "warm_p95_ms": 0.14197054999999997,
+ "warm_p99_ms": 0.15036738,
+ "samples_ms": [
+ 0.137657,
+ 0.141569,
+ 0.133886,
+ 0.130044,
+ 0.132,
+ 0.134374,
+ 0.134445,
+ 0.133467,
+ 0.1496,
+ 0.134444,
+ 0.132489,
+ 0.130533,
+ 0.133466,
+ 0.132908,
+ 0.132,
+ 0.137169,
+ 0.137239,
+ 0.134794,
+ 0.132001,
+ 0.133467,
+ 0.133816,
+ 0.132,
+ 0.131581,
+ 0.131092,
+ 0.130533,
+ 0.132001,
+ 0.132419,
+ 0.131511,
+ 0.131511,
+ 0.133397,
+ 0.132,
+ 0.132908,
+ 0.131093,
+ 0.131022,
+ 0.131022,
+ 0.131511,
+ 0.132,
+ 0.132001,
+ 0.150858,
+ 0.130603
+ ],
+ "warm_rows_per_second": 3872626.8814764395
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 531770,
+ "unique_rows": 3862,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3862
+ },
+ "source_float32_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b",
+ "decoded_float32_sha256": "7211c59801c5795c3b3925ab3fa757b1cee00caea511a0242254ee11e45d4c96",
+ "table_sha256": "9691fd208b593744b9dbba163a604afc4372d52ac2d57d134df4eb774def4a8e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.0665085930377245,
+ "representatives_sha256": "a75e3508a0ffe6b1bde4d8e97191ec9cb936367875475d634e0557b84f75c9a3",
+ "structurally_merged_rows": 1471,
+ "backend_artifact_bytes": 533011,
+ "backend_with_key_registry_bytes": 603945,
+ "backend_with_datrie_key_index_bytes": 576067,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9038088300906122,
+ "mean_cosine": 0.9987095735022165,
+ "p05_cosine": 0.9931008619725634,
+ "relative_l2_mean": 0.02805887765538132,
+ "relative_l2_p95": 0.11731227593917865,
+ "relative_l2_max": 0.4387194163405977,
+ "weighted_relative_l2_mean": 0.015091432431308172,
+ "weighted_relative_frobenius": 0.015538193567423851,
+ "weighted_mean_cosine": 0.9994786605194113,
+ "max_relative_norm_change": 0.011010627250046198
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9036987430160472,
+ "mean_cosine": 0.9987190532983201,
+ "p05_cosine": 0.9931055175354562,
+ "relative_l2_mean": 0.023495856889297635,
+ "relative_l2_p95": 0.11749266590090074,
+ "relative_l2_max": 0.43882451952200247,
+ "weighted_relative_l2_mean": 0.00951363456066782,
+ "weighted_relative_frobenius": 0.01415687281307621,
+ "weighted_mean_cosine": 0.9994941974605697,
+ "max_relative_norm_change": 0.010054287587219024
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.185801,
+ "warm_p50_ms": 0.13245400000000002,
+ "warm_p95_ms": 0.1397071,
+ "warm_p99_ms": 0.14891707,
+ "samples_ms": [
+ 0.137657,
+ 0.142825,
+ 0.134375,
+ 0.132489,
+ 0.132,
+ 0.13619,
+ 0.133956,
+ 0.132489,
+ 0.131511,
+ 0.133886,
+ 0.133397,
+ 0.152812,
+ 0.133327,
+ 0.132489,
+ 0.133886,
+ 0.132978,
+ 0.139543,
+ 0.132978,
+ 0.135701,
+ 0.132,
+ 0.132419,
+ 0.130115,
+ 0.131512,
+ 0.130603,
+ 0.135282,
+ 0.132,
+ 0.129695,
+ 0.130603,
+ 0.129695,
+ 0.130603,
+ 0.131511,
+ 0.135213,
+ 0.133467,
+ 0.132001,
+ 0.132,
+ 0.130044,
+ 0.130603,
+ 0.130114,
+ 0.132,
+ 0.132419
+ ],
+ "warm_rows_per_second": 3865492.925845954
+ }
+ }
+ ],
+ "baseline_quality": {},
+ "training_row_sha256": "8e02a804ab58387d46d2ffc3cf20b2e4afe2a2d05b24dca43ec60087753a574b"
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 11,
+ "rows": 5333,
+ "dim": 128,
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "source_float32_bytes": 2730496,
+ "key_registry_bytes": 70934,
+ "unchanged_backbone_artifact_bytes": 0,
+ "alternative_datrie_key_index_bytes": 43056,
+ "query_rows": 512,
+ "query_ids_sha256": "87677d489e681e6c0a39b8460231cf990090e3318abb13c912a002c527931d01",
+ "frequency_trace_sha256": "0e7a36c623f2fc5cc1e4fd51bc27b7b98d60f0c1b31a56fb66253e951084f56a",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 512,
+ "storage": {
+ "artifact_bytes": 2730496,
+ "unique_rows": 5333,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "table_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.015308177098631859,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2731299,
+ "backend_with_key_registry_bytes": 2802233,
+ "backend_with_datrie_key_index_bytes": 2774355,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.340197,
+ "warm_p50_ms": 0.062019,
+ "warm_p95_ms": 0.06301795,
+ "warm_p99_ms": 0.06401258,
+ "samples_ms": [
+ 0.064394,
+ 0.062508,
+ 0.062508,
+ 0.062927,
+ 0.061879,
+ 0.06153,
+ 0.062997,
+ 0.062508,
+ 0.060553,
+ 0.060972,
+ 0.061949,
+ 0.062019,
+ 0.062508,
+ 0.062508,
+ 0.06188,
+ 0.06153,
+ 0.060552,
+ 0.063416,
+ 0.062508,
+ 0.061531,
+ 0.06195,
+ 0.062368,
+ 0.062508,
+ 0.062438,
+ 0.0616,
+ 0.061461,
+ 0.061949,
+ 0.060622,
+ 0.060134,
+ 0.062438,
+ 0.062508,
+ 0.062508,
+ 0.062019,
+ 0.06202,
+ 0.062368,
+ 0.061949,
+ 0.061041,
+ 0.061111,
+ 0.062857,
+ 0.06188
+ ],
+ "warm_rows_per_second": 8255534.594237249
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 2607383,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 334,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 334,
+ "zlib_blocks": 298,
+ "xor_blocks": 36,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "table_sha256": "4465813fb78d28bdba11b7014025de4b6859d2738003af396758fd432063ecab",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.23725302796810865,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2608587,
+ "backend_with_key_registry_bytes": 2679521,
+ "backend_with_datrie_key_index_bytes": 2651643,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.498284,
+ "warm_p50_ms": 0.070051,
+ "warm_p95_ms": 0.0757534,
+ "warm_p99_ms": 0.09110106,
+ "samples_ms": [
+ 0.076616,
+ 0.073823,
+ 0.075708,
+ 0.073752,
+ 0.073822,
+ 0.071937,
+ 0.073823,
+ 0.075219,
+ 0.071937,
+ 0.073334,
+ 0.071029,
+ 0.100362,
+ 0.07515,
+ 0.071937,
+ 0.07054,
+ 0.070051,
+ 0.069004,
+ 0.069981,
+ 0.069561,
+ 0.070051,
+ 0.070051,
+ 0.069492,
+ 0.069981,
+ 0.070051,
+ 0.069073,
+ 0.070889,
+ 0.069562,
+ 0.069632,
+ 0.069562,
+ 0.069073,
+ 0.069562,
+ 0.069632,
+ 0.069073,
+ 0.069004,
+ 0.069981,
+ 0.070539,
+ 0.069562,
+ 0.070958,
+ 0.0704,
+ 0.069493
+ ],
+ "warm_rows_per_second": 7308960.614409502
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 2607383,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 334,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 334,
+ "zlib_blocks": 298,
+ "xor_blocks": 36,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "table_sha256": "4465813fb78d28bdba11b7014025de4b6859d2738003af396758fd432063ecab",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2622648940887302,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2608586,
+ "backend_with_key_registry_bytes": 2679520,
+ "backend_with_datrie_key_index_bytes": 2651642,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.52629,
+ "warm_p50_ms": 0.070959,
+ "warm_p95_ms": 0.07661945,
+ "warm_p99_ms": 0.09082968,
+ "samples_ms": [
+ 0.076685,
+ 0.076616,
+ 0.0748,
+ 0.099873,
+ 0.074311,
+ 0.074171,
+ 0.073822,
+ 0.073403,
+ 0.071936,
+ 0.073822,
+ 0.072914,
+ 0.073822,
+ 0.073752,
+ 0.069981,
+ 0.069981,
+ 0.07047,
+ 0.07061,
+ 0.069632,
+ 0.071378,
+ 0.069003,
+ 0.07054,
+ 0.07005,
+ 0.069073,
+ 0.069422,
+ 0.069142,
+ 0.071029,
+ 0.070889,
+ 0.069981,
+ 0.070958,
+ 0.070959,
+ 0.070959,
+ 0.070959,
+ 0.070051,
+ 0.070959,
+ 0.071448,
+ 0.071378,
+ 0.070889,
+ 0.070889,
+ 0.070889,
+ 0.070051
+ ],
+ "warm_rows_per_second": 7215434.264857171
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 2607383,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 334,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 334,
+ "zlib_blocks": 298,
+ "xor_blocks": 36,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "table_sha256": "4465813fb78d28bdba11b7014025de4b6859d2738003af396758fd432063ecab",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.23869595117866993,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2608586,
+ "backend_with_key_registry_bytes": 2679520,
+ "backend_with_datrie_key_index_bytes": 2651642,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.503522,
+ "warm_p50_ms": 0.0700505,
+ "warm_p95_ms": 0.073822,
+ "warm_p99_ms": 0.08421761999999999,
+ "samples_ms": [
+ 0.090864,
+ 0.073333,
+ 0.071936,
+ 0.071937,
+ 0.073334,
+ 0.073753,
+ 0.073822,
+ 0.072426,
+ 0.071937,
+ 0.073333,
+ 0.071936,
+ 0.069562,
+ 0.069073,
+ 0.069073,
+ 0.069073,
+ 0.069562,
+ 0.073822,
+ 0.070958,
+ 0.07047,
+ 0.070959,
+ 0.07047,
+ 0.068654,
+ 0.070959,
+ 0.071378,
+ 0.069073,
+ 0.07012,
+ 0.069562,
+ 0.07047,
+ 0.069492,
+ 0.069143,
+ 0.068654,
+ 0.069004,
+ 0.069143,
+ 0.068584,
+ 0.069492,
+ 0.068654,
+ 0.069003,
+ 0.069981,
+ 0.069143,
+ 0.069981
+ ],
+ "warm_rows_per_second": 7309012.783634664
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1968073,
+ "unique_rows": 3998,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 250,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 250,
+ "zlib_blocks": 250,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3998
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "8eb928aaa7073a9ca338dd2537d05b79bf8d542536c211384f987218d2e09d51",
+ "table_sha256": "32675e6148c95f00069758bb3f78daac16c7eb7578877c6c34642e708b01a11b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.18388285115361214,
+ "representatives_sha256": "1c84eef9686f2d387df0529d6f8c244ea07ca48509de04df67e0e50c6d680702",
+ "structurally_merged_rows": 1335,
+ "backend_artifact_bytes": 1969275,
+ "backend_with_key_registry_bytes": 2040209,
+ "backend_with_datrie_key_index_bytes": 2012331,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900325358107495,
+ "mean_cosine": 0.9991851815172184,
+ "p05_cosine": 0.9947963921319865,
+ "relative_l2_mean": 0.018818578252097564,
+ "relative_l2_p95": 0.1020233298539952,
+ "relative_l2_max": 0.14137685298673708,
+ "weighted_relative_l2_mean": 0.0075909892553223425,
+ "weighted_relative_frobenius": 0.006788218213522045,
+ "weighted_mean_cosine": 0.9997044562858881,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900325358107495,
+ "mean_cosine": 0.9991851815172184,
+ "p05_cosine": 0.9947963921319865,
+ "relative_l2_mean": 0.018818578252097564,
+ "relative_l2_p95": 0.1020233298539952,
+ "relative_l2_max": 0.14137685298673708,
+ "weighted_relative_l2_mean": 0.0075909892553223425,
+ "weighted_relative_frobenius": 0.006788218213522045,
+ "weighted_mean_cosine": 0.9997044562858881,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.617445,
+ "warm_p50_ms": 0.070784,
+ "warm_p95_ms": 0.0742934,
+ "warm_p99_ms": 0.07554459,
+ "samples_ms": [
+ 0.075289,
+ 0.073822,
+ 0.072774,
+ 0.073333,
+ 0.073334,
+ 0.073334,
+ 0.073822,
+ 0.073334,
+ 0.074241,
+ 0.072355,
+ 0.07061,
+ 0.072495,
+ 0.073334,
+ 0.075708,
+ 0.074241,
+ 0.07061,
+ 0.070958,
+ 0.070959,
+ 0.072845,
+ 0.073264,
+ 0.071447,
+ 0.07054,
+ 0.069004,
+ 0.069073,
+ 0.070539,
+ 0.071029,
+ 0.070121,
+ 0.070051,
+ 0.069004,
+ 0.070051,
+ 0.070051,
+ 0.069493,
+ 0.069143,
+ 0.069562,
+ 0.069492,
+ 0.069003,
+ 0.069493,
+ 0.069632,
+ 0.070051,
+ 0.069981
+ ],
+ "warm_rows_per_second": 7233273.056057866
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1912055,
+ "unique_rows": 3882,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 243,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 243,
+ "zlib_blocks": 243,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3882
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "e165804434e8b2612590d199b2805e2c962eecfe00fa342f9cf42695b81c1fbf",
+ "table_sha256": "6ccc91d2a792298155061479d627108bae5c69e627fa69d211579841315b1445",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1820334519725293,
+ "representatives_sha256": "526b2f53594f83bee7561ae6618b98fff3af1c262afe84aceb614f36c693f4bd",
+ "structurally_merged_rows": 1451,
+ "backend_artifact_bytes": 1913257,
+ "backend_with_key_registry_bytes": 1984191,
+ "backend_with_datrie_key_index_bytes": 1956313,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.980050565596075,
+ "mean_cosine": 0.9988970098862606,
+ "p05_cosine": 0.9930372589094937,
+ "relative_l2_mean": 0.02234897656608579,
+ "relative_l2_p95": 0.11789988934867349,
+ "relative_l2_max": 0.19904708995990741,
+ "weighted_relative_l2_mean": 0.008612040314153803,
+ "weighted_relative_frobenius": 0.007105333389532846,
+ "weighted_mean_cosine": 0.9996212089362276,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.980050565596075,
+ "mean_cosine": 0.9988970098862606,
+ "p05_cosine": 0.9930372589094937,
+ "relative_l2_mean": 0.02234897656608579,
+ "relative_l2_p95": 0.11789988934867349,
+ "relative_l2_max": 0.19904708995990741,
+ "weighted_relative_l2_mean": 0.008612040314153803,
+ "weighted_relative_frobenius": 0.007105333389532846,
+ "weighted_mean_cosine": 0.9996212089362276,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.144759,
+ "warm_p50_ms": 0.071448,
+ "warm_p95_ms": 0.08397744999999998,
+ "warm_p99_ms": 0.10604113999999999,
+ "samples_ms": [
+ 0.074242,
+ 0.073822,
+ 0.074241,
+ 0.073334,
+ 0.074731,
+ 0.071518,
+ 0.072006,
+ 0.070959,
+ 0.071448,
+ 0.071517,
+ 0.071378,
+ 0.069073,
+ 0.069073,
+ 0.071517,
+ 0.069143,
+ 0.069073,
+ 0.069004,
+ 0.070121,
+ 0.072914,
+ 0.071517,
+ 0.071029,
+ 0.071518,
+ 0.117263,
+ 0.073333,
+ 0.071378,
+ 0.070889,
+ 0.071866,
+ 0.070889,
+ 0.070889,
+ 0.070539,
+ 0.071448,
+ 0.071517,
+ 0.071448,
+ 0.070959,
+ 0.072426,
+ 0.070539,
+ 0.070121,
+ 0.088489,
+ 0.08374,
+ 0.071447
+ ],
+ "warm_rows_per_second": 7166050.834173105
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1908982,
+ "unique_rows": 3876,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 243,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 243,
+ "zlib_blocks": 243,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3876
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "d94ce50a3aaa6506480e876528b4f8bb8bfc273d02285a4c6253bc232f78b662",
+ "table_sha256": "78ed4ab50acd0e1953318e8d063724ce18ae3ccd996b9f79f7d4e7b9abd2a5df",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2236357107758522,
+ "representatives_sha256": "adaf5dd35d194f4bd19e7f348d2d4e6f88cff8790b2d2bdb8d7d8eb345aac93f",
+ "structurally_merged_rows": 1457,
+ "backend_artifact_bytes": 1910184,
+ "backend_with_key_registry_bytes": 1981118,
+ "backend_with_datrie_key_index_bytes": 1953240,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.974592749547151,
+ "mean_cosine": 0.9988713550106368,
+ "p05_cosine": 0.9929501043360548,
+ "relative_l2_mean": 0.022588941888132666,
+ "relative_l2_p95": 0.11875619066074654,
+ "relative_l2_max": 0.2246727709380708,
+ "weighted_relative_l2_mean": 0.008680336560416798,
+ "weighted_relative_frobenius": 0.007114297581872774,
+ "weighted_mean_cosine": 0.9996139073324438,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.974592749547151,
+ "mean_cosine": 0.9988713550106368,
+ "p05_cosine": 0.9929501043360548,
+ "relative_l2_mean": 0.022588941888132666,
+ "relative_l2_p95": 0.11875619066074654,
+ "relative_l2_max": 0.2246727709380708,
+ "weighted_relative_l2_mean": 0.008680336560416798,
+ "weighted_relative_frobenius": 0.007114297581872774,
+ "weighted_mean_cosine": 0.9996139073324438,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.247844,
+ "warm_p50_ms": 0.07092399999999999,
+ "warm_p95_ms": 0.07503045,
+ "warm_p99_ms": 0.08090045,
+ "samples_ms": [
+ 0.079409,
+ 0.0748,
+ 0.073822,
+ 0.073822,
+ 0.073333,
+ 0.073333,
+ 0.072844,
+ 0.073333,
+ 0.072845,
+ 0.071099,
+ 0.072006,
+ 0.072355,
+ 0.071518,
+ 0.071029,
+ 0.069981,
+ 0.070889,
+ 0.069492,
+ 0.071378,
+ 0.070959,
+ 0.069143,
+ 0.069492,
+ 0.073893,
+ 0.069073,
+ 0.069492,
+ 0.071028,
+ 0.068584,
+ 0.081854,
+ 0.069562,
+ 0.070889,
+ 0.069423,
+ 0.070121,
+ 0.069073,
+ 0.068514,
+ 0.071517,
+ 0.069073,
+ 0.069073,
+ 0.069562,
+ 0.068654,
+ 0.068095,
+ 0.069143
+ ],
+ "warm_rows_per_second": 7218994.980542554
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1904168,
+ "unique_rows": 3866,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 242,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 242,
+ "zlib_blocks": 242,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3866
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "55ea65a0c372435523a39ff35007088ad1ed00eb1d96f5ef67498de6c2250f0b",
+ "table_sha256": "323a389e405126fa74b4ccfce7f273d0be0103079868fcb3cdc9fca02ad37cd5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.18119738181121647,
+ "representatives_sha256": "c8dc8fcd1fcfc72a5977c5b2a2e679a5707784d65f7950540f6c4f47c865e933",
+ "structurally_merged_rows": 1467,
+ "backend_artifact_bytes": 1905370,
+ "backend_with_key_registry_bytes": 1976304,
+ "backend_with_datrie_key_index_bytes": 1948426,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9510872301154241,
+ "mean_cosine": 0.9987918492822377,
+ "p05_cosine": 0.9927808823560548,
+ "relative_l2_mean": 0.023153683395507205,
+ "relative_l2_p95": 0.12037670247217518,
+ "relative_l2_max": 0.31151735584289914,
+ "weighted_relative_l2_mean": 0.00908691072317337,
+ "weighted_relative_frobenius": 0.008899206643852799,
+ "weighted_mean_cosine": 0.9995569912219132,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9510872301154241,
+ "mean_cosine": 0.9987918492822377,
+ "p05_cosine": 0.9927808823560548,
+ "relative_l2_mean": 0.023153683395507205,
+ "relative_l2_p95": 0.12037670247217518,
+ "relative_l2_max": 0.31151735584289914,
+ "weighted_relative_l2_mean": 0.00908691072317337,
+ "weighted_relative_frobenius": 0.008899206643852799,
+ "weighted_mean_cosine": 0.9995569912219132,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.751971,
+ "warm_p50_ms": 0.069632,
+ "warm_p95_ms": 0.07729035,
+ "warm_p99_ms": 0.08680057999999999,
+ "samples_ms": [
+ 0.077245,
+ 0.09233,
+ 0.078152,
+ 0.074242,
+ 0.073753,
+ 0.073822,
+ 0.072007,
+ 0.070889,
+ 0.070889,
+ 0.072845,
+ 0.071378,
+ 0.069492,
+ 0.070051,
+ 0.069143,
+ 0.069562,
+ 0.068095,
+ 0.069632,
+ 0.068584,
+ 0.069492,
+ 0.068584,
+ 0.069981,
+ 0.069492,
+ 0.069143,
+ 0.069143,
+ 0.069004,
+ 0.069143,
+ 0.069073,
+ 0.069004,
+ 0.068654,
+ 0.068165,
+ 0.069562,
+ 0.069631,
+ 0.069981,
+ 0.0704,
+ 0.069981,
+ 0.069073,
+ 0.07005,
+ 0.069981,
+ 0.069981,
+ 0.069632
+ ],
+ "warm_rows_per_second": 7352941.176470588
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2730496,
+ "artifact_bytes": 1901839,
+ "unique_rows": 3861,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 242,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 242,
+ "zlib_blocks": 242,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3861
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "68eaa11999ab3221c4db3a9e438703b55d59bd2a50c27d15bd5cb056bcc6241d",
+ "table_sha256": "0197dd4a6473f21e2eac1ca153165eedc05861570240e9d5e4e159f2c6e1dff2",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.180918226018548,
+ "representatives_sha256": "1075cc01ecee84c61b5d73b08234d943406ba73bfffd6143b6358618207444a4",
+ "structurally_merged_rows": 1472,
+ "backend_artifact_bytes": 1903041,
+ "backend_with_key_registry_bytes": 1973975,
+ "backend_with_datrie_key_index_bytes": 1946097,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9127493431159585,
+ "mean_cosine": 0.9986919900464016,
+ "p05_cosine": 0.9925640858758106,
+ "relative_l2_mean": 0.023674874425416603,
+ "relative_l2_p95": 0.12177663040843287,
+ "relative_l2_max": 0.4187222387947898,
+ "weighted_relative_l2_mean": 0.009475765678451318,
+ "weighted_relative_frobenius": 0.013167753291854253,
+ "weighted_mean_cosine": 0.9994853902004286,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9127493431159585,
+ "mean_cosine": 0.9986919900464016,
+ "p05_cosine": 0.9925640858758106,
+ "relative_l2_mean": 0.023674874425416603,
+ "relative_l2_p95": 0.12177663040843287,
+ "relative_l2_max": 0.4187222387947898,
+ "weighted_relative_l2_mean": 0.009475765678451318,
+ "weighted_relative_frobenius": 0.013167753291854253,
+ "weighted_mean_cosine": 0.9994853902004286,
+ "max_relative_norm_change": 0.010066885851052261
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.805959,
+ "warm_p50_ms": 0.070959,
+ "warm_p95_ms": 0.0743783,
+ "warm_p99_ms": 0.075638,
+ "samples_ms": [
+ 0.075638,
+ 0.073752,
+ 0.075638,
+ 0.073823,
+ 0.073264,
+ 0.073334,
+ 0.074312,
+ 0.072426,
+ 0.072844,
+ 0.072355,
+ 0.072844,
+ 0.072356,
+ 0.073334,
+ 0.072915,
+ 0.069981,
+ 0.073752,
+ 0.071936,
+ 0.072915,
+ 0.070959,
+ 0.070539,
+ 0.070959,
+ 0.0704,
+ 0.070051,
+ 0.071029,
+ 0.07054,
+ 0.070121,
+ 0.070051,
+ 0.069981,
+ 0.069562,
+ 0.07005,
+ 0.07054,
+ 0.07047,
+ 0.069003,
+ 0.069562,
+ 0.069562,
+ 0.069911,
+ 0.069562,
+ 0.07054,
+ 0.070051,
+ 0.070959
+ ],
+ "warm_rows_per_second": 7215434.264857171
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 1365248,
+ "unique_rows": 5333,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "31d7dd18c6916ca11b7d50daa0b11adb6c6761085215b2407cd691841766649b",
+ "table_sha256": "0ed0935ce9f5e17b24267182eca235c452de95389f0197f3762fa1b231326bce",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.017962428042665124,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1366052,
+ "backend_with_key_registry_bytes": 1436986,
+ "backend_with_datrie_key_index_bytes": 1409108,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999644045328,
+ "mean_cosine": 0.9999999789970508,
+ "p05_cosine": 0.9999999733072,
+ "relative_l2_mean": 0.00020670589716848589,
+ "relative_l2_p95": 0.00023385651072794117,
+ "relative_l2_max": 0.00026765789895794544,
+ "weighted_relative_l2_mean": 0.00020683610612723765,
+ "weighted_relative_frobenius": 0.00020894692652536232,
+ "weighted_mean_cosine": 0.9999999789186156,
+ "max_relative_norm_change": 0.00012501920765291697
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.450895,
+ "warm_p50_ms": 0.2200005,
+ "warm_p95_ms": 0.2322719,
+ "warm_p99_ms": 0.24794164999999999,
+ "samples_ms": [
+ 0.222305,
+ 0.257366,
+ 0.223772,
+ 0.221886,
+ 0.220419,
+ 0.220419,
+ 0.21993,
+ 0.220001,
+ 0.220001,
+ 0.22,
+ 0.22,
+ 0.219023,
+ 0.219023,
+ 0.220001,
+ 0.219581,
+ 0.220978,
+ 0.219092,
+ 0.219581,
+ 0.219581,
+ 0.233201,
+ 0.221467,
+ 0.219022,
+ 0.219511,
+ 0.22,
+ 0.219511,
+ 0.219511,
+ 0.221397,
+ 0.222864,
+ 0.219511,
+ 0.219511,
+ 0.220489,
+ 0.220978,
+ 0.22049,
+ 0.219581,
+ 0.219093,
+ 0.219162,
+ 0.219023,
+ 0.232223,
+ 0.220489,
+ 0.22007
+ ],
+ "warm_rows_per_second": 2327267.4380285502
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1328953,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 167,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 167,
+ "zlib_blocks": 146,
+ "xor_blocks": 21,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "31d7dd18c6916ca11b7d50daa0b11adb6c6761085215b2407cd691841766649b",
+ "table_sha256": "ed4a1926028c10f5f42578dfcb839d0a5f1e1ec0218903327b75f854b6db8e60",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.12861166195943952,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1330158,
+ "backend_with_key_registry_bytes": 1401092,
+ "backend_with_datrie_key_index_bytes": 1373214,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999644045328,
+ "mean_cosine": 0.9999999789970508,
+ "p05_cosine": 0.9999999733072,
+ "relative_l2_mean": 0.00020670589716848589,
+ "relative_l2_p95": 0.00023385651072794117,
+ "relative_l2_max": 0.00026765789895794544,
+ "weighted_relative_l2_mean": 0.00020683610612723765,
+ "weighted_relative_frobenius": 0.00020894692652536232,
+ "weighted_mean_cosine": 0.9999999789186156,
+ "max_relative_norm_change": 0.00012501920765291697
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.093018,
+ "warm_p50_ms": 0.228521,
+ "warm_p95_ms": 0.23618624999999996,
+ "warm_p99_ms": 0.2516672,
+ "samples_ms": [
+ 0.232362,
+ 0.230336,
+ 0.235645,
+ 0.231874,
+ 0.230896,
+ 0.230336,
+ 0.230476,
+ 0.22901,
+ 0.229009,
+ 0.228032,
+ 0.228102,
+ 0.227054,
+ 0.227054,
+ 0.228451,
+ 0.228521,
+ 0.227962,
+ 0.25499,
+ 0.230407,
+ 0.228521,
+ 0.229499,
+ 0.229009,
+ 0.228102,
+ 0.227124,
+ 0.228591,
+ 0.22901,
+ 0.228521,
+ 0.228102,
+ 0.22859,
+ 0.228102,
+ 0.226705,
+ 0.227124,
+ 0.228032,
+ 0.227963,
+ 0.227962,
+ 0.24647,
+ 0.229429,
+ 0.227473,
+ 0.228032,
+ 0.229499,
+ 0.228032
+ ],
+ "warm_rows_per_second": 2240494.309056936
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1328953,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 167,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 167,
+ "zlib_blocks": 146,
+ "xor_blocks": 21,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "31d7dd18c6916ca11b7d50daa0b11adb6c6761085215b2407cd691841766649b",
+ "table_sha256": "ed4a1926028c10f5f42578dfcb839d0a5f1e1ec0218903327b75f854b6db8e60",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.15969895594753325,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1330157,
+ "backend_with_key_registry_bytes": 1401091,
+ "backend_with_datrie_key_index_bytes": 1373213,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999644045328,
+ "mean_cosine": 0.9999999789970508,
+ "p05_cosine": 0.9999999733072,
+ "relative_l2_mean": 0.00020670589716848589,
+ "relative_l2_p95": 0.00023385651072794117,
+ "relative_l2_max": 0.00026765789895794544,
+ "weighted_relative_l2_mean": 0.00020683610612723765,
+ "weighted_relative_frobenius": 0.00020894692652536232,
+ "weighted_mean_cosine": 0.9999999789186156,
+ "max_relative_norm_change": 0.00012501920765291697
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.446904,
+ "warm_p50_ms": 0.277759,
+ "warm_p95_ms": 0.2962492,
+ "warm_p99_ms": 0.31126753999999995,
+ "samples_ms": [
+ 0.230895,
+ 0.230337,
+ 0.22859,
+ 0.229499,
+ 0.258762,
+ 0.282369,
+ 0.278178,
+ 0.277759,
+ 0.275804,
+ 0.282508,
+ 0.278667,
+ 0.298571,
+ 0.280064,
+ 0.278737,
+ 0.280553,
+ 0.277689,
+ 0.296127,
+ 0.241232,
+ 0.229499,
+ 0.228521,
+ 0.227613,
+ 0.229009,
+ 0.228032,
+ 0.228521,
+ 0.228521,
+ 0.228102,
+ 0.227125,
+ 0.319385,
+ 0.279156,
+ 0.280133,
+ 0.282019,
+ 0.277759,
+ 0.279575,
+ 0.278178,
+ 0.279156,
+ 0.275455,
+ 0.277759,
+ 0.273499,
+ 0.279644,
+ 0.28195
+ ],
+ "warm_rows_per_second": 1843324.6087435512
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1328953,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 167,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 167,
+ "zlib_blocks": 146,
+ "xor_blocks": 21,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "31d7dd18c6916ca11b7d50daa0b11adb6c6761085215b2407cd691841766649b",
+ "table_sha256": "ed4a1926028c10f5f42578dfcb839d0a5f1e1ec0218903327b75f854b6db8e60",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1359516391530633,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1330157,
+ "backend_with_key_registry_bytes": 1401091,
+ "backend_with_datrie_key_index_bytes": 1373213,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999644045328,
+ "mean_cosine": 0.9999999789970508,
+ "p05_cosine": 0.9999999733072,
+ "relative_l2_mean": 0.00020670589716848589,
+ "relative_l2_p95": 0.00023385651072794117,
+ "relative_l2_max": 0.00026765789895794544,
+ "weighted_relative_l2_mean": 0.00020683610612723765,
+ "weighted_relative_frobenius": 0.00020894692652536232,
+ "weighted_mean_cosine": 0.9999999789186156,
+ "max_relative_norm_change": 0.00012501920765291697
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.161113,
+ "warm_p50_ms": 0.22782249999999998,
+ "warm_p95_ms": 0.2777449,
+ "warm_p99_ms": 0.29136984,
+ "samples_ms": [
+ 0.231873,
+ 0.22894,
+ 0.227543,
+ 0.245981,
+ 0.231385,
+ 0.229079,
+ 0.22908,
+ 0.286768,
+ 0.227683,
+ 0.229987,
+ 0.227613,
+ 0.227124,
+ 0.228521,
+ 0.228451,
+ 0.227544,
+ 0.22901,
+ 0.227474,
+ 0.227962,
+ 0.227124,
+ 0.227613,
+ 0.294312,
+ 0.235155,
+ 0.229988,
+ 0.231314,
+ 0.226636,
+ 0.227613,
+ 0.227683,
+ 0.227613,
+ 0.227543,
+ 0.228032,
+ 0.226077,
+ 0.227543,
+ 0.227124,
+ 0.227543,
+ 0.227613,
+ 0.227543,
+ 0.232781,
+ 0.27727,
+ 0.232711,
+ 0.227543
+ ],
+ "warm_rows_per_second": 2247363.627385355
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 1009598,
+ "unique_rows": 3998,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 125,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 125,
+ "zlib_blocks": 125,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3998
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "d82a65c6e1e5eaa45e3dec8e0290e765ae7e1907061d70e69f0cf90c0cd6fea7",
+ "table_sha256": "993bbb7fcf9cf986c66605694c2fc4593c7e071cff3183528670fc2614cd5584",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.11485298606567085,
+ "representatives_sha256": "440e2d6bdd0874af8f64f811c85840996a519b5535ca13cd99366b0737b6ceae",
+ "structurally_merged_rows": 1335,
+ "backend_artifact_bytes": 1010800,
+ "backend_with_key_registry_bytes": 1081734,
+ "backend_with_datrie_key_index_bytes": 1053856,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900297489440841,
+ "mean_cosine": 0.9991859754494833,
+ "p05_cosine": 0.9948008103489178,
+ "relative_l2_mean": 0.018964019676226754,
+ "relative_l2_p95": 0.1019787228960663,
+ "relative_l2_max": 0.14139286698093315,
+ "weighted_relative_l2_mean": 0.007774702292468263,
+ "weighted_relative_frobenius": 0.006792836185212061,
+ "weighted_mean_cosine": 0.9997045343265375,
+ "max_relative_norm_change": 0.010144653073261482
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900292416446087,
+ "mean_cosine": 0.9991859895150045,
+ "p05_cosine": 0.9948020602316103,
+ "relative_l2_mean": 0.018809038392007316,
+ "relative_l2_p95": 0.10196921350794233,
+ "relative_l2_max": 0.14139712152680467,
+ "weighted_relative_l2_mean": 0.007592158104580025,
+ "weighted_relative_frobenius": 0.0067896740613937155,
+ "weighted_mean_cosine": 0.9997045524928819,
+ "max_relative_norm_change": 0.010084202899738793
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.216859,
+ "warm_p50_ms": 0.228276,
+ "warm_p95_ms": 0.27095269999999994,
+ "warm_p99_ms": 0.31252341999999994,
+ "samples_ms": [
+ 0.229918,
+ 0.229498,
+ 0.230407,
+ 0.228032,
+ 0.227613,
+ 0.227613,
+ 0.226635,
+ 0.227054,
+ 0.227962,
+ 0.228032,
+ 0.227613,
+ 0.227194,
+ 0.329302,
+ 0.23327,
+ 0.228591,
+ 0.22852,
+ 0.228032,
+ 0.232781,
+ 0.233759,
+ 0.230896,
+ 0.230476,
+ 0.227543,
+ 0.228521,
+ 0.228521,
+ 0.227194,
+ 0.226147,
+ 0.227473,
+ 0.229429,
+ 0.228521,
+ 0.270146,
+ 0.230825,
+ 0.230407,
+ 0.22901,
+ 0.28628,
+ 0.226635,
+ 0.227683,
+ 0.228032,
+ 0.227543,
+ 0.226565,
+ 0.228032
+ ],
+ "warm_rows_per_second": 2242898.946888854
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 981368,
+ "unique_rows": 3882,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 122,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 122,
+ "zlib_blocks": 122,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3882
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "89f4f162ebd0c7b68354a2308b72e8317a36060395f1556936fc2274ec48ab62",
+ "table_sha256": "14be1afd3c9cf04650f676d4a8a403d19863c5a28aedb7b6799a22533f939520",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10346870590001345,
+ "representatives_sha256": "1861b5297294a7d87f4d7b3d1e73fd35fcb9f66791b206e956f8f266e337f62d",
+ "structurally_merged_rows": 1451,
+ "backend_artifact_bytes": 982569,
+ "backend_with_key_registry_bytes": 1053503,
+ "backend_with_datrie_key_index_bytes": 1025625,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800553769769432,
+ "mean_cosine": 0.9988970173074616,
+ "p05_cosine": 0.9930376882360196,
+ "relative_l2_mean": 0.022494958949884042,
+ "relative_l2_p95": 0.11788731479341936,
+ "relative_l2_max": 0.19902064111331721,
+ "weighted_relative_l2_mean": 0.0087958824132922,
+ "weighted_relative_frobenius": 0.007109698929287029,
+ "weighted_mean_cosine": 0.9996210633671456,
+ "max_relative_norm_change": 0.010144653073261482
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800556007168431,
+ "mean_cosine": 0.9988970348161804,
+ "p05_cosine": 0.9930386649816408,
+ "relative_l2_mean": 0.02234447086639264,
+ "relative_l2_p95": 0.11788397443690124,
+ "relative_l2_max": 0.19901867636766674,
+ "weighted_relative_l2_mean": 0.008614642592702672,
+ "weighted_relative_frobenius": 0.007106717676646518,
+ "weighted_mean_cosine": 0.9996210821053663,
+ "max_relative_norm_change": 0.010084202899738793
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.148693,
+ "warm_p50_ms": 0.228521,
+ "warm_p95_ms": 0.2543934499999999,
+ "warm_p99_ms": 0.29506654,
+ "samples_ms": [
+ 0.239905,
+ 0.231803,
+ 0.229359,
+ 0.233689,
+ 0.232781,
+ 0.299479,
+ 0.288165,
+ 0.22901,
+ 0.227613,
+ 0.22901,
+ 0.228101,
+ 0.228032,
+ 0.227054,
+ 0.230895,
+ 0.229498,
+ 0.230826,
+ 0.229499,
+ 0.228521,
+ 0.227543,
+ 0.230895,
+ 0.231803,
+ 0.22901,
+ 0.252616,
+ 0.232293,
+ 0.22894,
+ 0.228102,
+ 0.227543,
+ 0.228102,
+ 0.228101,
+ 0.226636,
+ 0.227962,
+ 0.227543,
+ 0.228451,
+ 0.227543,
+ 0.227613,
+ 0.228521,
+ 0.227543,
+ 0.228102,
+ 0.227124,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2240494.309056936
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 979908,
+ "unique_rows": 3876,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 122,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 122,
+ "zlib_blocks": 122,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3876
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "206a08fd9da8392d4fc54e6f38a599574975df4d6f5e22ebe4f3be43d722ec68",
+ "table_sha256": "9388381cca8a68f49a1dfe4cce8cc8c12f19e2130a928103b6dd228e5f8070c2",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10100568109191954,
+ "representatives_sha256": "4935ba9823018254815ad15642eaa99c21801c525d4d04b4494377ed343292bf",
+ "structurally_merged_rows": 1457,
+ "backend_artifact_bytes": 981109,
+ "backend_with_key_registry_bytes": 1052043,
+ "backend_with_datrie_key_index_bytes": 1024165,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9745920179413181,
+ "mean_cosine": 0.9988713623700916,
+ "p05_cosine": 0.9929502053254349,
+ "relative_l2_mean": 0.022734702380910424,
+ "relative_l2_p95": 0.11875486523989011,
+ "relative_l2_max": 0.22467937737027865,
+ "weighted_relative_l2_mean": 0.008864115507414503,
+ "weighted_relative_frobenius": 0.007118657593170506,
+ "weighted_mean_cosine": 0.9996137617457883,
+ "max_relative_norm_change": 0.010144653073261482
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9745930835830295,
+ "mean_cosine": 0.9988713824533704,
+ "p05_cosine": 0.9929487131771021,
+ "relative_l2_mean": 0.022584422313542345,
+ "relative_l2_p95": 0.11876937698750527,
+ "relative_l2_max": 0.22467708936171987,
+ "weighted_relative_l2_mean": 0.008682934890047607,
+ "weighted_relative_frobenius": 0.00711567916296217,
+ "weighted_mean_cosine": 0.9996137812167515,
+ "max_relative_norm_change": 0.010084202899738793
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.985963,
+ "warm_p50_ms": 0.228591,
+ "warm_p95_ms": 0.23533834999999992,
+ "warm_p99_ms": 0.26056118,
+ "samples_ms": [
+ 0.232292,
+ 0.234179,
+ 0.232292,
+ 0.229918,
+ 0.22901,
+ 0.22894,
+ 0.22894,
+ 0.229847,
+ 0.229988,
+ 0.229499,
+ 0.22908,
+ 0.257366,
+ 0.2332,
+ 0.228521,
+ 0.226705,
+ 0.228591,
+ 0.228032,
+ 0.227543,
+ 0.227054,
+ 0.226635,
+ 0.228032,
+ 0.227613,
+ 0.228591,
+ 0.228521,
+ 0.226147,
+ 0.22894,
+ 0.229359,
+ 0.232781,
+ 0.262604,
+ 0.232293,
+ 0.227543,
+ 0.227543,
+ 0.229429,
+ 0.227473,
+ 0.228451,
+ 0.226565,
+ 0.226635,
+ 0.228032,
+ 0.226705,
+ 0.228032
+ ],
+ "warm_rows_per_second": 2239808.216421469
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 977365,
+ "unique_rows": 3866,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 121,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 121,
+ "zlib_blocks": 121,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3866
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "7efe602c6977216fc4abf2dca03750a0bebb93ff8e17327e8ee30e95772584a4",
+ "table_sha256": "151081f64e92a899110e41c9aef2788262d113c66230a7bf2a911a08dde0ee29",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10517758410423994,
+ "representatives_sha256": "e2028acf3fb33a1d82863d1469d2e27d2193ec91593a1f620da3f2f4aacda754",
+ "structurally_merged_rows": 1467,
+ "backend_artifact_bytes": 978566,
+ "backend_with_key_registry_bytes": 1049500,
+ "backend_with_datrie_key_index_bytes": 1021622,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9510868764099407,
+ "mean_cosine": 0.9987918573459835,
+ "p05_cosine": 0.9927821767369956,
+ "relative_l2_mean": 0.023299050710395626,
+ "relative_l2_p95": 0.12037584768935825,
+ "relative_l2_max": 0.3115171417508406,
+ "weighted_relative_l2_mean": 0.009270404592701302,
+ "weighted_relative_frobenius": 0.008902606932115491,
+ "weighted_mean_cosine": 0.9995568479061023,
+ "max_relative_norm_change": 0.010144653073261482
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9510758394745452,
+ "mean_cosine": 0.9987918781979689,
+ "p05_cosine": 0.992780348192306,
+ "relative_l2_mean": 0.02314915489090952,
+ "relative_l2_p95": 0.1203808277603667,
+ "relative_l2_max": 0.3115519347010528,
+ "weighted_relative_l2_mean": 0.009089482196330162,
+ "weighted_relative_frobenius": 0.008900056852594328,
+ "weighted_mean_cosine": 0.9995568722623082,
+ "max_relative_norm_change": 0.010084202899738793
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.922966,
+ "warm_p50_ms": 0.229045,
+ "warm_p95_ms": 0.24441994999999997,
+ "warm_p99_ms": 0.27214144,
+ "samples_ms": [
+ 0.233759,
+ 0.232292,
+ 0.231314,
+ 0.230407,
+ 0.229498,
+ 0.229499,
+ 0.22894,
+ 0.229429,
+ 0.227962,
+ 0.22894,
+ 0.22901,
+ 0.229917,
+ 0.228521,
+ 0.228591,
+ 0.286768,
+ 0.244165,
+ 0.229988,
+ 0.229918,
+ 0.233759,
+ 0.228102,
+ 0.228032,
+ 0.227474,
+ 0.228032,
+ 0.228521,
+ 0.228032,
+ 0.227543,
+ 0.228521,
+ 0.228521,
+ 0.227055,
+ 0.22894,
+ 0.228939,
+ 0.231384,
+ 0.249264,
+ 0.230966,
+ 0.23362,
+ 0.232781,
+ 0.229988,
+ 0.22908,
+ 0.226635,
+ 0.227054
+ ],
+ "warm_rows_per_second": 2235368.595690803
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1365248,
+ "artifact_bytes": 976156,
+ "unique_rows": 3861,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 121,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 121,
+ "zlib_blocks": 121,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3861
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "05ff2c9ba69f563e1b4ebe145fcd35517d25e37a623cc252734285239838132c",
+ "table_sha256": "87034ecba5c013dc118a949b28a05e2c518a2c0af2f1834c000c6bf8db772f8f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10555849899537861,
+ "representatives_sha256": "35458ec71543f84bb961a8eb01e723c2636d5a7e881f6ebb28ac2fc7803d893e",
+ "structurally_merged_rows": 1472,
+ "backend_artifact_bytes": 977357,
+ "backend_with_key_registry_bytes": 1048291,
+ "backend_with_datrie_key_index_bytes": 1020413,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9127490251613625,
+ "mean_cosine": 0.9986919996624374,
+ "p05_cosine": 0.9925628236350986,
+ "relative_l2_mean": 0.023820038942016522,
+ "relative_l2_p95": 0.12178604963188834,
+ "relative_l2_max": 0.41871917428766875,
+ "weighted_relative_l2_mean": 0.009659084390710736,
+ "weighted_relative_frobenius": 0.013169823033350842,
+ "weighted_mean_cosine": 0.9994852500344135,
+ "max_relative_norm_change": 0.010144653073261482
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9127456807937895,
+ "mean_cosine": 0.9986920245008379,
+ "p05_cosine": 0.9925641423895385,
+ "relative_l2_mean": 0.023670324181361044,
+ "relative_l2_p95": 0.12176897934386684,
+ "relative_l2_max": 0.41873337973536934,
+ "weighted_relative_l2_mean": 0.009478314565333906,
+ "weighted_relative_frobenius": 0.01316775731103199,
+ "weighted_mean_cosine": 0.9994852786825348,
+ "max_relative_norm_change": 0.010084202899738793
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.896006,
+ "warm_p50_ms": 0.22901,
+ "warm_p95_ms": 0.23533379999999998,
+ "warm_p99_ms": 0.25027555999999995,
+ "samples_ms": [
+ 0.233759,
+ 0.235016,
+ 0.23334,
+ 0.230406,
+ 0.231803,
+ 0.229918,
+ 0.229918,
+ 0.22852,
+ 0.228521,
+ 0.229498,
+ 0.231385,
+ 0.230896,
+ 0.228451,
+ 0.22901,
+ 0.227962,
+ 0.228521,
+ 0.255968,
+ 0.234178,
+ 0.229499,
+ 0.227473,
+ 0.228521,
+ 0.228521,
+ 0.22901,
+ 0.23327,
+ 0.227124,
+ 0.227054,
+ 0.228032,
+ 0.231803,
+ 0.230895,
+ 0.231315,
+ 0.228451,
+ 0.229429,
+ 0.22894,
+ 0.241372,
+ 0.22901,
+ 0.228032,
+ 0.228591,
+ 0.228591,
+ 0.228032,
+ 0.228102
+ ],
+ "warm_rows_per_second": 2235710.2309942795
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 132,
+ "storage": {
+ "artifact_bytes": 703956,
+ "unique_rows": 5333,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "49c40289aa5709af39591af7618b4d2796626664399a25a1da1c9f1231474746",
+ "table_sha256": "39a5bfb46dcda9abae79229e8fb09380005212fafa61770793c05a633311fbe3",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.013785704970359802,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 704803,
+ "backend_with_key_registry_bytes": 775737,
+ "backend_with_datrie_key_index_bytes": 747859,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999462992105821,
+ "mean_cosine": 0.9999792931538555,
+ "p05_cosine": 0.999968339662035,
+ "relative_l2_mean": 0.006403173405174251,
+ "relative_l2_p95": 0.007979164492074683,
+ "relative_l2_max": 0.010381556328689853,
+ "weighted_relative_l2_mean": 0.006403568186156278,
+ "weighted_relative_frobenius": 0.006504947549050006,
+ "weighted_mean_cosine": 0.9999792801327788,
+ "max_relative_norm_change": 0.002201939701089459
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.319874,
+ "warm_p50_ms": 0.122013,
+ "warm_p95_ms": 0.13200005,
+ "warm_p99_ms": 0.13493998,
+ "samples_ms": [
+ 0.12348,
+ 0.122501,
+ 0.12348,
+ 0.121105,
+ 0.121593,
+ 0.122502,
+ 0.122013,
+ 0.132,
+ 0.132001,
+ 0.121524,
+ 0.122013,
+ 0.121035,
+ 0.120127,
+ 0.122012,
+ 0.122152,
+ 0.122153,
+ 0.136819,
+ 0.121594,
+ 0.121105,
+ 0.122502,
+ 0.122083,
+ 0.121105,
+ 0.121594,
+ 0.120616,
+ 0.121175,
+ 0.121105,
+ 0.122083,
+ 0.121105,
+ 0.122502,
+ 0.12299,
+ 0.121105,
+ 0.121104,
+ 0.121175,
+ 0.120266,
+ 0.122502,
+ 0.121594,
+ 0.122571,
+ 0.122572,
+ 0.12299,
+ 0.12299
+ ],
+ "warm_rows_per_second": 4196274.167506741
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 712456,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 87,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 86,
+ "zlib_blocks": 37,
+ "xor_blocks": 49,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "49c40289aa5709af39591af7618b4d2796626664399a25a1da1c9f1231474746",
+ "table_sha256": "67b0076714b293ec1ce3a042ba79a1d455bff8813b1654e3d68f5c7df4f0ea40",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.07455829693935812,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 713700,
+ "backend_with_key_registry_bytes": 784634,
+ "backend_with_datrie_key_index_bytes": 756756,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999462992105821,
+ "mean_cosine": 0.9999792931538555,
+ "p05_cosine": 0.999968339662035,
+ "relative_l2_mean": 0.006403173405174251,
+ "relative_l2_p95": 0.007979164492074683,
+ "relative_l2_max": 0.010381556328689853,
+ "weighted_relative_l2_mean": 0.006403568186156278,
+ "weighted_relative_frobenius": 0.006504947549050006,
+ "weighted_mean_cosine": 0.9999792801327788,
+ "max_relative_norm_change": 0.002201939701089459
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.456565,
+ "warm_p50_ms": 0.1310575,
+ "warm_p95_ms": 0.14261905,
+ "warm_p99_ms": 0.16575328999999997,
+ "samples_ms": [
+ 0.135771,
+ 0.142406,
+ 0.132978,
+ 0.132489,
+ 0.131022,
+ 0.131022,
+ 0.130603,
+ 0.132489,
+ 0.131092,
+ 0.177956,
+ 0.132489,
+ 0.132001,
+ 0.130603,
+ 0.130184,
+ 0.130185,
+ 0.129696,
+ 0.138635,
+ 0.139124,
+ 0.130534,
+ 0.131023,
+ 0.130115,
+ 0.129696,
+ 0.131582,
+ 0.131023,
+ 0.131511,
+ 0.130114,
+ 0.130603,
+ 0.131022,
+ 0.132,
+ 0.129625,
+ 0.131511,
+ 0.131092,
+ 0.130673,
+ 0.130533,
+ 0.134934,
+ 0.129206,
+ 0.129695,
+ 0.134375,
+ 0.146667,
+ 0.131581
+ ],
+ "warm_rows_per_second": 3906682.181485226
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 712456,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 87,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 86,
+ "zlib_blocks": 37,
+ "xor_blocks": 49,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "49c40289aa5709af39591af7618b4d2796626664399a25a1da1c9f1231474746",
+ "table_sha256": "67b0076714b293ec1ce3a042ba79a1d455bff8813b1654e3d68f5c7df4f0ea40",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.09629292599856853,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 713699,
+ "backend_with_key_registry_bytes": 784633,
+ "backend_with_datrie_key_index_bytes": 756755,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999462992105821,
+ "mean_cosine": 0.9999792931538555,
+ "p05_cosine": 0.999968339662035,
+ "relative_l2_mean": 0.006403173405174251,
+ "relative_l2_p95": 0.007979164492074683,
+ "relative_l2_max": 0.010381556328689853,
+ "weighted_relative_l2_mean": 0.006403568186156278,
+ "weighted_relative_frobenius": 0.006504947549050006,
+ "weighted_mean_cosine": 0.9999792801327788,
+ "max_relative_norm_change": 0.002201939701089459
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.487714,
+ "warm_p50_ms": 0.131023,
+ "warm_p95_ms": 0.1367044,
+ "warm_p99_ms": 0.1536563,
+ "samples_ms": [
+ 0.135353,
+ 0.13668,
+ 0.133467,
+ 0.129695,
+ 0.136261,
+ 0.132,
+ 0.130114,
+ 0.132419,
+ 0.131511,
+ 0.131093,
+ 0.132908,
+ 0.137168,
+ 0.130114,
+ 0.131022,
+ 0.131022,
+ 0.130533,
+ 0.130603,
+ 0.130603,
+ 0.130533,
+ 0.132,
+ 0.129626,
+ 0.164198,
+ 0.132,
+ 0.131511,
+ 0.131511,
+ 0.131581,
+ 0.132,
+ 0.130604,
+ 0.130115,
+ 0.131023,
+ 0.130115,
+ 0.130603,
+ 0.130603,
+ 0.130045,
+ 0.131512,
+ 0.130185,
+ 0.130604,
+ 0.131023,
+ 0.130115,
+ 0.131023
+ ],
+ "warm_rows_per_second": 3907710.8599253567
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 712456,
+ "unique_rows": 5333,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 87,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 86,
+ "zlib_blocks": 37,
+ "xor_blocks": 49,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 5333
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "49c40289aa5709af39591af7618b4d2796626664399a25a1da1c9f1231474746",
+ "table_sha256": "67b0076714b293ec1ce3a042ba79a1d455bff8813b1654e3d68f5c7df4f0ea40",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.07648577890358865,
+ "representatives_sha256": "b87cb171f9ce05babadb9abe917f696021522d5fafe1c497d777370a59cf220d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 713699,
+ "backend_with_key_registry_bytes": 784633,
+ "backend_with_datrie_key_index_bytes": 756755,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999462992105821,
+ "mean_cosine": 0.9999792931538555,
+ "p05_cosine": 0.999968339662035,
+ "relative_l2_mean": 0.006403173405174251,
+ "relative_l2_p95": 0.007979164492074683,
+ "relative_l2_max": 0.010381556328689853,
+ "weighted_relative_l2_mean": 0.006403568186156278,
+ "weighted_relative_frobenius": 0.006504947549050006,
+ "weighted_mean_cosine": 0.9999792801327788,
+ "max_relative_norm_change": 0.002201939701089459
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.521378,
+ "warm_p50_ms": 0.131092,
+ "warm_p95_ms": 0.1372169,
+ "warm_p99_ms": 0.15403710999999998,
+ "samples_ms": [
+ 0.138146,
+ 0.134863,
+ 0.132,
+ 0.132,
+ 0.130533,
+ 0.131511,
+ 0.131093,
+ 0.130534,
+ 0.135283,
+ 0.132489,
+ 0.133886,
+ 0.132489,
+ 0.130533,
+ 0.131092,
+ 0.131022,
+ 0.13633,
+ 0.130603,
+ 0.131022,
+ 0.131511,
+ 0.132001,
+ 0.131512,
+ 0.130115,
+ 0.130115,
+ 0.130604,
+ 0.164197,
+ 0.137168,
+ 0.130603,
+ 0.130114,
+ 0.130533,
+ 0.130044,
+ 0.129137,
+ 0.130534,
+ 0.129137,
+ 0.128718,
+ 0.130114,
+ 0.130603,
+ 0.131581,
+ 0.13633,
+ 0.131092,
+ 0.132
+ ],
+ "warm_rows_per_second": 3905654.0444878405
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 549156,
+ "unique_rows": 3998,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 65,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 65,
+ "zlib_blocks": 65,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3998
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "aae123abfc4f71a68e7ec0af3535ffc6617e0834f1e0b1028025359d5084208c",
+ "table_sha256": "3c28cf8d5727899c9f383550f7241c82ac43bd61e46d4d4251dd8252acb30a83",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06409340887330472,
+ "representatives_sha256": "4b481ab0d27e04463dcd26f8162ab7756547ef8ca42c11f8f54d90b15ebf1883",
+ "structurally_merged_rows": 1335,
+ "backend_artifact_bytes": 550397,
+ "backend_with_key_registry_bytes": 621331,
+ "backend_with_datrie_key_index_bytes": 593453,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9899934222289732,
+ "mean_cosine": 0.9991632929825155,
+ "p05_cosine": 0.9947887591206397,
+ "relative_l2_mean": 0.023713495632551772,
+ "relative_l2_p95": 0.10212735134839708,
+ "relative_l2_max": 0.1417056444563921,
+ "weighted_relative_l2_mean": 0.013278828741218295,
+ "weighted_relative_frobenius": 0.00940051238979808,
+ "weighted_mean_cosine": 0.9996842272959379,
+ "max_relative_norm_change": 0.01097164479864802
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9900189817284225,
+ "mean_cosine": 0.9991737332185844,
+ "p05_cosine": 0.9947844121822593,
+ "relative_l2_mean": 0.01899638009320118,
+ "relative_l2_p95": 0.10218295670728121,
+ "relative_l2_max": 0.14142651092374273,
+ "weighted_relative_l2_mean": 0.00767327899498125,
+ "weighted_relative_frobenius": 0.006914346925094502,
+ "weighted_mean_cosine": 0.999700063737909,
+ "max_relative_norm_change": 0.010031623350154936
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.264164,
+ "warm_p50_ms": 0.131512,
+ "warm_p95_ms": 0.13653605,
+ "warm_p99_ms": 0.14164416,
+ "samples_ms": [
+ 0.140451,
+ 0.135772,
+ 0.133956,
+ 0.133816,
+ 0.13633,
+ 0.132908,
+ 0.132978,
+ 0.130603,
+ 0.131022,
+ 0.131512,
+ 0.132908,
+ 0.132978,
+ 0.132,
+ 0.130604,
+ 0.131512,
+ 0.131512,
+ 0.131022,
+ 0.132,
+ 0.131022,
+ 0.131092,
+ 0.132,
+ 0.131023,
+ 0.131511,
+ 0.132489,
+ 0.131092,
+ 0.131092,
+ 0.13626,
+ 0.132001,
+ 0.142407,
+ 0.131512,
+ 0.130604,
+ 0.131163,
+ 0.132489,
+ 0.131022,
+ 0.131022,
+ 0.130184,
+ 0.131092,
+ 0.129136,
+ 0.130533,
+ 0.130114
+ ],
+ "warm_rows_per_second": 3893180.850416692
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 534807,
+ "unique_rows": 3885,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3885
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "28ac3a00592d6854a1800e8d1c425a58515b10fe0f050b78fce37d6a702884c2",
+ "table_sha256": "1683f4fcdeece02634106cbcf8e679bf09d615129da3444cdb90789326bbb95d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.0614638130646199,
+ "representatives_sha256": "e2ff6c1c1b12ba8084d1d0766d5491ea4619d8c7f76acb95dab2e151b9bc08ca",
+ "structurally_merged_rows": 1448,
+ "backend_artifact_bytes": 536048,
+ "backend_with_key_registry_bytes": 606982,
+ "backend_with_datrie_key_index_bytes": 579104,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800401744851072,
+ "mean_cosine": 0.9988806200285881,
+ "p05_cosine": 0.9930299198404723,
+ "relative_l2_mean": 0.02703043177233258,
+ "relative_l2_p95": 0.11799639536476662,
+ "relative_l2_max": 0.19908866256523516,
+ "weighted_relative_l2_mean": 0.014238363469704323,
+ "weighted_relative_frobenius": 0.009627358087898327,
+ "weighted_mean_cosine": 0.99960255758857,
+ "max_relative_norm_change": 0.01097164479864802
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9800988294273547,
+ "mean_cosine": 0.998890430134469,
+ "p05_cosine": 0.9929888872159995,
+ "relative_l2_mean": 0.02245089322008093,
+ "relative_l2_p95": 0.11850476828542113,
+ "relative_l2_max": 0.1987932467735768,
+ "weighted_relative_l2_mean": 0.008672756110095822,
+ "weighted_relative_frobenius": 0.007221971183909667,
+ "weighted_mean_cosine": 0.9996181977174599,
+ "max_relative_norm_change": 0.010031623350154936
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.222119,
+ "warm_p50_ms": 0.1322095,
+ "warm_p95_ms": 0.1349094,
+ "warm_p99_ms": 0.15426171,
+ "samples_ms": [
+ 0.134863,
+ 0.132908,
+ 0.134374,
+ 0.132,
+ 0.132489,
+ 0.132489,
+ 0.130603,
+ 0.135772,
+ 0.132978,
+ 0.134864,
+ 0.132419,
+ 0.130184,
+ 0.132978,
+ 0.132,
+ 0.131093,
+ 0.130534,
+ 0.130603,
+ 0.133886,
+ 0.132,
+ 0.131023,
+ 0.131512,
+ 0.131023,
+ 0.131022,
+ 0.134375,
+ 0.133886,
+ 0.131512,
+ 0.134794,
+ 0.134444,
+ 0.166083,
+ 0.133467,
+ 0.131162,
+ 0.132489,
+ 0.133956,
+ 0.131092,
+ 0.130533,
+ 0.130603,
+ 0.130603,
+ 0.130603,
+ 0.129625,
+ 0.132978
+ ],
+ "warm_rows_per_second": 3872641.527273002
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 534042,
+ "unique_rows": 3879,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3879
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "f69b24dbe40c19ba537d061ae1827fb052af652887b23876ca007bd0447136e5",
+ "table_sha256": "c1d551af235707630b731567b459b997766dc2e7c68dfdc452aafc7438ffc103",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06319161807186902,
+ "representatives_sha256": "dcec11d97205c1f7cd9e55c9be1a1ddfced8d21268c04e7cb3c02dd07fa9e577",
+ "structurally_merged_rows": 1454,
+ "backend_artifact_bytes": 535283,
+ "backend_with_key_registry_bytes": 606217,
+ "backend_with_datrie_key_index_bytes": 578339,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9746593100882213,
+ "mean_cosine": 0.9988549663087168,
+ "p05_cosine": 0.9929140091542047,
+ "relative_l2_mean": 0.027263932824013213,
+ "relative_l2_p95": 0.11901936063559732,
+ "relative_l2_max": 0.22429710615850762,
+ "weighted_relative_l2_mean": 0.014304819927630079,
+ "weighted_relative_frobenius": 0.009633974231074247,
+ "weighted_mean_cosine": 0.9995952563137235,
+ "max_relative_norm_change": 0.01097164479864802
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9747256332809409,
+ "mean_cosine": 0.9988647192842288,
+ "p05_cosine": 0.9929013771709432,
+ "relative_l2_mean": 0.02269106920580281,
+ "relative_l2_p95": 0.11919248437636946,
+ "relative_l2_max": 0.22395884382247594,
+ "weighted_relative_l2_mean": 0.008741112313097998,
+ "weighted_relative_frobenius": 0.007230809611545858,
+ "weighted_mean_cosine": 0.9996108801828067,
+ "max_relative_norm_change": 0.010031623350154936
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.142081,
+ "warm_p50_ms": 0.132245,
+ "warm_p95_ms": 0.13864179999999998,
+ "warm_p99_ms": 0.15049662,
+ "samples_ms": [
+ 0.140101,
+ 0.136261,
+ 0.138565,
+ 0.134724,
+ 0.135283,
+ 0.133467,
+ 0.133886,
+ 0.133467,
+ 0.13633,
+ 0.135212,
+ 0.135702,
+ 0.134864,
+ 0.13668,
+ 0.132909,
+ 0.132489,
+ 0.157143,
+ 0.134863,
+ 0.132,
+ 0.132,
+ 0.130115,
+ 0.132001,
+ 0.131093,
+ 0.132,
+ 0.130603,
+ 0.132,
+ 0.134375,
+ 0.132,
+ 0.131022,
+ 0.132977,
+ 0.132,
+ 0.133397,
+ 0.131022,
+ 0.130114,
+ 0.131512,
+ 0.130604,
+ 0.131512,
+ 0.131511,
+ 0.130603,
+ 0.131581,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3871601.9509244207
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 532560,
+ "unique_rows": 3867,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 63,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3867
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "7b8a07f1b071e86d8514bb4249c7eeec40dc7ea14c98d8e8970c4c4a795f7c1a",
+ "table_sha256": "7a360b583fe1f338120c00165101d62a4216fde4a27a6198320713e01225ca80",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06164735509082675,
+ "representatives_sha256": "6640075131d383beb880b09f5208eb1b47e3ca8fcd0b3886ec27fcbe6b7d1ab8",
+ "structurally_merged_rows": 1466,
+ "backend_artifact_bytes": 533801,
+ "backend_with_key_registry_bytes": 604735,
+ "backend_with_datrie_key_index_bytes": 576857,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.9509876848478194,
+ "mean_cosine": 0.9987614785689874,
+ "p05_cosine": 0.9927806778831033,
+ "relative_l2_mean": 0.027916173556634682,
+ "relative_l2_p95": 0.1201832795122255,
+ "relative_l2_max": 0.311851238175594,
+ "weighted_relative_l2_mean": 0.014731416817497978,
+ "weighted_relative_frobenius": 0.011073839881895566,
+ "weighted_mean_cosine": 0.9995343569310857,
+ "max_relative_norm_change": 0.01097164479864802
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9509615279952359,
+ "mean_cosine": 0.9987712563715393,
+ "p05_cosine": 0.9927357309980902,
+ "relative_l2_mean": 0.023357600049515495,
+ "relative_l2_p95": 0.12065559752209504,
+ "relative_l2_max": 0.31193165760070524,
+ "weighted_relative_l2_mean": 0.009175777532520312,
+ "weighted_relative_frobenius": 0.009049998438847075,
+ "weighted_mean_cosine": 0.99955017708866,
+ "max_relative_norm_change": 0.010031623350154936
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.187129,
+ "warm_p50_ms": 0.1336415,
+ "warm_p95_ms": 0.1372834,
+ "warm_p99_ms": 0.14449426999999998,
+ "samples_ms": [
+ 0.137168,
+ 0.133397,
+ 0.137238,
+ 0.134794,
+ 0.133466,
+ 0.133886,
+ 0.132489,
+ 0.132419,
+ 0.135283,
+ 0.138146,
+ 0.135702,
+ 0.134864,
+ 0.132489,
+ 0.135772,
+ 0.132419,
+ 0.148553,
+ 0.133816,
+ 0.133467,
+ 0.132001,
+ 0.132489,
+ 0.135353,
+ 0.134864,
+ 0.135282,
+ 0.132978,
+ 0.131092,
+ 0.132489,
+ 0.132978,
+ 0.133397,
+ 0.135283,
+ 0.134864,
+ 0.133816,
+ 0.133397,
+ 0.135771,
+ 0.132978,
+ 0.132908,
+ 0.13675,
+ 0.134864,
+ 0.13242,
+ 0.131581,
+ 0.131511
+ ],
+ "warm_rows_per_second": 3831145.2655051015
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 5333,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 703956,
+ "artifact_bytes": 531805,
+ "unique_rows": 3861,
+ "row_map_bytes": 42664,
+ "n_rows": 5333,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 63,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 63,
+ "zlib_blocks": 62,
+ "xor_blocks": 1,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 3861
+ },
+ "source_float32_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966",
+ "decoded_float32_sha256": "e824ba3aae4a0129d26cdbf7d29d7fe22d50454ef80c15da374388e67f0bc0b7",
+ "table_sha256": "3864b7a9fc3c2be506a9fca7d7131b003f94db0fbbe851b34b63b60305fdec1b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.0668308399617672,
+ "representatives_sha256": "f50613b9c916ca0c6b2ba4bcb1489ab2196e96f509a08f96ddfa1b4f7b6941e1",
+ "structurally_merged_rows": 1472,
+ "backend_artifact_bytes": 533046,
+ "backend_with_key_registry_bytes": 603980,
+ "backend_with_datrie_key_index_bytes": 576102,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 5333,
+ "minimum_cosine": 0.8996025209240703,
+ "mean_cosine": 0.9986426261248615,
+ "p05_cosine": 0.9925659301041087,
+ "relative_l2_mean": 0.02851507216503056,
+ "relative_l2_p95": 0.12180405322443522,
+ "relative_l2_max": 0.450476772178466,
+ "weighted_relative_l2_mean": 0.015210267558143175,
+ "weighted_relative_frobenius": 0.014919251451837928,
+ "weighted_mean_cosine": 0.999441083525298,
+ "max_relative_norm_change": 0.01097164479864802
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 5333,
+ "minimum_cosine": 0.9001643705394677,
+ "mean_cosine": 0.9986525094715527,
+ "p05_cosine": 0.9925633786418048,
+ "relative_l2_mean": 0.023963313369431437,
+ "relative_l2_p95": 0.12183598817005416,
+ "relative_l2_max": 0.44917041932412277,
+ "weighted_relative_l2_mean": 0.009661088397133939,
+ "weighted_relative_frobenius": 0.013479473656458102,
+ "weighted_mean_cosine": 0.9994570772808833,
+ "max_relative_norm_change": 0.010031623350154936
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.178119,
+ "warm_p50_ms": 0.132489,
+ "warm_p95_ms": 0.13957354999999996,
+ "warm_p99_ms": 0.15823072,
+ "samples_ms": [
+ 0.135702,
+ 0.133397,
+ 0.133397,
+ 0.132978,
+ 0.131511,
+ 0.131511,
+ 0.148134,
+ 0.137168,
+ 0.139123,
+ 0.132908,
+ 0.133956,
+ 0.133816,
+ 0.132489,
+ 0.131581,
+ 0.133467,
+ 0.131512,
+ 0.133467,
+ 0.132489,
+ 0.134375,
+ 0.134374,
+ 0.131511,
+ 0.131511,
+ 0.135282,
+ 0.132489,
+ 0.132489,
+ 0.131022,
+ 0.130533,
+ 0.132489,
+ 0.130603,
+ 0.131092,
+ 0.130114,
+ 0.132001,
+ 0.133816,
+ 0.131512,
+ 0.132,
+ 0.164686,
+ 0.134794,
+ 0.133397,
+ 0.132419,
+ 0.130603
+ ],
+ "warm_rows_per_second": 3864471.7674674885
+ }
+ }
+ ],
+ "baseline_quality": {},
+ "training_row_sha256": "36ddfdc77bbd89e4a6d29ace0eb8d7434db7c6aea2f962b09a109ebd23d40966"
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 3,
+ "rows": 13424,
+ "dim": 128,
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "source_float32_bytes": 6873088,
+ "key_registry_bytes": 203948,
+ "unchanged_backbone_artifact_bytes": 0,
+ "alternative_datrie_key_index_bytes": 108592,
+ "query_rows": 512,
+ "query_ids_sha256": "80ab83de1ecc982405671e980d9e9a66070e51aead43121126faabb6b44a50cd",
+ "frequency_trace_sha256": "1a0389197253d049e85973cd9812529ea0dc3eae69dcdb62085a34b49c0d90b8",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 512,
+ "storage": {
+ "artifact_bytes": 6873088,
+ "unique_rows": 13424,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "table_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.036947192158550024,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6873893,
+ "backend_with_key_registry_bytes": 7077841,
+ "backend_with_datrie_key_index_bytes": 6982485,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.363315,
+ "warm_p50_ms": 0.060553,
+ "warm_p95_ms": 0.06190695,
+ "warm_p99_ms": 0.07662476999999998,
+ "samples_ms": [
+ 0.061879,
+ 0.060553,
+ 0.085695,
+ 0.060063,
+ 0.060063,
+ 0.060063,
+ 0.06146,
+ 0.062438,
+ 0.060552,
+ 0.060971,
+ 0.061111,
+ 0.06146,
+ 0.060063,
+ 0.060134,
+ 0.060553,
+ 0.059715,
+ 0.06153,
+ 0.060553,
+ 0.061041,
+ 0.059644,
+ 0.06146,
+ 0.061041,
+ 0.060552,
+ 0.059085,
+ 0.059575,
+ 0.060483,
+ 0.0616,
+ 0.060553,
+ 0.061042,
+ 0.060623,
+ 0.060553,
+ 0.060134,
+ 0.060063,
+ 0.060553,
+ 0.061042,
+ 0.060553,
+ 0.060553,
+ 0.0616,
+ 0.059644,
+ 0.060901
+ ],
+ "warm_rows_per_second": 8455402.705068287
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 6480560,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 839,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 839,
+ "zlib_blocks": 56,
+ "xor_blocks": 783,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "table_sha256": "1c958153046d8ba8ddd207bf7597464d496d4246381783d9c4446b6e678e8de0",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6004768090788275,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6481769,
+ "backend_with_key_registry_bytes": 6685717,
+ "backend_with_datrie_key_index_bytes": 6590361,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.405151,
+ "warm_p50_ms": 0.0724255,
+ "warm_p95_ms": 0.0771983,
+ "warm_p99_ms": 0.09488110999999999,
+ "samples_ms": [
+ 0.07899,
+ 0.076756,
+ 0.075708,
+ 0.105041,
+ 0.077104,
+ 0.075289,
+ 0.07473,
+ 0.075289,
+ 0.074312,
+ 0.074241,
+ 0.074311,
+ 0.075219,
+ 0.071936,
+ 0.071867,
+ 0.072426,
+ 0.072495,
+ 0.072006,
+ 0.071448,
+ 0.07047,
+ 0.073822,
+ 0.071447,
+ 0.071518,
+ 0.071029,
+ 0.071867,
+ 0.072356,
+ 0.072426,
+ 0.072426,
+ 0.072425,
+ 0.072495,
+ 0.071448,
+ 0.071937,
+ 0.071028,
+ 0.072915,
+ 0.072356,
+ 0.071937,
+ 0.072006,
+ 0.07054,
+ 0.071517,
+ 0.07054,
+ 0.076127
+ ],
+ "warm_rows_per_second": 7069333.314923611
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 6480560,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 839,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 839,
+ "zlib_blocks": 56,
+ "xor_blocks": 783,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "table_sha256": "1c958153046d8ba8ddd207bf7597464d496d4246381783d9c4446b6e678e8de0",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6389819530304521,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6481768,
+ "backend_with_key_registry_bytes": 6685716,
+ "backend_with_datrie_key_index_bytes": 6590360,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.377843,
+ "warm_p50_ms": 0.072006,
+ "warm_p95_ms": 0.07664395,
+ "warm_p99_ms": 0.09272450999999998,
+ "samples_ms": [
+ 0.077175,
+ 0.075219,
+ 0.075288,
+ 0.072426,
+ 0.074312,
+ 0.0748,
+ 0.07473,
+ 0.074241,
+ 0.074311,
+ 0.07473,
+ 0.074242,
+ 0.074242,
+ 0.072356,
+ 0.071797,
+ 0.072355,
+ 0.071866,
+ 0.070889,
+ 0.102666,
+ 0.072356,
+ 0.071937,
+ 0.070051,
+ 0.072426,
+ 0.070051,
+ 0.072006,
+ 0.069981,
+ 0.071866,
+ 0.076616,
+ 0.070051,
+ 0.071448,
+ 0.070051,
+ 0.072006,
+ 0.071448,
+ 0.07047,
+ 0.069911,
+ 0.071867,
+ 0.072355,
+ 0.069981,
+ 0.071867,
+ 0.071447,
+ 0.069632
+ ],
+ "warm_rows_per_second": 7110518.56789712
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 6480560,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 839,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 839,
+ "zlib_blocks": 56,
+ "xor_blocks": 783,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "table_sha256": "1c958153046d8ba8ddd207bf7597464d496d4246381783d9c4446b6e678e8de0",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5854424799326807,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6481768,
+ "backend_with_key_registry_bytes": 6685716,
+ "backend_with_datrie_key_index_bytes": 6590360,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.37484,
+ "warm_p50_ms": 0.071867,
+ "warm_p95_ms": 0.07780369999999999,
+ "warm_p99_ms": 0.08846669,
+ "samples_ms": [
+ 0.077664,
+ 0.080458,
+ 0.074311,
+ 0.07473,
+ 0.093587,
+ 0.077174,
+ 0.075219,
+ 0.071937,
+ 0.072356,
+ 0.073823,
+ 0.073334,
+ 0.073823,
+ 0.072006,
+ 0.072844,
+ 0.071517,
+ 0.073892,
+ 0.072426,
+ 0.071867,
+ 0.071936,
+ 0.072426,
+ 0.070959,
+ 0.071028,
+ 0.070051,
+ 0.071448,
+ 0.070051,
+ 0.071867,
+ 0.071448,
+ 0.069981,
+ 0.0704,
+ 0.069981,
+ 0.071028,
+ 0.0704,
+ 0.071377,
+ 0.07047,
+ 0.070889,
+ 0.071448,
+ 0.07054,
+ 0.071028,
+ 0.07054,
+ 0.073822
+ ],
+ "warm_rows_per_second": 7124271.223231804
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 4096126,
+ "unique_rows": 8305,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 520,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 519,
+ "zlib_blocks": 370,
+ "xor_blocks": 149,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8305
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "dc605d0a3b37e596c145e6d11c388507ab6c4bdbbea2a957dd48b38b2c13cc64",
+ "table_sha256": "458ad7e65187ad8d1b501429bd99d2ca31f6c807f52ba83e2d7d0c5a23d46755",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.37950568506494164,
+ "representatives_sha256": "25fa7cbbb9d1e2e5d9cc53c8fe066b825fab9b925285db1c3e5875fab20eaa86",
+ "structurally_merged_rows": 5119,
+ "backend_artifact_bytes": 4097333,
+ "backend_with_key_registry_bytes": 4301281,
+ "backend_with_datrie_key_index_bytes": 4205925,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900126594367357,
+ "mean_cosine": 0.9985937157194625,
+ "p05_cosine": 0.9930711762913733,
+ "relative_l2_mean": 0.03088048812596691,
+ "relative_l2_p95": 0.11792570917396121,
+ "relative_l2_max": 0.14211742856778714,
+ "weighted_relative_l2_mean": 0.017521503643948114,
+ "weighted_relative_frobenius": 0.009587863324407826,
+ "weighted_mean_cosine": 0.9992425122949266,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900126594367357,
+ "mean_cosine": 0.9985937157194625,
+ "p05_cosine": 0.9930711762913733,
+ "relative_l2_mean": 0.03088048812596691,
+ "relative_l2_p95": 0.11792570917396121,
+ "relative_l2_max": 0.14211742856778714,
+ "weighted_relative_l2_mean": 0.017521503643948114,
+ "weighted_relative_frobenius": 0.009587863324407826,
+ "weighted_mean_cosine": 0.9992425122949266,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.31445,
+ "warm_p50_ms": 0.069632,
+ "warm_p95_ms": 0.07613045,
+ "warm_p99_ms": 0.09212980999999999,
+ "samples_ms": [
+ 0.076196,
+ 0.073333,
+ 0.102317,
+ 0.076127,
+ 0.075219,
+ 0.076127,
+ 0.073823,
+ 0.073264,
+ 0.072356,
+ 0.072915,
+ 0.072425,
+ 0.072844,
+ 0.070121,
+ 0.07054,
+ 0.069981,
+ 0.068166,
+ 0.070051,
+ 0.071028,
+ 0.069073,
+ 0.069003,
+ 0.069073,
+ 0.069073,
+ 0.069562,
+ 0.069143,
+ 0.069074,
+ 0.069003,
+ 0.069632,
+ 0.069074,
+ 0.069073,
+ 0.069562,
+ 0.069004,
+ 0.070051,
+ 0.069632,
+ 0.068165,
+ 0.068584,
+ 0.069073,
+ 0.069143,
+ 0.069003,
+ 0.069492,
+ 0.07012
+ ],
+ "warm_rows_per_second": 7352941.176470588
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3847377,
+ "unique_rows": 7785,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 487,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 487,
+ "zlib_blocks": 388,
+ "xor_blocks": 99,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7785
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "12223f8457285e033aa2aff4b723c3ca69d08120454890e32e1f1342c7859782",
+ "table_sha256": "2739ed9bd81b3f30b6293fd71b0cbcebf573f5ad5025994d16110cca7cdd7fac",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3562793149612844,
+ "representatives_sha256": "51abef867e22e4b21b773a2aaf8f2176e35911b200214cc09b84b2ff51efb3a1",
+ "structurally_merged_rows": 5639,
+ "backend_artifact_bytes": 3848583,
+ "backend_with_key_registry_bytes": 4052531,
+ "backend_with_datrie_key_index_bytes": 3957175,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9801116849489457,
+ "mean_cosine": 0.9980497479485829,
+ "p05_cosine": 0.9906083903349477,
+ "relative_l2_mean": 0.03735267712132519,
+ "relative_l2_p95": 0.13703607338904397,
+ "relative_l2_max": 0.20024584710789572,
+ "weighted_relative_l2_mean": 0.020662773018602778,
+ "weighted_relative_frobenius": 0.010050265018434344,
+ "weighted_mean_cosine": 0.9989785122367599,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9801116849489457,
+ "mean_cosine": 0.9980497479485829,
+ "p05_cosine": 0.9906083903349477,
+ "relative_l2_mean": 0.03735267712132519,
+ "relative_l2_p95": 0.13703607338904397,
+ "relative_l2_max": 0.20024584710789572,
+ "weighted_relative_l2_mean": 0.020662773018602778,
+ "weighted_relative_frobenius": 0.010050265018434344,
+ "weighted_mean_cosine": 0.9989785122367599,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.490381,
+ "warm_p50_ms": 0.069981,
+ "warm_p95_ms": 0.07344919999999999,
+ "warm_p99_ms": 0.0756807,
+ "samples_ms": [
+ 0.075638,
+ 0.073334,
+ 0.075708,
+ 0.072845,
+ 0.072845,
+ 0.072845,
+ 0.072356,
+ 0.07047,
+ 0.071448,
+ 0.071377,
+ 0.071378,
+ 0.071448,
+ 0.070539,
+ 0.071029,
+ 0.071028,
+ 0.07054,
+ 0.068585,
+ 0.069562,
+ 0.069003,
+ 0.069003,
+ 0.069143,
+ 0.069073,
+ 0.069003,
+ 0.069074,
+ 0.069073,
+ 0.068584,
+ 0.069073,
+ 0.068096,
+ 0.068654,
+ 0.069073,
+ 0.069981,
+ 0.069492,
+ 0.069562,
+ 0.070539,
+ 0.07047,
+ 0.069492,
+ 0.07054,
+ 0.069981,
+ 0.069981,
+ 0.069492
+ ],
+ "warm_rows_per_second": 7316271.559423272
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3814746,
+ "unique_rows": 7716,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 483,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 483,
+ "zlib_blocks": 383,
+ "xor_blocks": 100,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7716
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "c1a3d5cd5433063d3d18270293cda41808482290a5b8ce0cfb9919b812db7933",
+ "table_sha256": "8aab45d929724120ca9235ab2e8a16e5676eb92a106e183c7686cc680cf24f15",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3821442211046815,
+ "representatives_sha256": "571db57df46df6ae8c314e1e6b2e66dea12151fbaf5130c80f961307e39296da",
+ "structurally_merged_rows": 5708,
+ "backend_artifact_bytes": 3815953,
+ "backend_with_key_registry_bytes": 4019901,
+ "backend_with_datrie_key_index_bytes": 3924545,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.970343081661601,
+ "mean_cosine": 0.9979239946075346,
+ "p05_cosine": 0.9900978443322698,
+ "relative_l2_mean": 0.03848967424519778,
+ "relative_l2_p95": 0.14086231295128412,
+ "relative_l2_max": 0.24497659089464754,
+ "weighted_relative_l2_mean": 0.021211546951327244,
+ "weighted_relative_frobenius": 0.010113759652953795,
+ "weighted_mean_cosine": 0.99891781713554,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.970343081661601,
+ "mean_cosine": 0.9979239946075346,
+ "p05_cosine": 0.9900978443322698,
+ "relative_l2_mean": 0.03848967424519778,
+ "relative_l2_p95": 0.14086231295128412,
+ "relative_l2_max": 0.24497659089464754,
+ "weighted_relative_l2_mean": 0.021211546951327244,
+ "weighted_relative_frobenius": 0.010113759652953795,
+ "weighted_mean_cosine": 0.99891781713554,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.277993,
+ "warm_p50_ms": 0.0710285,
+ "warm_p95_ms": 0.0747789,
+ "warm_p99_ms": 0.08895780999999998,
+ "samples_ms": [
+ 0.075708,
+ 0.07473,
+ 0.07473,
+ 0.072844,
+ 0.072355,
+ 0.072355,
+ 0.071937,
+ 0.072356,
+ 0.071936,
+ 0.072006,
+ 0.072495,
+ 0.074241,
+ 0.071937,
+ 0.071029,
+ 0.071028,
+ 0.070121,
+ 0.071028,
+ 0.071378,
+ 0.071378,
+ 0.070889,
+ 0.069981,
+ 0.069981,
+ 0.069632,
+ 0.069981,
+ 0.069981,
+ 0.071867,
+ 0.069981,
+ 0.070051,
+ 0.070051,
+ 0.069492,
+ 0.097429,
+ 0.072425,
+ 0.070051,
+ 0.069492,
+ 0.069981,
+ 0.071029,
+ 0.070051,
+ 0.070959,
+ 0.069492,
+ 0.069911
+ ],
+ "warm_rows_per_second": 7208374.10335288
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3804432,
+ "unique_rows": 7696,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 481,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 481,
+ "zlib_blocks": 381,
+ "xor_blocks": 100,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7696
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "c7f75a54b1f5d74974e2b6334235f06471c9bb162f668cde6cd109269e64a611",
+ "table_sha256": "59e5334f1e93fe7fedf5cec6d9222c38ee6cae22ff80ee758577ba9c84647f46",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3475266571622342,
+ "representatives_sha256": "2e9e6d353edbf6a33317e108cefb746a100f4890e40169f11fd6c96afef18d88",
+ "structurally_merged_rows": 5728,
+ "backend_artifact_bytes": 3805639,
+ "backend_with_key_registry_bytes": 4009587,
+ "backend_with_datrie_key_index_bytes": 3914231,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9515267263969356,
+ "mean_cosine": 0.9978276087481274,
+ "p05_cosine": 0.989916908259933,
+ "relative_l2_mean": 0.03913300939645236,
+ "relative_l2_p95": 0.14200964850446474,
+ "relative_l2_max": 0.31035210487769166,
+ "weighted_relative_l2_mean": 0.0215610588496456,
+ "weighted_relative_frobenius": 0.011062901641116845,
+ "weighted_mean_cosine": 0.9988659972406398,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9515267263969356,
+ "mean_cosine": 0.9978276087481274,
+ "p05_cosine": 0.989916908259933,
+ "relative_l2_mean": 0.03913300939645236,
+ "relative_l2_p95": 0.14200964850446474,
+ "relative_l2_max": 0.31035210487769166,
+ "weighted_relative_l2_mean": 0.0215610588496456,
+ "weighted_relative_frobenius": 0.011062901641116845,
+ "weighted_mean_cosine": 0.9988659972406398,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.476272,
+ "warm_p50_ms": 0.070435,
+ "warm_p95_ms": 0.07289389999999998,
+ "warm_p99_ms": 0.09405913999999999,
+ "samples_ms": [
+ 0.073823,
+ 0.071378,
+ 0.072425,
+ 0.071866,
+ 0.071448,
+ 0.072356,
+ 0.072845,
+ 0.07047,
+ 0.071937,
+ 0.072845,
+ 0.072356,
+ 0.071378,
+ 0.070889,
+ 0.071448,
+ 0.069143,
+ 0.069073,
+ 0.106997,
+ 0.070051,
+ 0.07047,
+ 0.068165,
+ 0.070051,
+ 0.07047,
+ 0.0704,
+ 0.069981,
+ 0.069632,
+ 0.069073,
+ 0.069911,
+ 0.068235,
+ 0.068654,
+ 0.070889,
+ 0.070959,
+ 0.069632,
+ 0.070959,
+ 0.069562,
+ 0.068095,
+ 0.069143,
+ 0.069562,
+ 0.069492,
+ 0.070121,
+ 0.069562
+ ],
+ "warm_rows_per_second": 7269113.366934053
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3798092,
+ "unique_rows": 7682,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 481,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 481,
+ "zlib_blocks": 391,
+ "xor_blocks": 90,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7682
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "24bed3551c44c8f6a7587e26c7fa086581fe8706a07bb5672f4f3aa298697be7",
+ "table_sha256": "8d1a5bb633ab680ffcbb3fa549a2f65ede3aa2d56cee75db9fe171ff5c96dc89",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4390689739957452,
+ "representatives_sha256": "02853bc63111f33dfd64232ee44c0a3dac0182bc7fffccb4618903cecbcd5a0f",
+ "structurally_merged_rows": 5742,
+ "backend_artifact_bytes": 3799298,
+ "backend_with_key_registry_bytes": 4003246,
+ "backend_with_datrie_key_index_bytes": 3907890,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.901636957934259,
+ "mean_cosine": 0.9976914672684307,
+ "p05_cosine": 0.9897294932832881,
+ "relative_l2_mean": 0.03979940049850905,
+ "relative_l2_p95": 0.14344509893530952,
+ "relative_l2_max": 0.44557719917122357,
+ "weighted_relative_l2_mean": 0.022108816306860343,
+ "weighted_relative_frobenius": 0.01276195516775435,
+ "weighted_mean_cosine": 0.998756420601568,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.901636957934259,
+ "mean_cosine": 0.9976914672684307,
+ "p05_cosine": 0.9897294932832881,
+ "relative_l2_mean": 0.03979940049850905,
+ "relative_l2_p95": 0.14344509893530952,
+ "relative_l2_max": 0.44557719917122357,
+ "weighted_relative_l2_mean": 0.022108816306860343,
+ "weighted_relative_frobenius": 0.01276195516775435,
+ "weighted_mean_cosine": 0.998756420601568,
+ "max_relative_norm_change": 0.010085628995169935
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.45546,
+ "warm_p50_ms": 0.07047,
+ "warm_p95_ms": 0.0729593,
+ "warm_p99_ms": 0.08933676999999998,
+ "samples_ms": [
+ 0.07515,
+ 0.072844,
+ 0.072425,
+ 0.072356,
+ 0.0704,
+ 0.071867,
+ 0.071378,
+ 0.071378,
+ 0.071866,
+ 0.071867,
+ 0.072775,
+ 0.072844,
+ 0.072425,
+ 0.072355,
+ 0.071936,
+ 0.069073,
+ 0.0704,
+ 0.07047,
+ 0.068235,
+ 0.07047,
+ 0.07047,
+ 0.070051,
+ 0.07054,
+ 0.070051,
+ 0.069562,
+ 0.069911,
+ 0.07047,
+ 0.071029,
+ 0.07012,
+ 0.07054,
+ 0.069981,
+ 0.098407,
+ 0.071867,
+ 0.0704,
+ 0.07047,
+ 0.069561,
+ 0.070051,
+ 0.069073,
+ 0.069981,
+ 0.069632
+ ],
+ "warm_rows_per_second": 7265503.050943663
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 3436544,
+ "unique_rows": 13424,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "df8b79955d7468b7b2f0161394c5f6357e57ff0ab38d908a1ce9d750765b1361",
+ "table_sha256": "6246b2b9b270b1f329de11d92199f3fb5bb751123b26cb5a9f7b789465490a8b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.042963535990566015,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3437350,
+ "backend_with_key_registry_bytes": 3641298,
+ "backend_with_datrie_key_index_bytes": 3545942,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999651577072,
+ "mean_cosine": 0.9999999790112526,
+ "p05_cosine": 0.9999999732767585,
+ "relative_l2_mean": 0.00020655615851127175,
+ "relative_l2_p95": 0.00023390014156990472,
+ "relative_l2_max": 0.0002665638639023805,
+ "weighted_relative_l2_mean": 0.00020622079510946033,
+ "weighted_relative_frobenius": 0.00020596156618784485,
+ "weighted_mean_cosine": 0.9999999790907921,
+ "max_relative_norm_change": 0.0001125115961258942
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.486375,
+ "warm_p50_ms": 0.22,
+ "warm_p95_ms": 0.23145480000000002,
+ "warm_p99_ms": 0.24982195999999995,
+ "samples_ms": [
+ 0.220978,
+ 0.220489,
+ 0.260717,
+ 0.221467,
+ 0.220978,
+ 0.22475,
+ 0.219023,
+ 0.219931,
+ 0.219023,
+ 0.22,
+ 0.219512,
+ 0.220001,
+ 0.218674,
+ 0.220001,
+ 0.220489,
+ 0.218604,
+ 0.220419,
+ 0.21993,
+ 0.22426,
+ 0.231385,
+ 0.220908,
+ 0.219022,
+ 0.219092,
+ 0.219092,
+ 0.218603,
+ 0.22007,
+ 0.219511,
+ 0.219022,
+ 0.219022,
+ 0.219511,
+ 0.22,
+ 0.22,
+ 0.218114,
+ 0.220489,
+ 0.219511,
+ 0.219441,
+ 0.219511,
+ 0.232781,
+ 0.221886,
+ 0.221886
+ ],
+ "warm_rows_per_second": 2327272.727272727
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 3247131,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 420,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 420,
+ "zlib_blocks": 0,
+ "xor_blocks": 420,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "df8b79955d7468b7b2f0161394c5f6357e57ff0ab38d908a1ce9d750765b1361",
+ "table_sha256": "f868590cc6cca6b19dcc63977294a94461d481250dc1489640b26f6f449582fc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.33485457277856767,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3248340,
+ "backend_with_key_registry_bytes": 3452288,
+ "backend_with_datrie_key_index_bytes": 3356932,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999651577072,
+ "mean_cosine": 0.9999999790112526,
+ "p05_cosine": 0.9999999732767585,
+ "relative_l2_mean": 0.00020655615851127175,
+ "relative_l2_p95": 0.00023390014156990472,
+ "relative_l2_max": 0.0002665638639023805,
+ "weighted_relative_l2_mean": 0.00020622079510946033,
+ "weighted_relative_frobenius": 0.00020596156618784485,
+ "weighted_mean_cosine": 0.9999999790907921,
+ "max_relative_norm_change": 0.0001125115961258942
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.104495,
+ "warm_p50_ms": 0.229499,
+ "warm_p95_ms": 0.23576705,
+ "warm_p99_ms": 0.24346335,
+ "samples_ms": [
+ 0.231385,
+ 0.230826,
+ 0.229918,
+ 0.230336,
+ 0.229499,
+ 0.234667,
+ 0.230825,
+ 0.230896,
+ 0.22894,
+ 0.235575,
+ 0.229988,
+ 0.22894,
+ 0.228939,
+ 0.228032,
+ 0.239416,
+ 0.231804,
+ 0.230406,
+ 0.229987,
+ 0.22894,
+ 0.228451,
+ 0.22894,
+ 0.22901,
+ 0.235086,
+ 0.228521,
+ 0.228521,
+ 0.22901,
+ 0.235575,
+ 0.230896,
+ 0.229499,
+ 0.228591,
+ 0.228591,
+ 0.246051,
+ 0.231384,
+ 0.22894,
+ 0.228451,
+ 0.227473,
+ 0.228521,
+ 0.22901,
+ 0.22901,
+ 0.235086
+ ],
+ "warm_rows_per_second": 2230946.54007207
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 3247131,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 420,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 420,
+ "zlib_blocks": 0,
+ "xor_blocks": 420,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "df8b79955d7468b7b2f0161394c5f6357e57ff0ab38d908a1ce9d750765b1361",
+ "table_sha256": "f868590cc6cca6b19dcc63977294a94461d481250dc1489640b26f6f449582fc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.37537303287535906,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3248339,
+ "backend_with_key_registry_bytes": 3452287,
+ "backend_with_datrie_key_index_bytes": 3356931,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999651577072,
+ "mean_cosine": 0.9999999790112526,
+ "p05_cosine": 0.9999999732767585,
+ "relative_l2_mean": 0.00020655615851127175,
+ "relative_l2_p95": 0.00023390014156990472,
+ "relative_l2_max": 0.0002665638639023805,
+ "weighted_relative_l2_mean": 0.00020622079510946033,
+ "weighted_relative_frobenius": 0.00020596156618784485,
+ "weighted_mean_cosine": 0.9999999790907921,
+ "max_relative_norm_change": 0.0001125115961258942
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.122096,
+ "warm_p50_ms": 0.2292195,
+ "warm_p95_ms": 0.23530269999999998,
+ "warm_p99_ms": 0.25095052,
+ "samples_ms": [
+ 0.232362,
+ 0.231314,
+ 0.230406,
+ 0.230407,
+ 0.234597,
+ 0.230895,
+ 0.230406,
+ 0.229988,
+ 0.254083,
+ 0.234737,
+ 0.229988,
+ 0.229917,
+ 0.228451,
+ 0.229429,
+ 0.22894,
+ 0.228032,
+ 0.234597,
+ 0.228032,
+ 0.228102,
+ 0.22859,
+ 0.22901,
+ 0.232711,
+ 0.228032,
+ 0.227962,
+ 0.227543,
+ 0.22852,
+ 0.246051,
+ 0.230406,
+ 0.228521,
+ 0.228591,
+ 0.228521,
+ 0.228591,
+ 0.228521,
+ 0.228591,
+ 0.230406,
+ 0.229499,
+ 0.22901,
+ 0.228591,
+ 0.234736,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2233666.856441097
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 3247131,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 420,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 420,
+ "zlib_blocks": 0,
+ "xor_blocks": 420,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "df8b79955d7468b7b2f0161394c5f6357e57ff0ab38d908a1ce9d750765b1361",
+ "table_sha256": "f868590cc6cca6b19dcc63977294a94461d481250dc1489640b26f6f449582fc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3818764500319958,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3248339,
+ "backend_with_key_registry_bytes": 3452287,
+ "backend_with_datrie_key_index_bytes": 3356931,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999651577072,
+ "mean_cosine": 0.9999999790112526,
+ "p05_cosine": 0.9999999732767585,
+ "relative_l2_mean": 0.00020655615851127175,
+ "relative_l2_p95": 0.00023390014156990472,
+ "relative_l2_max": 0.0002665638639023805,
+ "weighted_relative_l2_mean": 0.00020622079510946033,
+ "weighted_relative_frobenius": 0.00020596156618784485,
+ "weighted_mean_cosine": 0.9999999790907921,
+ "max_relative_norm_change": 0.0001125115961258942
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.069994,
+ "warm_p50_ms": 0.229499,
+ "warm_p95_ms": 0.23416409999999996,
+ "warm_p99_ms": 0.25996763,
+ "samples_ms": [
+ 0.232292,
+ 0.230337,
+ 0.230477,
+ 0.229848,
+ 0.22894,
+ 0.229918,
+ 0.230895,
+ 0.22901,
+ 0.229429,
+ 0.229428,
+ 0.229429,
+ 0.228451,
+ 0.229429,
+ 0.233759,
+ 0.230477,
+ 0.241861,
+ 0.231384,
+ 0.229499,
+ 0.229429,
+ 0.229499,
+ 0.230337,
+ 0.22894,
+ 0.22894,
+ 0.227613,
+ 0.229918,
+ 0.229499,
+ 0.229498,
+ 0.228591,
+ 0.229429,
+ 0.228102,
+ 0.230406,
+ 0.229499,
+ 0.271544,
+ 0.232781,
+ 0.229988,
+ 0.227613,
+ 0.233759,
+ 0.228032,
+ 0.229429,
+ 0.22894
+ ],
+ "warm_rows_per_second": 2230946.54007207
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 2103141,
+ "unique_rows": 8304,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 260,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 260,
+ "zlib_blocks": 120,
+ "xor_blocks": 140,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8304
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "c647a6f03ff7d8da1493c385fa42746f8c9039f23e5ece75eaa237d5dbf9a646",
+ "table_sha256": "33122735c092e5b915a27d2fb2a526daf8c59506737d12440e551b43cc0ab2a9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2212739558890462,
+ "representatives_sha256": "f339349bb2cc6141baa1aed94f7e1bb07d53085b7f1721899cf616772aaac17e",
+ "structurally_merged_rows": 5120,
+ "backend_artifact_bytes": 2104348,
+ "backend_with_key_registry_bytes": 2308296,
+ "backend_with_datrie_key_index_bytes": 2212940,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.989995664652372,
+ "mean_cosine": 0.9985917131225744,
+ "p05_cosine": 0.9930612603811456,
+ "relative_l2_mean": 0.03102973626034901,
+ "relative_l2_p95": 0.11799753123837675,
+ "relative_l2_max": 0.14211255515914975,
+ "weighted_relative_l2_mean": 0.017689505161649466,
+ "weighted_relative_frobenius": 0.009591286480164959,
+ "weighted_mean_cosine": 0.9992415348598541,
+ "max_relative_norm_change": 0.01011448738647392
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900020103581858,
+ "mean_cosine": 0.9985917223456484,
+ "p05_cosine": 0.9930604868870606,
+ "relative_l2_mean": 0.0309019783063906,
+ "relative_l2_p95": 0.1179895543875637,
+ "relative_l2_max": 0.14213321561707243,
+ "weighted_relative_l2_mean": 0.017532013358679342,
+ "weighted_relative_frobenius": 0.009589247498575619,
+ "weighted_mean_cosine": 0.999241545341953,
+ "max_relative_norm_change": 0.010090500650870148
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.969376,
+ "warm_p50_ms": 0.228032,
+ "warm_p95_ms": 0.23362954999999996,
+ "warm_p99_ms": 0.25533849,
+ "samples_ms": [
+ 0.230895,
+ 0.22901,
+ 0.228032,
+ 0.227543,
+ 0.227055,
+ 0.229429,
+ 0.229009,
+ 0.228451,
+ 0.229428,
+ 0.264,
+ 0.2332,
+ 0.228102,
+ 0.229429,
+ 0.228032,
+ 0.228032,
+ 0.228521,
+ 0.226146,
+ 0.227683,
+ 0.227124,
+ 0.228032,
+ 0.227963,
+ 0.232711,
+ 0.228032,
+ 0.228101,
+ 0.227613,
+ 0.228102,
+ 0.228032,
+ 0.241791,
+ 0.228452,
+ 0.227613,
+ 0.228032,
+ 0.22901,
+ 0.226216,
+ 0.227124,
+ 0.232782,
+ 0.227194,
+ 0.227124,
+ 0.227124,
+ 0.226565,
+ 0.22852
+ ],
+ "warm_rows_per_second": 2245298.9054167834
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1981949,
+ "unique_rows": 7789,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 244,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 244,
+ "zlib_blocks": 143,
+ "xor_blocks": 101,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7789
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "ab0bb6a4d9becd73d5eba51dc6803ed34e58748286692c8621a31a98fff5db40",
+ "table_sha256": "ac87a3e6bfad8be0da8a26f5034b1f44aa92c96cfdfddf7c310b644585d871e7",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3681252470705658,
+ "representatives_sha256": "f25871770dac13fbc1e80ab01dc281b8096a2ddd742b8ef54af18ea84a9a057d",
+ "structurally_merged_rows": 5635,
+ "backend_artifact_bytes": 1983156,
+ "backend_with_key_registry_bytes": 2187104,
+ "backend_with_datrie_key_index_bytes": 2091748,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9801100636412576,
+ "mean_cosine": 0.99805187141531,
+ "p05_cosine": 0.9906073219999877,
+ "relative_l2_mean": 0.037439422532174416,
+ "relative_l2_p95": 0.1369488420509971,
+ "relative_l2_max": 0.20023657268276374,
+ "weighted_relative_l2_mean": 0.020800583754423056,
+ "weighted_relative_frobenius": 0.010050334109087993,
+ "weighted_mean_cosine": 0.9989795261793699,
+ "max_relative_norm_change": 0.010117918800127386
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9801145590004429,
+ "mean_cosine": 0.9980518783295492,
+ "p05_cosine": 0.9906054580284465,
+ "relative_l2_mean": 0.03731964535769372,
+ "relative_l2_p95": 0.13695811856336337,
+ "relative_l2_max": 0.20022589783209746,
+ "weighted_relative_l2_mean": 0.020646967196899225,
+ "weighted_relative_frobenius": 0.01004836675400962,
+ "weighted_mean_cosine": 0.9989795356859189,
+ "max_relative_norm_change": 0.010099054117758131
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.729051,
+ "warm_p50_ms": 0.2289745,
+ "warm_p95_ms": 0.23418489999999997,
+ "warm_p99_ms": 0.25485357,
+ "samples_ms": [
+ 0.232781,
+ 0.231803,
+ 0.2332,
+ 0.229987,
+ 0.229918,
+ 0.228521,
+ 0.229429,
+ 0.230407,
+ 0.22894,
+ 0.22894,
+ 0.228521,
+ 0.243607,
+ 0.230896,
+ 0.229499,
+ 0.229498,
+ 0.228102,
+ 0.229499,
+ 0.228521,
+ 0.229917,
+ 0.229918,
+ 0.228521,
+ 0.22894,
+ 0.227544,
+ 0.228521,
+ 0.229009,
+ 0.228032,
+ 0.228102,
+ 0.229499,
+ 0.262044,
+ 0.232851,
+ 0.229079,
+ 0.228102,
+ 0.233689,
+ 0.22894,
+ 0.226984,
+ 0.228521,
+ 0.22894,
+ 0.227054,
+ 0.228451,
+ 0.228451
+ ],
+ "warm_rows_per_second": 2236056.8534924197
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1965068,
+ "unique_rows": 7719,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 242,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 242,
+ "zlib_blocks": 150,
+ "xor_blocks": 92,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7719
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "ff83a0d8fb9627fde316cd06428c16c905df3ca890ce92dbcfc81cce194ac942",
+ "table_sha256": "ce5e9b2d5599c74b4bc1d18c3aec2bffadb026b0e6eecf9e2f7d9335e1653af8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21476851403713226,
+ "representatives_sha256": "25a359b0eccdd8b4ffebd880b7df57c500c29652d51ebad842d5f3939af99ba7",
+ "structurally_merged_rows": 5705,
+ "backend_artifact_bytes": 1966274,
+ "backend_with_key_registry_bytes": 2170222,
+ "backend_with_datrie_key_index_bytes": 2074866,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9703426200710272,
+ "mean_cosine": 0.9979244323553459,
+ "p05_cosine": 0.9900997123354504,
+ "relative_l2_mean": 0.03859114289792424,
+ "relative_l2_p95": 0.14087544569897528,
+ "relative_l2_max": 0.24497565577229352,
+ "weighted_relative_l2_mean": 0.021356463889245822,
+ "weighted_relative_frobenius": 0.010114557173851743,
+ "weighted_mean_cosine": 0.998918017462548,
+ "max_relative_norm_change": 0.010117918800127386
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9703459889593165,
+ "mean_cosine": 0.9979244394645114,
+ "p05_cosine": 0.9900974012823818,
+ "relative_l2_mean": 0.038472432653053085,
+ "relative_l2_p95": 0.14086727251138315,
+ "relative_l2_max": 0.2449596840063457,
+ "weighted_relative_l2_mean": 0.021203362287428976,
+ "weighted_relative_frobenius": 0.01011260477785047,
+ "weighted_mean_cosine": 0.998918027063179,
+ "max_relative_norm_change": 0.010099054117758131
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.417699,
+ "warm_p50_ms": 0.22893950000000002,
+ "warm_p95_ms": 0.2384914,
+ "warm_p99_ms": 0.25379566000000003,
+ "samples_ms": [
+ 0.23802,
+ 0.230896,
+ 0.229428,
+ 0.22901,
+ 0.22894,
+ 0.22901,
+ 0.229429,
+ 0.228939,
+ 0.233689,
+ 0.257854,
+ 0.230406,
+ 0.229988,
+ 0.226705,
+ 0.22901,
+ 0.229009,
+ 0.228102,
+ 0.226705,
+ 0.227963,
+ 0.228451,
+ 0.22894,
+ 0.228032,
+ 0.232711,
+ 0.226705,
+ 0.227613,
+ 0.226635,
+ 0.228451,
+ 0.247448,
+ 0.229499,
+ 0.22901,
+ 0.228032,
+ 0.231874,
+ 0.226984,
+ 0.227124,
+ 0.227613,
+ 0.232292,
+ 0.227962,
+ 0.228032,
+ 0.226635,
+ 0.228102,
+ 0.228591
+ ],
+ "warm_rows_per_second": 2236398.6992196627
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1960815,
+ "unique_rows": 7699,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 241,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 241,
+ "zlib_blocks": 159,
+ "xor_blocks": 82,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7699
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "045e517281546731b546c8b1d7adfc51d1c9ea926c5bca5a72a5ce04409a0f8c",
+ "table_sha256": "36f2ccb7e6feecafea2be2c1c5aa7f5343e4b2ccd0c5640a3f0b9dfe2b1e614b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21026584203355014,
+ "representatives_sha256": "1d15fcb1d67eedef0defa6c3a30e986b0b9b200376c53338020b637f2e31bc7d",
+ "structurally_merged_rows": 5725,
+ "backend_artifact_bytes": 1962021,
+ "backend_with_key_registry_bytes": 2165969,
+ "backend_with_datrie_key_index_bytes": 2070613,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9515207734429368,
+ "mean_cosine": 0.997828042121488,
+ "p05_cosine": 0.9899147410588622,
+ "relative_l2_mean": 0.03923419065524245,
+ "relative_l2_p95": 0.14203493723420965,
+ "relative_l2_max": 0.31037335562002444,
+ "weighted_relative_l2_mean": 0.02170580614399317,
+ "weighted_relative_frobenius": 0.011063628330345916,
+ "weighted_mean_cosine": 0.9988661954256461,
+ "max_relative_norm_change": 0.010117918800127386
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9515114032704579,
+ "mean_cosine": 0.9978280484143693,
+ "p05_cosine": 0.9899122601784989,
+ "relative_l2_mean": 0.03911578474741316,
+ "relative_l2_p95": 0.14202421468116166,
+ "relative_l2_max": 0.31040553239053054,
+ "weighted_relative_l2_mean": 0.0215528822799628,
+ "weighted_relative_frobenius": 0.011061795795406735,
+ "weighted_mean_cosine": 0.9988662048193726,
+ "max_relative_norm_change": 0.010099054117758131
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.614581,
+ "warm_p50_ms": 0.22806700000000002,
+ "warm_p95_ms": 0.23920269999999993,
+ "warm_p99_ms": 0.27026492,
+ "samples_ms": [
+ 0.231804,
+ 0.232781,
+ 0.228032,
+ 0.227962,
+ 0.226566,
+ 0.22901,
+ 0.274895,
+ 0.263023,
+ 0.229917,
+ 0.22894,
+ 0.227543,
+ 0.22887,
+ 0.227054,
+ 0.228521,
+ 0.228032,
+ 0.228521,
+ 0.227124,
+ 0.228102,
+ 0.233759,
+ 0.226565,
+ 0.228032,
+ 0.227543,
+ 0.227124,
+ 0.237949,
+ 0.22901,
+ 0.228521,
+ 0.228102,
+ 0.228102,
+ 0.226705,
+ 0.228032,
+ 0.227124,
+ 0.232293,
+ 0.22901,
+ 0.227613,
+ 0.227682,
+ 0.228032,
+ 0.2332,
+ 0.227124,
+ 0.227194,
+ 0.227124
+ ],
+ "warm_rows_per_second": 2244954.3335949522
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1957354,
+ "unique_rows": 7685,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 241,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 241,
+ "zlib_blocks": 156,
+ "xor_blocks": 85,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7685
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "ee86f39371f052d826061c20167a6abfefa15b99ce9576dc2503a4130acdb9cc",
+ "table_sha256": "e6e1d7d625eca2be032067ada24771cf4090ba89c350102b498e11347412b7da",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.20897321705706418,
+ "representatives_sha256": "759730df622c90b1ed104e6406f6e8efdcbdb5b3405c9dd821ba0080e465ed99",
+ "structurally_merged_rows": 5739,
+ "backend_artifact_bytes": 1958560,
+ "backend_with_key_registry_bytes": 2162508,
+ "backend_with_datrie_key_index_bytes": 2067152,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9016499744140217,
+ "mean_cosine": 0.997691905635778,
+ "p05_cosine": 0.9897283669000854,
+ "relative_l2_mean": 0.039900345172597676,
+ "relative_l2_p95": 0.14344021411172275,
+ "relative_l2_max": 0.4455450884412612,
+ "weighted_relative_l2_mean": 0.022253324510674376,
+ "weighted_relative_frobenius": 0.012762492197345335,
+ "weighted_mean_cosine": 0.9987566249052615,
+ "max_relative_norm_change": 0.010117918800127386
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9016575948166327,
+ "mean_cosine": 0.9976919122693515,
+ "p05_cosine": 0.9897270242059109,
+ "relative_l2_mean": 0.03978215263776933,
+ "relative_l2_p95": 0.14342778425751362,
+ "relative_l2_max": 0.44554379064698335,
+ "weighted_relative_l2_mean": 0.022100618010424723,
+ "weighted_relative_frobenius": 0.012760904632781796,
+ "weighted_mean_cosine": 0.9987566336903675,
+ "max_relative_norm_change": 0.010099054117758131
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.496759,
+ "warm_p50_ms": 0.22852099999999997,
+ "warm_p95_ms": 0.23967089999999996,
+ "warm_p99_ms": 0.25179923,
+ "samples_ms": [
+ 0.234597,
+ 0.234248,
+ 0.256457,
+ 0.235574,
+ 0.230407,
+ 0.229429,
+ 0.228451,
+ 0.226565,
+ 0.22901,
+ 0.228102,
+ 0.226705,
+ 0.227613,
+ 0.227543,
+ 0.227543,
+ 0.227054,
+ 0.227543,
+ 0.228102,
+ 0.229499,
+ 0.230825,
+ 0.244514,
+ 0.229498,
+ 0.227962,
+ 0.231873,
+ 0.231803,
+ 0.231873,
+ 0.22887,
+ 0.228451,
+ 0.233759,
+ 0.227194,
+ 0.226636,
+ 0.227054,
+ 0.231873,
+ 0.228591,
+ 0.227543,
+ 0.227613,
+ 0.227194,
+ 0.239416,
+ 0.22894,
+ 0.228101,
+ 0.226705
+ ],
+ "warm_rows_per_second": 2240494.309056936
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 132,
+ "storage": {
+ "artifact_bytes": 1771968,
+ "unique_rows": 13424,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "9f777ab7bd00dedb4404ea8845e359fbaf82aaa095e98cfef4783fbcfeb33df4",
+ "table_sha256": "7c01f03bbb40873f368ce12e55598733cbd8271f035cb510761f8f5794d8d841",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.035551971988752484,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1772818,
+ "backend_with_key_registry_bytes": 1976766,
+ "backend_with_datrie_key_index_bytes": 1881410,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999442322587219,
+ "mean_cosine": 0.9999795979612697,
+ "p05_cosine": 0.9999690851376799,
+ "relative_l2_mean": 0.006358420584391994,
+ "relative_l2_p95": 0.00788237035989277,
+ "relative_l2_max": 0.01059356282798076,
+ "weighted_relative_l2_mean": 0.006371080776756557,
+ "weighted_relative_frobenius": 0.006401343301882059,
+ "weighted_mean_cosine": 0.9999795156917556,
+ "max_relative_norm_change": 0.003028731873775844
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.380845,
+ "warm_p50_ms": 0.122991,
+ "warm_p95_ms": 0.12555354999999999,
+ "warm_p99_ms": 0.1346748,
+ "samples_ms": [
+ 0.125365,
+ 0.123899,
+ 0.12348,
+ 0.123061,
+ 0.129136,
+ 0.122502,
+ 0.12348,
+ 0.12355,
+ 0.122572,
+ 0.123479,
+ 0.12299,
+ 0.122991,
+ 0.122991,
+ 0.122991,
+ 0.123968,
+ 0.12348,
+ 0.122571,
+ 0.123479,
+ 0.122991,
+ 0.122013,
+ 0.122502,
+ 0.122501,
+ 0.123061,
+ 0.122572,
+ 0.122082,
+ 0.122991,
+ 0.121594,
+ 0.12348,
+ 0.122082,
+ 0.138216,
+ 0.123968,
+ 0.121524,
+ 0.122991,
+ 0.121175,
+ 0.121663,
+ 0.12306,
+ 0.12348,
+ 0.12348,
+ 0.122432,
+ 0.12299
+ ],
+ "warm_rows_per_second": 4162906.2289110585
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1655350,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 217,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 217,
+ "zlib_blocks": 0,
+ "xor_blocks": 217,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "9f777ab7bd00dedb4404ea8845e359fbaf82aaa095e98cfef4783fbcfeb33df4",
+ "table_sha256": "51281896ec1aa4b107e865f42c1f991de94ffca03446b5e95bcde52b96a7ec93",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25538244703784585,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1656603,
+ "backend_with_key_registry_bytes": 1860551,
+ "backend_with_datrie_key_index_bytes": 1765195,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999442322587219,
+ "mean_cosine": 0.9999795979612697,
+ "p05_cosine": 0.9999690851376799,
+ "relative_l2_mean": 0.006358420584391994,
+ "relative_l2_p95": 0.00788237035989277,
+ "relative_l2_max": 0.01059356282798076,
+ "weighted_relative_l2_mean": 0.006371080776756557,
+ "weighted_relative_frobenius": 0.006401343301882059,
+ "weighted_mean_cosine": 0.9999795156917556,
+ "max_relative_norm_change": 0.003028731873775844
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.024469,
+ "warm_p50_ms": 0.132489,
+ "warm_p95_ms": 0.13857554999999994,
+ "warm_p99_ms": 0.16364547999999998,
+ "samples_ms": [
+ 0.154698,
+ 0.134374,
+ 0.132978,
+ 0.133886,
+ 0.132489,
+ 0.134375,
+ 0.132978,
+ 0.132908,
+ 0.132,
+ 0.132908,
+ 0.132978,
+ 0.132978,
+ 0.131512,
+ 0.133467,
+ 0.132489,
+ 0.131511,
+ 0.131511,
+ 0.132489,
+ 0.132,
+ 0.132,
+ 0.131512,
+ 0.137169,
+ 0.132,
+ 0.131511,
+ 0.132489,
+ 0.132489,
+ 0.132,
+ 0.131511,
+ 0.137727,
+ 0.169366,
+ 0.136261,
+ 0.132001,
+ 0.132001,
+ 0.133397,
+ 0.132489,
+ 0.130603,
+ 0.131581,
+ 0.131582,
+ 0.131511,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3864471.7674674885
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1655350,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 217,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 217,
+ "zlib_blocks": 0,
+ "xor_blocks": 217,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "9f777ab7bd00dedb4404ea8845e359fbaf82aaa095e98cfef4783fbcfeb33df4",
+ "table_sha256": "51281896ec1aa4b107e865f42c1f991de94ffca03446b5e95bcde52b96a7ec93",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.23377129808068275,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1656602,
+ "backend_with_key_registry_bytes": 1860550,
+ "backend_with_datrie_key_index_bytes": 1765194,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999442322587219,
+ "mean_cosine": 0.9999795979612697,
+ "p05_cosine": 0.9999690851376799,
+ "relative_l2_mean": 0.006358420584391994,
+ "relative_l2_p95": 0.00788237035989277,
+ "relative_l2_max": 0.01059356282798076,
+ "weighted_relative_l2_mean": 0.006371080776756557,
+ "weighted_relative_frobenius": 0.006401343301882059,
+ "weighted_mean_cosine": 0.9999795156917556,
+ "max_relative_norm_change": 0.003028731873775844
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.984659,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.13584419999999997,
+ "warm_p99_ms": 0.14261235,
+ "samples_ms": [
+ 0.135701,
+ 0.134794,
+ 0.132,
+ 0.130534,
+ 0.131093,
+ 0.132978,
+ 0.132,
+ 0.131092,
+ 0.132,
+ 0.132489,
+ 0.131581,
+ 0.131092,
+ 0.131092,
+ 0.131511,
+ 0.131022,
+ 0.130604,
+ 0.1452,
+ 0.138565,
+ 0.131581,
+ 0.134375,
+ 0.134305,
+ 0.132978,
+ 0.131092,
+ 0.130115,
+ 0.130604,
+ 0.131511,
+ 0.131511,
+ 0.130114,
+ 0.131511,
+ 0.131022,
+ 0.131511,
+ 0.130603,
+ 0.131511,
+ 0.133886,
+ 0.131581,
+ 0.129695,
+ 0.130603,
+ 0.130603,
+ 0.131581,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1655350,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 217,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 217,
+ "zlib_blocks": 0,
+ "xor_blocks": 217,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "9f777ab7bd00dedb4404ea8845e359fbaf82aaa095e98cfef4783fbcfeb33df4",
+ "table_sha256": "51281896ec1aa4b107e865f42c1f991de94ffca03446b5e95bcde52b96a7ec93",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1943037409801036,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1656602,
+ "backend_with_key_registry_bytes": 1860550,
+ "backend_with_datrie_key_index_bytes": 1765194,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999442322587219,
+ "mean_cosine": 0.9999795979612697,
+ "p05_cosine": 0.9999690851376799,
+ "relative_l2_mean": 0.006358420584391994,
+ "relative_l2_p95": 0.00788237035989277,
+ "relative_l2_max": 0.01059356282798076,
+ "weighted_relative_l2_mean": 0.006371080776756557,
+ "weighted_relative_frobenius": 0.006401343301882059,
+ "weighted_mean_cosine": 0.9999795156917556,
+ "max_relative_norm_change": 0.003028731873775844
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.934094,
+ "warm_p50_ms": 0.13245400000000002,
+ "warm_p95_ms": 0.14004234999999998,
+ "warm_p99_ms": 0.15817857999999999,
+ "samples_ms": [
+ 0.13626,
+ 0.135352,
+ 0.132489,
+ 0.135841,
+ 0.134793,
+ 0.134864,
+ 0.132489,
+ 0.132,
+ 0.135282,
+ 0.139543,
+ 0.163708,
+ 0.132978,
+ 0.135352,
+ 0.131022,
+ 0.130603,
+ 0.131022,
+ 0.131511,
+ 0.130114,
+ 0.131022,
+ 0.131511,
+ 0.132978,
+ 0.132908,
+ 0.132,
+ 0.134444,
+ 0.132908,
+ 0.133955,
+ 0.132419,
+ 0.131022,
+ 0.130603,
+ 0.132,
+ 0.130115,
+ 0.135771,
+ 0.130533,
+ 0.131023,
+ 0.131023,
+ 0.133467,
+ 0.130603,
+ 0.132,
+ 0.130114,
+ 0.14953
+ ],
+ "warm_rows_per_second": 3865492.925845954
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1129211,
+ "unique_rows": 8311,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 135,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 134,
+ "zlib_blocks": 1,
+ "xor_blocks": 133,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8311
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "1bdf0ad9b3f1f224ceaa52430fb36f9c423cd334ad1cc2673d744369700cc577",
+ "table_sha256": "35ce753c687e470c242945aff294cec188500c6a6784cd7668d8d257e4b73658",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13416230399161577,
+ "representatives_sha256": "f8a053eb28da44da2bedb7269b459de91eab848d6ba2f892a9993216eb1a3d16",
+ "structurally_merged_rows": 5113,
+ "backend_artifact_bytes": 1130460,
+ "backend_with_key_registry_bytes": 1334408,
+ "backend_with_datrie_key_index_bytes": 1239052,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9899136811807194,
+ "mean_cosine": 0.9985749964000988,
+ "p05_cosine": 0.9930476127303063,
+ "relative_l2_mean": 0.0348989236734183,
+ "relative_l2_p95": 0.11797927934547961,
+ "relative_l2_max": 0.14306266325996833,
+ "weighted_relative_l2_mean": 0.02245632026244996,
+ "weighted_relative_frobenius": 0.011525639114851247,
+ "weighted_mean_cosine": 0.999222689535393,
+ "max_relative_norm_change": 0.010991590798041326
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900110489149397,
+ "mean_cosine": 0.9985798293373181,
+ "p05_cosine": 0.9930382461113353,
+ "relative_l2_mean": 0.031064004534487093,
+ "relative_l2_p95": 0.11811641388146403,
+ "relative_l2_max": 0.14220749334411578,
+ "weighted_relative_l2_mean": 0.017657962537942273,
+ "weighted_relative_frobenius": 0.009713811108145775,
+ "weighted_mean_cosine": 0.9992335472616621,
+ "max_relative_norm_change": 0.01010024019379909
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.240173,
+ "warm_p50_ms": 0.1315115,
+ "warm_p95_ms": 0.13361014999999998,
+ "warm_p99_ms": 0.15648195999999998,
+ "samples_ms": [
+ 0.133397,
+ 0.133467,
+ 0.133397,
+ 0.133467,
+ 0.131511,
+ 0.132978,
+ 0.130533,
+ 0.131092,
+ 0.132908,
+ 0.131023,
+ 0.131023,
+ 0.130534,
+ 0.131581,
+ 0.130604,
+ 0.131023,
+ 0.133467,
+ 0.131511,
+ 0.129626,
+ 0.130534,
+ 0.129626,
+ 0.131512,
+ 0.169366,
+ 0.132489,
+ 0.132,
+ 0.131022,
+ 0.132,
+ 0.132908,
+ 0.13633,
+ 0.131581,
+ 0.131512,
+ 0.130534,
+ 0.132838,
+ 0.131022,
+ 0.131511,
+ 0.130604,
+ 0.13242,
+ 0.130115,
+ 0.131512,
+ 0.131511,
+ 0.131511
+ ],
+ "warm_rows_per_second": 3893195.6520912615
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1070809,
+ "unique_rows": 7790,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 126,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 126,
+ "zlib_blocks": 3,
+ "xor_blocks": 123,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7790
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "0ed19c226bc08bca8268d27f1e94ef3177d42a35930d4a26e31d384f2eed8f04",
+ "table_sha256": "8c803988d779ed98ddc748a63b024fb8b9331a5d0ed698777705c9f8e4bec12c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13391485600732267,
+ "representatives_sha256": "ced14b422e7247d7fb744385596f5f3dfbe060303f4f5b132525ea5d30c9cefb",
+ "structurally_merged_rows": 5634,
+ "backend_artifact_bytes": 1072058,
+ "backend_with_key_registry_bytes": 1276006,
+ "backend_with_datrie_key_index_bytes": 1180650,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9800819990811546,
+ "mean_cosine": 0.9980303391823193,
+ "p05_cosine": 0.9905511398594655,
+ "relative_l2_mean": 0.04114180380308397,
+ "relative_l2_p95": 0.1376814432707729,
+ "relative_l2_max": 0.2003393432717317,
+ "weighted_relative_l2_mean": 0.02548624139235245,
+ "weighted_relative_frobenius": 0.011914136039216528,
+ "weighted_mean_cosine": 0.9989583493846503,
+ "max_relative_norm_change": 0.010991590798041326
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9800028673484945,
+ "mean_cosine": 0.99803364650307,
+ "p05_cosine": 0.9905125901414213,
+ "relative_l2_mean": 0.03755827903409694,
+ "relative_l2_p95": 0.13786314874743147,
+ "relative_l2_max": 0.2006707042295582,
+ "weighted_relative_l2_mean": 0.020809976491343307,
+ "weighted_relative_frobenius": 0.01017311955010261,
+ "weighted_mean_cosine": 0.9989684651491306,
+ "max_relative_norm_change": 0.01010024019379909
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.921278,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.13810749999999997,
+ "warm_p99_ms": 0.16200422999999997,
+ "samples_ms": [
+ 0.135772,
+ 0.135213,
+ 0.132001,
+ 0.131022,
+ 0.130603,
+ 0.132,
+ 0.132001,
+ 0.17181,
+ 0.137657,
+ 0.134794,
+ 0.133886,
+ 0.131512,
+ 0.132,
+ 0.132978,
+ 0.130603,
+ 0.131092,
+ 0.132489,
+ 0.132,
+ 0.131511,
+ 0.130044,
+ 0.131022,
+ 0.131512,
+ 0.132001,
+ 0.130534,
+ 0.130115,
+ 0.130604,
+ 0.132978,
+ 0.130603,
+ 0.131092,
+ 0.131511,
+ 0.130533,
+ 0.130604,
+ 0.132,
+ 0.132419,
+ 0.131511,
+ 0.134794,
+ 0.146667,
+ 0.132489,
+ 0.131022,
+ 0.135353
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1062349,
+ "unique_rows": 7715,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 125,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 125,
+ "zlib_blocks": 3,
+ "xor_blocks": 122,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7715
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "a65ed6f04c6eb38ba974601bea3914635bc1977aa885be2c34a7ef1252e5a0a0",
+ "table_sha256": "db471a073504d12ba7edf777c074f618650e74d07c93eaafc75195ee9fc107a5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13173098699189723,
+ "representatives_sha256": "2c651c7431fd40185e50e6a43e931eddacfe49168f0011ce791f44a0fa4ae7fb",
+ "structurally_merged_rows": 5709,
+ "backend_artifact_bytes": 1063598,
+ "backend_with_key_registry_bytes": 1267546,
+ "backend_with_datrie_key_index_bytes": 1172190,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9701953282035927,
+ "mean_cosine": 0.9978960395945996,
+ "p05_cosine": 0.9900424388819883,
+ "relative_l2_mean": 0.04233024132743534,
+ "relative_l2_p95": 0.1413128536710294,
+ "relative_l2_max": 0.2456655597397933,
+ "weighted_relative_l2_mean": 0.026059843137108237,
+ "weighted_relative_frobenius": 0.011959636682586174,
+ "weighted_mean_cosine": 0.9988935294204053,
+ "max_relative_norm_change": 0.010991590798041326
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9701616210721082,
+ "mean_cosine": 0.9978989440226749,
+ "p05_cosine": 0.9900253707939715,
+ "relative_l2_mean": 0.03878435322246367,
+ "relative_l2_p95": 0.1413036081012778,
+ "relative_l2_max": 0.245592669302795,
+ "weighted_relative_l2_mean": 0.021401743647156593,
+ "weighted_relative_frobenius": 0.010226536183463626,
+ "weighted_mean_cosine": 0.9989034507279311,
+ "max_relative_norm_change": 0.01010024019379909
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.099094,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.1344483,
+ "warm_p99_ms": 0.14103881,
+ "samples_ms": [
+ 0.133886,
+ 0.135841,
+ 0.131511,
+ 0.131581,
+ 0.134375,
+ 0.132,
+ 0.131512,
+ 0.132489,
+ 0.132489,
+ 0.132,
+ 0.131512,
+ 0.13249,
+ 0.130604,
+ 0.129207,
+ 0.130603,
+ 0.131511,
+ 0.130045,
+ 0.131512,
+ 0.131582,
+ 0.144362,
+ 0.132,
+ 0.130603,
+ 0.131022,
+ 0.131511,
+ 0.131022,
+ 0.131092,
+ 0.131581,
+ 0.131093,
+ 0.131023,
+ 0.131511,
+ 0.130114,
+ 0.131581,
+ 0.131092,
+ 0.130603,
+ 0.131511,
+ 0.131022,
+ 0.130603,
+ 0.131162,
+ 0.129137,
+ 0.129625
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1060046,
+ "unique_rows": 7693,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 125,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 124,
+ "zlib_blocks": 3,
+ "xor_blocks": 121,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7693
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "b66c6be47db4d19b81db87d089815be279d67672e8f73e822b283861f5b7a554",
+ "table_sha256": "d0a8b9d19b68d5939d6742797b4254dbdbbabb180e9ba18a1fe7337f33d137d5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1281924748327583,
+ "representatives_sha256": "34059e133bf2367ed0352f5df1f47ea57baab36d0f1b6bec9f94453ff2b02761",
+ "structurally_merged_rows": 5731,
+ "backend_artifact_bytes": 1061295,
+ "backend_with_key_registry_bytes": 1265243,
+ "backend_with_datrie_key_index_bytes": 1169887,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9513409007200618,
+ "mean_cosine": 0.997793466896943,
+ "p05_cosine": 0.9898396814040983,
+ "relative_l2_mean": 0.04300487377165273,
+ "relative_l2_p95": 0.1425172665914292,
+ "relative_l2_max": 0.3109355148579674,
+ "weighted_relative_l2_mean": 0.026423577148642827,
+ "weighted_relative_frobenius": 0.01277455427633077,
+ "weighted_mean_cosine": 0.9988387228072758,
+ "max_relative_norm_change": 0.010991590798041326
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9513650714582464,
+ "mean_cosine": 0.9977962797082564,
+ "p05_cosine": 0.9898302935962124,
+ "relative_l2_mean": 0.039469799009364724,
+ "relative_l2_p95": 0.14263249893132462,
+ "relative_l2_max": 0.3112027514607818,
+ "weighted_relative_l2_mean": 0.021771616550171242,
+ "weighted_relative_frobenius": 0.01116869855378381,
+ "weighted_mean_cosine": 0.9988485868475183,
+ "max_relative_norm_change": 0.01010024019379909
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.182904,
+ "warm_p50_ms": 0.132489,
+ "warm_p95_ms": 0.13544725,
+ "warm_p99_ms": 0.14328798,
+ "samples_ms": [
+ 0.134375,
+ 0.132909,
+ 0.132489,
+ 0.132908,
+ 0.134794,
+ 0.134794,
+ 0.137238,
+ 0.133886,
+ 0.134794,
+ 0.134794,
+ 0.134934,
+ 0.133397,
+ 0.130603,
+ 0.132489,
+ 0.132977,
+ 0.131511,
+ 0.130114,
+ 0.132,
+ 0.131022,
+ 0.132908,
+ 0.131512,
+ 0.131023,
+ 0.130045,
+ 0.131023,
+ 0.131092,
+ 0.147156,
+ 0.133467,
+ 0.130114,
+ 0.130114,
+ 0.130603,
+ 0.130603,
+ 0.131511,
+ 0.132419,
+ 0.135353,
+ 0.134864,
+ 0.132977,
+ 0.133397,
+ 0.131022,
+ 0.129625,
+ 0.129625
+ ],
+ "warm_rows_per_second": 3864471.7674674885
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1058542,
+ "unique_rows": 7681,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 124,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 124,
+ "zlib_blocks": 3,
+ "xor_blocks": 121,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7681
+ },
+ "source_float32_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07",
+ "decoded_float32_sha256": "c798df084039f9f563c0202f4a79c4ca176c0e7a9ba4355c1cc5b182966b3305",
+ "table_sha256": "4c0650eba6aba8af28ecbf7958f4b514c47d67af80085f593d8789966fcc184e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13557358807884157,
+ "representatives_sha256": "75da31e2b4ff5f94336fd5386b6afd1572da78ae180b3ae9e3a8ddd16ea4ee2d",
+ "structurally_merged_rows": 5743,
+ "backend_artifact_bytes": 1059791,
+ "backend_with_key_registry_bytes": 1263739,
+ "backend_with_datrie_key_index_bytes": 1168383,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9013489996890685,
+ "mean_cosine": 0.997651994391575,
+ "p05_cosine": 0.9896481483675127,
+ "relative_l2_mean": 0.04368167817732234,
+ "relative_l2_p95": 0.14425391081360175,
+ "relative_l2_max": 0.4464244868253807,
+ "weighted_relative_l2_mean": 0.02698530975733708,
+ "weighted_relative_frobenius": 0.01432390571017968,
+ "weighted_mean_cosine": 0.9987239127258772,
+ "max_relative_norm_change": 0.010991590798041326
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9014350454777781,
+ "mean_cosine": 0.9976546353753513,
+ "p05_cosine": 0.9896478765184749,
+ "relative_l2_mean": 0.04015194510655089,
+ "relative_l2_p95": 0.14421017659095295,
+ "relative_l2_max": 0.4461840001692348,
+ "weighted_relative_l2_mean": 0.022339067423681105,
+ "weighted_relative_frobenius": 0.012914232859430198,
+ "weighted_mean_cosine": 0.9987336847848998,
+ "max_relative_norm_change": 0.01010024019379909
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.951519,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.13439945,
+ "warm_p99_ms": 0.15241613999999998,
+ "samples_ms": [
+ 0.133816,
+ 0.134375,
+ 0.132978,
+ 0.132,
+ 0.132,
+ 0.132489,
+ 0.132,
+ 0.132489,
+ 0.134375,
+ 0.133816,
+ 0.132,
+ 0.131512,
+ 0.133886,
+ 0.132978,
+ 0.130534,
+ 0.132,
+ 0.130114,
+ 0.130533,
+ 0.130044,
+ 0.130114,
+ 0.131022,
+ 0.131511,
+ 0.130114,
+ 0.131092,
+ 0.163638,
+ 0.134864,
+ 0.132908,
+ 0.132001,
+ 0.130534,
+ 0.132489,
+ 0.134375,
+ 0.133397,
+ 0.131511,
+ 0.131511,
+ 0.131022,
+ 0.132,
+ 0.130603,
+ 0.131023,
+ 0.131511,
+ 0.132
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ }
+ ],
+ "baseline_quality": {},
+ "training_row_sha256": "88ef8b5ba65079c79087b3a393cdf85dea46b1782995d5a6bda1289c567d0a07"
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 7,
+ "rows": 13424,
+ "dim": 128,
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "source_float32_bytes": 6873088,
+ "key_registry_bytes": 203948,
+ "unchanged_backbone_artifact_bytes": 0,
+ "alternative_datrie_key_index_bytes": 108592,
+ "query_rows": 512,
+ "query_ids_sha256": "80ab83de1ecc982405671e980d9e9a66070e51aead43121126faabb6b44a50cd",
+ "frequency_trace_sha256": "1a0389197253d049e85973cd9812529ea0dc3eae69dcdb62085a34b49c0d90b8",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 512,
+ "storage": {
+ "artifact_bytes": 6873088,
+ "unique_rows": 13424,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "table_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.03638713504187763,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6873893,
+ "backend_with_key_registry_bytes": 7077841,
+ "backend_with_datrie_key_index_bytes": 6982485,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.375607,
+ "warm_p50_ms": 0.060971,
+ "warm_p95_ms": 0.06530545,
+ "warm_p99_ms": 0.08922931999999999,
+ "samples_ms": [
+ 0.065302,
+ 0.104483,
+ 0.065371,
+ 0.062927,
+ 0.062928,
+ 0.06146,
+ 0.061949,
+ 0.063346,
+ 0.061042,
+ 0.062927,
+ 0.063416,
+ 0.062508,
+ 0.062019,
+ 0.061949,
+ 0.062927,
+ 0.060552,
+ 0.060482,
+ 0.060971,
+ 0.060063,
+ 0.059644,
+ 0.060971,
+ 0.060622,
+ 0.060064,
+ 0.060553,
+ 0.061112,
+ 0.060483,
+ 0.060134,
+ 0.061041,
+ 0.060553,
+ 0.060623,
+ 0.060133,
+ 0.060552,
+ 0.059085,
+ 0.058667,
+ 0.058667,
+ 0.060552,
+ 0.06153,
+ 0.061041,
+ 0.060133,
+ 0.059644
+ ],
+ "warm_rows_per_second": 8397434.846074363
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 6482626,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 839,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 839,
+ "zlib_blocks": 46,
+ "xor_blocks": 793,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "table_sha256": "9ed6c5c269513323a75b48c96f80fd98f7c7cbc77060148bafda48caa4abed5b",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5779507379047573,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6483835,
+ "backend_with_key_registry_bytes": 6687783,
+ "backend_with_datrie_key_index_bytes": 6592427,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.391392,
+ "warm_p50_ms": 0.0719715,
+ "warm_p95_ms": 0.07622145,
+ "warm_p99_ms": 0.07779376,
+ "samples_ms": [
+ 0.078502,
+ 0.075708,
+ 0.0748,
+ 0.076197,
+ 0.076686,
+ 0.075708,
+ 0.0748,
+ 0.07473,
+ 0.073822,
+ 0.07515,
+ 0.074731,
+ 0.074312,
+ 0.074241,
+ 0.072426,
+ 0.071029,
+ 0.07054,
+ 0.070889,
+ 0.072844,
+ 0.072844,
+ 0.072775,
+ 0.070051,
+ 0.07047,
+ 0.071867,
+ 0.071378,
+ 0.071377,
+ 0.071378,
+ 0.070959,
+ 0.071936,
+ 0.070959,
+ 0.071937,
+ 0.071447,
+ 0.071518,
+ 0.072426,
+ 0.072006,
+ 0.071029,
+ 0.071448,
+ 0.071028,
+ 0.07047,
+ 0.070889,
+ 0.073822
+ ],
+ "warm_rows_per_second": 7113927.040564668
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 6482626,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 839,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 839,
+ "zlib_blocks": 46,
+ "xor_blocks": 793,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "table_sha256": "9ed6c5c269513323a75b48c96f80fd98f7c7cbc77060148bafda48caa4abed5b",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6452370139304549,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6483834,
+ "backend_with_key_registry_bytes": 6687782,
+ "backend_with_datrie_key_index_bytes": 6592426,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.670618,
+ "warm_p50_ms": 0.07047,
+ "warm_p95_ms": 0.076616,
+ "warm_p99_ms": 0.09191052999999999,
+ "samples_ms": [
+ 0.076616,
+ 0.076616,
+ 0.076616,
+ 0.073823,
+ 0.074731,
+ 0.07473,
+ 0.07473,
+ 0.07473,
+ 0.07473,
+ 0.074311,
+ 0.073823,
+ 0.069981,
+ 0.07047,
+ 0.069562,
+ 0.07054,
+ 0.069142,
+ 0.069911,
+ 0.0704,
+ 0.069143,
+ 0.069073,
+ 0.069003,
+ 0.07047,
+ 0.070051,
+ 0.07047,
+ 0.071447,
+ 0.101689,
+ 0.073333,
+ 0.070889,
+ 0.07047,
+ 0.069143,
+ 0.068585,
+ 0.069981,
+ 0.069562,
+ 0.069632,
+ 0.069003,
+ 0.0704,
+ 0.069003,
+ 0.070539,
+ 0.070121,
+ 0.069492
+ ],
+ "warm_rows_per_second": 7265503.050943663
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 6482626,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 839,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 839,
+ "zlib_blocks": 46,
+ "xor_blocks": 793,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "table_sha256": "9ed6c5c269513323a75b48c96f80fd98f7c7cbc77060148bafda48caa4abed5b",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6974755520932376,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6483834,
+ "backend_with_key_registry_bytes": 6687782,
+ "backend_with_datrie_key_index_bytes": 6592426,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.409901,
+ "warm_p50_ms": 0.07099349999999999,
+ "warm_p95_ms": 0.07664045,
+ "warm_p99_ms": 0.09095139,
+ "samples_ms": [
+ 0.077105,
+ 0.075219,
+ 0.076616,
+ 0.07473,
+ 0.074241,
+ 0.07473,
+ 0.075219,
+ 0.072356,
+ 0.074731,
+ 0.074731,
+ 0.072356,
+ 0.071378,
+ 0.072356,
+ 0.07047,
+ 0.07047,
+ 0.070959,
+ 0.071448,
+ 0.070051,
+ 0.069562,
+ 0.069982,
+ 0.069981,
+ 0.071028,
+ 0.07054,
+ 0.069562,
+ 0.099804,
+ 0.074312,
+ 0.071448,
+ 0.071448,
+ 0.070051,
+ 0.070051,
+ 0.069981,
+ 0.07047,
+ 0.070959,
+ 0.070051,
+ 0.07047,
+ 0.07054,
+ 0.070958,
+ 0.07047,
+ 0.0704,
+ 0.071378
+ ],
+ "warm_rows_per_second": 7211927.852549883
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 4080768,
+ "unique_rows": 8271,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 517,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 517,
+ "zlib_blocks": 361,
+ "xor_blocks": 156,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8271
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "1c1d4481afeb929a6c60ffb2dda15be882e5821857e68e1591cdf86a8a23c0e4",
+ "table_sha256": "10c70f4eac16302ceaac2f35eb677e5a8323f44f32c9cba3c656de9460248654",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.38640943402424455,
+ "representatives_sha256": "5a0f623392ded7d50ec694118345b2dc49557b1bebd5e7e9870e49eb422a2185",
+ "structurally_merged_rows": 5153,
+ "backend_artifact_bytes": 4081975,
+ "backend_with_key_registry_bytes": 4285923,
+ "backend_with_datrie_key_index_bytes": 4190567,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900053980881236,
+ "mean_cosine": 0.9985802128091349,
+ "p05_cosine": 0.9930332680282773,
+ "relative_l2_mean": 0.03112516137545862,
+ "relative_l2_p95": 0.1181307410446335,
+ "relative_l2_max": 0.14236738964640794,
+ "weighted_relative_l2_mean": 0.017583539902024107,
+ "weighted_relative_frobenius": 0.009429837428214146,
+ "weighted_mean_cosine": 0.9992397497471761,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900053980881236,
+ "mean_cosine": 0.9985802128091349,
+ "p05_cosine": 0.9930332680282773,
+ "relative_l2_mean": 0.03112516137545862,
+ "relative_l2_p95": 0.1181307410446335,
+ "relative_l2_max": 0.14236738964640794,
+ "weighted_relative_l2_mean": 0.017583539902024107,
+ "weighted_relative_frobenius": 0.009429837428214146,
+ "weighted_mean_cosine": 0.9992397497471761,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.797962,
+ "warm_p50_ms": 0.070051,
+ "warm_p95_ms": 0.0748278,
+ "warm_p99_ms": 0.09078736999999999,
+ "samples_ms": [
+ 0.076686,
+ 0.07473,
+ 0.074312,
+ 0.072845,
+ 0.0704,
+ 0.07054,
+ 0.070051,
+ 0.07054,
+ 0.070051,
+ 0.070051,
+ 0.071378,
+ 0.07047,
+ 0.069143,
+ 0.069073,
+ 0.07047,
+ 0.069981,
+ 0.07047,
+ 0.069632,
+ 0.069492,
+ 0.069004,
+ 0.069632,
+ 0.068165,
+ 0.069004,
+ 0.070051,
+ 0.07054,
+ 0.069562,
+ 0.069981,
+ 0.069561,
+ 0.069073,
+ 0.07047,
+ 0.069492,
+ 0.068654,
+ 0.099803,
+ 0.071936,
+ 0.069562,
+ 0.069631,
+ 0.070051,
+ 0.069911,
+ 0.07054,
+ 0.069562
+ ],
+ "warm_rows_per_second": 7308960.614409502
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3865416,
+ "unique_rows": 7821,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 489,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 489,
+ "zlib_blocks": 378,
+ "xor_blocks": 111,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7821
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "93cf72dbdf8d27b18b36be0d472cbb4701d8b9a4780bf12c65971e4688b1dbbc",
+ "table_sha256": "2433e4e96782083f94e8266b09ac4dcd0cb92e153b8a2b2eba988288843284ca",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.41839837888255715,
+ "representatives_sha256": "66e0123ae35dfc3e15b0374c3ac98431eec0d321e0ded9a545a8f915195de474",
+ "structurally_merged_rows": 5603,
+ "backend_artifact_bytes": 3866623,
+ "backend_with_key_registry_bytes": 4070571,
+ "backend_with_datrie_key_index_bytes": 3975215,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.980085122495785,
+ "mean_cosine": 0.9981056136973616,
+ "p05_cosine": 0.9910258422608778,
+ "relative_l2_mean": 0.036756417401143396,
+ "relative_l2_p95": 0.13431279623149256,
+ "relative_l2_max": 0.19939341253996173,
+ "weighted_relative_l2_mean": 0.020450147654051058,
+ "weighted_relative_frobenius": 0.011197217894442387,
+ "weighted_mean_cosine": 0.998998288504402,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.980085122495785,
+ "mean_cosine": 0.9981056136973616,
+ "p05_cosine": 0.9910258422608778,
+ "relative_l2_mean": 0.036756417401143396,
+ "relative_l2_p95": 0.13431279623149256,
+ "relative_l2_max": 0.19939341253996173,
+ "weighted_relative_l2_mean": 0.020450147654051058,
+ "weighted_relative_frobenius": 0.011197217894442387,
+ "weighted_mean_cosine": 0.998998288504402,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.351396,
+ "warm_p50_ms": 0.07120299999999999,
+ "warm_p95_ms": 0.07812835,
+ "warm_p99_ms": 0.09402954999999999,
+ "samples_ms": [
+ 0.075638,
+ 0.075219,
+ 0.073333,
+ 0.073334,
+ 0.072775,
+ 0.073334,
+ 0.073893,
+ 0.078083,
+ 0.07899,
+ 0.076616,
+ 0.074312,
+ 0.073823,
+ 0.071867,
+ 0.070959,
+ 0.07047,
+ 0.070539,
+ 0.071448,
+ 0.07054,
+ 0.069562,
+ 0.072845,
+ 0.070539,
+ 0.07047,
+ 0.069562,
+ 0.071937,
+ 0.103645,
+ 0.075219,
+ 0.072775,
+ 0.071378,
+ 0.069073,
+ 0.069631,
+ 0.07054,
+ 0.07047,
+ 0.069422,
+ 0.070469,
+ 0.07054,
+ 0.071028,
+ 0.068584,
+ 0.069981,
+ 0.069981,
+ 0.07061
+ ],
+ "warm_rows_per_second": 7190708.256674579
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3836346,
+ "unique_rows": 7760,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 485,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 485,
+ "zlib_blocks": 393,
+ "xor_blocks": 92,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7760
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "3b8d1e499cfda1a058271bc38b401335d3adbba2f23ce3b7c755dd72a698a686",
+ "table_sha256": "7fcc8c95d329446ced6adfab0590a0925f73a2775a7cc06456fee5173d65462f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4337080221157521,
+ "representatives_sha256": "ffcdd62e49e4dc0401a65bc1dd4a3d1d17bdf69c328160ce16d379c2ce01d18a",
+ "structurally_merged_rows": 5664,
+ "backend_artifact_bytes": 3837552,
+ "backend_with_key_registry_bytes": 4041500,
+ "backend_with_datrie_key_index_bytes": 3946144,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.970170421501845,
+ "mean_cosine": 0.9979948273478729,
+ "p05_cosine": 0.9905955432170158,
+ "relative_l2_mean": 0.03776016573206639,
+ "relative_l2_p95": 0.1371075806604864,
+ "relative_l2_max": 0.24352079349033676,
+ "weighted_relative_l2_mean": 0.020934608790724923,
+ "weighted_relative_frobenius": 0.011236415060272766,
+ "weighted_mean_cosine": 0.9989448172515513,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.970170421501845,
+ "mean_cosine": 0.9979948273478729,
+ "p05_cosine": 0.9905955432170158,
+ "relative_l2_mean": 0.03776016573206639,
+ "relative_l2_p95": 0.1371075806604864,
+ "relative_l2_max": 0.24352079349033676,
+ "weighted_relative_l2_mean": 0.020934608790724923,
+ "weighted_relative_frobenius": 0.011236415060272766,
+ "weighted_mean_cosine": 0.9989448172515513,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.50372,
+ "warm_p50_ms": 0.07054,
+ "warm_p95_ms": 0.07667175,
+ "warm_p99_ms": 0.08187248999999999,
+ "samples_ms": [
+ 0.076127,
+ 0.076127,
+ 0.079061,
+ 0.076127,
+ 0.074311,
+ 0.073333,
+ 0.072845,
+ 0.072845,
+ 0.072915,
+ 0.073753,
+ 0.076546,
+ 0.073333,
+ 0.070889,
+ 0.071378,
+ 0.07012,
+ 0.068584,
+ 0.07047,
+ 0.07005,
+ 0.070121,
+ 0.069073,
+ 0.069004,
+ 0.07054,
+ 0.068654,
+ 0.072845,
+ 0.071378,
+ 0.069562,
+ 0.069981,
+ 0.069562,
+ 0.070051,
+ 0.070889,
+ 0.07047,
+ 0.069073,
+ 0.07054,
+ 0.071518,
+ 0.070539,
+ 0.07047,
+ 0.069981,
+ 0.068165,
+ 0.08367,
+ 0.069981
+ ],
+ "warm_rows_per_second": 7258293.166997449
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3830251,
+ "unique_rows": 7748,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 485,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 485,
+ "zlib_blocks": 388,
+ "xor_blocks": 97,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7748
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "5bc5de942dd7680276060d940810a0371630e3326ca5108ae9834d3ce44a6b4e",
+ "table_sha256": "2105fdb09947d30bce3a39f65685885452b87de9d161ac24b8543363364265dc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.36579897115007043,
+ "representatives_sha256": "f61c8cb0a451eacc74c0b072772ba972d2aba61682b94c5d0d847abf3fccc5ec",
+ "structurally_merged_rows": 5676,
+ "backend_artifact_bytes": 3831457,
+ "backend_with_key_registry_bytes": 4035405,
+ "backend_with_datrie_key_index_bytes": 3940049,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9541835879181099,
+ "mean_cosine": 0.9979284129801629,
+ "p05_cosine": 0.990434384937845,
+ "relative_l2_mean": 0.038211001031595915,
+ "relative_l2_p95": 0.13878562080257548,
+ "relative_l2_max": 0.30221514303946656,
+ "weighted_relative_l2_mean": 0.021162645136322683,
+ "weighted_relative_frobenius": 0.011927566717317878,
+ "weighted_mean_cosine": 0.9989112583547245,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9541835879181099,
+ "mean_cosine": 0.9979284129801629,
+ "p05_cosine": 0.990434384937845,
+ "relative_l2_mean": 0.038211001031595915,
+ "relative_l2_p95": 0.13878562080257548,
+ "relative_l2_max": 0.30221514303946656,
+ "weighted_relative_l2_mean": 0.021162645136322683,
+ "weighted_relative_frobenius": 0.011927566717317878,
+ "weighted_mean_cosine": 0.9989112583547245,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.601149,
+ "warm_p50_ms": 0.07050500000000001,
+ "warm_p95_ms": 0.07572895,
+ "warm_p99_ms": 0.09061206,
+ "samples_ms": [
+ 0.07473,
+ 0.074312,
+ 0.072426,
+ 0.071517,
+ 0.072844,
+ 0.072914,
+ 0.071937,
+ 0.073334,
+ 0.072844,
+ 0.071936,
+ 0.070959,
+ 0.075708,
+ 0.076127,
+ 0.0748,
+ 0.071867,
+ 0.070959,
+ 0.069073,
+ 0.07054,
+ 0.069632,
+ 0.07047,
+ 0.069492,
+ 0.069143,
+ 0.07047,
+ 0.069492,
+ 0.099873,
+ 0.072914,
+ 0.071029,
+ 0.0704,
+ 0.069073,
+ 0.068654,
+ 0.067607,
+ 0.068723,
+ 0.068096,
+ 0.069073,
+ 0.068654,
+ 0.069493,
+ 0.068165,
+ 0.070051,
+ 0.069981,
+ 0.0704
+ ],
+ "warm_rows_per_second": 7261896.319409969
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3829955,
+ "unique_rows": 7747,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 485,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 485,
+ "zlib_blocks": 375,
+ "xor_blocks": 110,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7747
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "a09f6b89ed66dd418433bf89edcba90e16e2a747a7f916bd356ca9134c909cbe",
+ "table_sha256": "39ff3027b70891a0029412f7089d882ba34b74c4c3c0f307f90065aa8fab2cd8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5347587501164526,
+ "representatives_sha256": "51455a8f97507a51849fcf8a5d576f6a407a8ecdbaaabde92e2d7f69b81cc634",
+ "structurally_merged_rows": 5677,
+ "backend_artifact_bytes": 3831162,
+ "backend_with_key_registry_bytes": 4035110,
+ "backend_with_datrie_key_index_bytes": 3939754,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9018038404486557,
+ "mean_cosine": 0.9978519567597138,
+ "p05_cosine": 0.9902982040263192,
+ "relative_l2_mean": 0.038567436474365495,
+ "relative_l2_p95": 0.13939661712678836,
+ "relative_l2_max": 0.4441858606483749,
+ "weighted_relative_l2_mean": 0.02147655604447714,
+ "weighted_relative_frobenius": 0.013740973656759222,
+ "weighted_mean_cosine": 0.9988484043357869,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9018038404486557,
+ "mean_cosine": 0.9978519567597138,
+ "p05_cosine": 0.9902982040263192,
+ "relative_l2_mean": 0.038567436474365495,
+ "relative_l2_p95": 0.13939661712678836,
+ "relative_l2_max": 0.4441858606483749,
+ "weighted_relative_l2_mean": 0.02147655604447714,
+ "weighted_relative_frobenius": 0.013740973656759222,
+ "weighted_mean_cosine": 0.9988484043357869,
+ "max_relative_norm_change": 0.010098739162329335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.636209,
+ "warm_p50_ms": 0.071937,
+ "warm_p95_ms": 0.07531329999999999,
+ "warm_p99_ms": 0.08264379999999999,
+ "samples_ms": [
+ 0.0748,
+ 0.072915,
+ 0.072425,
+ 0.077105,
+ 0.072356,
+ 0.074312,
+ 0.072845,
+ 0.072356,
+ 0.072914,
+ 0.072355,
+ 0.069562,
+ 0.07054,
+ 0.069981,
+ 0.069982,
+ 0.069981,
+ 0.07054,
+ 0.07061,
+ 0.071098,
+ 0.074241,
+ 0.072844,
+ 0.071937,
+ 0.069632,
+ 0.071937,
+ 0.086185,
+ 0.072426,
+ 0.071937,
+ 0.074241,
+ 0.072425,
+ 0.070959,
+ 0.075219,
+ 0.073333,
+ 0.071448,
+ 0.069911,
+ 0.069073,
+ 0.071517,
+ 0.071448,
+ 0.071029,
+ 0.068584,
+ 0.071378,
+ 0.069562
+ ],
+ "warm_rows_per_second": 7117338.7825458385
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 3436544,
+ "unique_rows": 13424,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "96287e438559898544cb2b53c3e3e1b1fc356fa0de75846da54b3e109092a525",
+ "table_sha256": "2b80ff97b7f3a0bdc95ac7eead482fad4fb6e9e7e25c67eb90f9c67d3d9aefef",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.046544023090973496,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3437350,
+ "backend_with_key_registry_bytes": 3641298,
+ "backend_with_datrie_key_index_bytes": 3545942,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999624473883,
+ "mean_cosine": 0.9999999790488479,
+ "p05_cosine": 0.9999999732577625,
+ "relative_l2_mean": 0.00020648776297690362,
+ "relative_l2_p95": 0.00023381194563882685,
+ "relative_l2_max": 0.0002755920071979578,
+ "weighted_relative_l2_mean": 0.0002061626673808047,
+ "weighted_relative_frobenius": 0.00020539489681704117,
+ "weighted_mean_cosine": 0.999999979135146,
+ "max_relative_norm_change": 0.00013809462340430532
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.532261,
+ "warm_p50_ms": 0.22,
+ "warm_p95_ms": 0.22404074999999996,
+ "warm_p99_ms": 0.24538284999999999,
+ "samples_ms": [
+ 0.223283,
+ 0.221886,
+ 0.220489,
+ 0.219022,
+ 0.22,
+ 0.218603,
+ 0.22,
+ 0.220489,
+ 0.22,
+ 0.220489,
+ 0.220001,
+ 0.220908,
+ 0.21993,
+ 0.238438,
+ 0.220978,
+ 0.219162,
+ 0.219581,
+ 0.219582,
+ 0.22042,
+ 0.219023,
+ 0.218953,
+ 0.219512,
+ 0.219512,
+ 0.220001,
+ 0.22,
+ 0.219162,
+ 0.219092,
+ 0.219023,
+ 0.218534,
+ 0.219512,
+ 0.22,
+ 0.249823,
+ 0.222864,
+ 0.220419,
+ 0.219093,
+ 0.219582,
+ 0.220559,
+ 0.220489,
+ 0.22,
+ 0.219022
+ ],
+ "warm_rows_per_second": 2327272.727272727
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 3248746,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 420,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 420,
+ "zlib_blocks": 0,
+ "xor_blocks": 420,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "96287e438559898544cb2b53c3e3e1b1fc356fa0de75846da54b3e109092a525",
+ "table_sha256": "ca615eea8d5329ec5c737f7f154aa51f20547cef4211595c8cc9433f3a5f8f4b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3372036158107221,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3249955,
+ "backend_with_key_registry_bytes": 3453903,
+ "backend_with_datrie_key_index_bytes": 3358547,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999624473883,
+ "mean_cosine": 0.9999999790488479,
+ "p05_cosine": 0.9999999732577625,
+ "relative_l2_mean": 0.00020648776297690362,
+ "relative_l2_p95": 0.00023381194563882685,
+ "relative_l2_max": 0.0002755920071979578,
+ "weighted_relative_l2_mean": 0.0002061626673808047,
+ "weighted_relative_frobenius": 0.00020539489681704117,
+ "weighted_mean_cosine": 0.999999979135146,
+ "max_relative_norm_change": 0.00013809462340430532
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.196826,
+ "warm_p50_ms": 0.23019699999999998,
+ "warm_p95_ms": 0.24275135,
+ "warm_p99_ms": 0.25060545,
+ "samples_ms": [
+ 0.232292,
+ 0.231385,
+ 0.254991,
+ 0.234667,
+ 0.229918,
+ 0.229429,
+ 0.229988,
+ 0.229499,
+ 0.228521,
+ 0.228521,
+ 0.229987,
+ 0.229499,
+ 0.229429,
+ 0.22901,
+ 0.22894,
+ 0.230476,
+ 0.228451,
+ 0.230895,
+ 0.229499,
+ 0.242699,
+ 0.230407,
+ 0.230476,
+ 0.230407,
+ 0.232292,
+ 0.231873,
+ 0.232292,
+ 0.230406,
+ 0.228521,
+ 0.22859,
+ 0.228102,
+ 0.229987,
+ 0.229918,
+ 0.229918,
+ 0.232712,
+ 0.231874,
+ 0.232363,
+ 0.230896,
+ 0.243746,
+ 0.231314,
+ 0.22901
+ ],
+ "warm_rows_per_second": 2224181.896375713
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 3248746,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 420,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 420,
+ "zlib_blocks": 0,
+ "xor_blocks": 420,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "96287e438559898544cb2b53c3e3e1b1fc356fa0de75846da54b3e109092a525",
+ "table_sha256": "ca615eea8d5329ec5c737f7f154aa51f20547cef4211595c8cc9433f3a5f8f4b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.37777592497877777,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3249954,
+ "backend_with_key_registry_bytes": 3453902,
+ "backend_with_datrie_key_index_bytes": 3358546,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999624473883,
+ "mean_cosine": 0.9999999790488479,
+ "p05_cosine": 0.9999999732577625,
+ "relative_l2_mean": 0.00020648776297690362,
+ "relative_l2_p95": 0.00023381194563882685,
+ "relative_l2_max": 0.0002755920071979578,
+ "weighted_relative_l2_mean": 0.0002061626673808047,
+ "weighted_relative_frobenius": 0.00020539489681704117,
+ "weighted_mean_cosine": 0.999999979135146,
+ "max_relative_norm_change": 0.00013809462340430532
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.136692,
+ "warm_p50_ms": 0.229499,
+ "warm_p95_ms": 0.23381474999999996,
+ "warm_p99_ms": 0.25277088,
+ "samples_ms": [
+ 0.232851,
+ 0.232222,
+ 0.230895,
+ 0.231454,
+ 0.231314,
+ 0.230407,
+ 0.229918,
+ 0.228939,
+ 0.258273,
+ 0.23327,
+ 0.22859,
+ 0.229499,
+ 0.229499,
+ 0.228521,
+ 0.228521,
+ 0.228521,
+ 0.229499,
+ 0.229917,
+ 0.229918,
+ 0.22894,
+ 0.229009,
+ 0.229499,
+ 0.228521,
+ 0.22901,
+ 0.229918,
+ 0.244165,
+ 0.231384,
+ 0.231873,
+ 0.22901,
+ 0.228102,
+ 0.230337,
+ 0.228521,
+ 0.229009,
+ 0.228521,
+ 0.229009,
+ 0.229429,
+ 0.229918,
+ 0.228451,
+ 0.229359,
+ 0.230407
+ ],
+ "warm_rows_per_second": 2230946.54007207
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 3248746,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 420,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 420,
+ "zlib_blocks": 0,
+ "xor_blocks": 420,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "96287e438559898544cb2b53c3e3e1b1fc356fa0de75846da54b3e109092a525",
+ "table_sha256": "ca615eea8d5329ec5c737f7f154aa51f20547cef4211595c8cc9433f3a5f8f4b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3295996389351785,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3249954,
+ "backend_with_key_registry_bytes": 3453902,
+ "backend_with_datrie_key_index_bytes": 3358546,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999624473883,
+ "mean_cosine": 0.9999999790488479,
+ "p05_cosine": 0.9999999732577625,
+ "relative_l2_mean": 0.00020648776297690362,
+ "relative_l2_p95": 0.00023381194563882685,
+ "relative_l2_max": 0.0002755920071979578,
+ "weighted_relative_l2_mean": 0.0002061626673808047,
+ "weighted_relative_frobenius": 0.00020539489681704117,
+ "weighted_mean_cosine": 0.999999979135146,
+ "max_relative_norm_change": 0.00013809462340430532
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.164629,
+ "warm_p50_ms": 0.23019699999999998,
+ "warm_p95_ms": 0.23499874999999998,
+ "warm_p99_ms": 0.24394329999999997,
+ "samples_ms": [
+ 0.234178,
+ 0.230966,
+ 0.22908,
+ 0.23327,
+ 0.231873,
+ 0.230895,
+ 0.229917,
+ 0.229918,
+ 0.229918,
+ 0.228451,
+ 0.231315,
+ 0.231874,
+ 0.229429,
+ 0.241302,
+ 0.232712,
+ 0.231804,
+ 0.232293,
+ 0.231385,
+ 0.228521,
+ 0.230895,
+ 0.229009,
+ 0.231385,
+ 0.234667,
+ 0.229918,
+ 0.228521,
+ 0.22894,
+ 0.229988,
+ 0.229498,
+ 0.229918,
+ 0.231315,
+ 0.245632,
+ 0.23327,
+ 0.230406,
+ 0.229499,
+ 0.22901,
+ 0.234178,
+ 0.228101,
+ 0.22908,
+ 0.229498,
+ 0.22908
+ ],
+ "warm_rows_per_second": 2224181.896375713
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 2095833,
+ "unique_rows": 8273,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 259,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 259,
+ "zlib_blocks": 117,
+ "xor_blocks": 142,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8273
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "56eaf00fd221231d5d3188c9ed0859c7ba0230673f6ccb33a1cace19834895fc",
+ "table_sha256": "92e55c0c7ba1f125aa517d3fa10a34f6fa879ed591f54e5519e4db8fe8fe9a5c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2195942709222436,
+ "representatives_sha256": "ed176ce71e84167bce823c740905ba40eabe50ed8eadf74fb5130a8e2c8c3fbf",
+ "structurally_merged_rows": 5151,
+ "backend_artifact_bytes": 2097040,
+ "backend_with_key_registry_bytes": 2300988,
+ "backend_with_datrie_key_index_bytes": 2205632,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900024178532458,
+ "mean_cosine": 0.9985803974928634,
+ "p05_cosine": 0.9930321598848577,
+ "relative_l2_mean": 0.031243688802523692,
+ "relative_l2_p95": 0.11812890288280987,
+ "relative_l2_max": 0.1423848947835388,
+ "weighted_relative_l2_mean": 0.017730440592170018,
+ "weighted_relative_frobenius": 0.009429329011061569,
+ "weighted_mean_cosine": 0.9992401131688717,
+ "max_relative_norm_change": 0.010144415505645151
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.990000929996695,
+ "mean_cosine": 0.9985803997905448,
+ "p05_cosine": 0.993032265603851,
+ "relative_l2_mean": 0.03111666172324363,
+ "relative_l2_p95": 0.11812310079870236,
+ "relative_l2_max": 0.14239308219003036,
+ "weighted_relative_l2_mean": 0.017573317428987902,
+ "weighted_relative_frobenius": 0.009427344865478818,
+ "weighted_mean_cosine": 0.9992401221857902,
+ "max_relative_norm_change": 0.01009549559175055
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.387586,
+ "warm_p50_ms": 0.228102,
+ "warm_p95_ms": 0.23965684999999998,
+ "warm_p99_ms": 0.25161237,
+ "samples_ms": [
+ 0.245563,
+ 0.231455,
+ 0.230477,
+ 0.229009,
+ 0.226635,
+ 0.229079,
+ 0.22901,
+ 0.22852,
+ 0.228102,
+ 0.232781,
+ 0.227474,
+ 0.228032,
+ 0.226635,
+ 0.229009,
+ 0.227683,
+ 0.227124,
+ 0.227543,
+ 0.25548,
+ 0.230895,
+ 0.227613,
+ 0.228101,
+ 0.228102,
+ 0.228102,
+ 0.228102,
+ 0.227124,
+ 0.227612,
+ 0.228451,
+ 0.227543,
+ 0.226985,
+ 0.227124,
+ 0.232292,
+ 0.232362,
+ 0.231314,
+ 0.228032,
+ 0.228101,
+ 0.239346,
+ 0.228451,
+ 0.228451,
+ 0.226565,
+ 0.23327
+ ],
+ "warm_rows_per_second": 2244609.86751541
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1989999,
+ "unique_rows": 7822,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 245,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 245,
+ "zlib_blocks": 142,
+ "xor_blocks": 103,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7822
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "f8f49b81a78fba298b146b9b3466b63acacabde65223e4c097a9958c7340fc89",
+ "table_sha256": "07df28a754d07375c6d041150d8ad49c0c212d4f435a8f0f21a4cd97efb42fa2",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21171365305781364,
+ "representatives_sha256": "c2fe82c139ea01fe20fd09fc6180a3ea7206d946373ee9ad388752b37ed24a6f",
+ "structurally_merged_rows": 5602,
+ "backend_artifact_bytes": 1991206,
+ "backend_with_key_registry_bytes": 2195154,
+ "backend_with_datrie_key_index_bytes": 2099798,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9800888265045965,
+ "mean_cosine": 0.9981042407515186,
+ "p05_cosine": 0.9910205144552796,
+ "relative_l2_mean": 0.03688405382576256,
+ "relative_l2_p95": 0.13436454643040865,
+ "relative_l2_max": 0.19937838717662318,
+ "weighted_relative_l2_mean": 0.0206012343269577,
+ "weighted_relative_frobenius": 0.011197578046986576,
+ "weighted_mean_cosine": 0.9989979012517765,
+ "max_relative_norm_change": 0.010144415505645151
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9800901317835918,
+ "mean_cosine": 0.998104240470195,
+ "p05_cosine": 0.9910197197482495,
+ "relative_l2_mean": 0.03676396341980959,
+ "relative_l2_p95": 0.13435857381636385,
+ "relative_l2_max": 0.19938918351590493,
+ "weighted_relative_l2_mean": 0.020447662594474707,
+ "weighted_relative_frobenius": 0.01119595314295581,
+ "weighted_mean_cosine": 0.998997908979883,
+ "max_relative_norm_change": 0.01009549559175055
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.706283,
+ "warm_p50_ms": 0.228591,
+ "warm_p95_ms": 0.24009024999999998,
+ "warm_p99_ms": 0.25523152,
+ "samples_ms": [
+ 0.233759,
+ 0.231384,
+ 0.228032,
+ 0.263512,
+ 0.231385,
+ 0.229988,
+ 0.229917,
+ 0.228521,
+ 0.235156,
+ 0.229429,
+ 0.229429,
+ 0.227124,
+ 0.228521,
+ 0.227613,
+ 0.228521,
+ 0.22901,
+ 0.227124,
+ 0.228521,
+ 0.228521,
+ 0.228102,
+ 0.22901,
+ 0.24228,
+ 0.228032,
+ 0.228521,
+ 0.22901,
+ 0.228521,
+ 0.229009,
+ 0.22901,
+ 0.228591,
+ 0.228591,
+ 0.228101,
+ 0.227543,
+ 0.227543,
+ 0.22894,
+ 0.228032,
+ 0.22894,
+ 0.228032,
+ 0.228032,
+ 0.239975,
+ 0.230477
+ ],
+ "warm_rows_per_second": 2239808.216421469
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1975641,
+ "unique_rows": 7761,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 243,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 243,
+ "zlib_blocks": 147,
+ "xor_blocks": 96,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7761
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "268022c62540b0526f85a737bbf91a69cf7b1c17d298b3e74a00a8428cff1ac9",
+ "table_sha256": "f8f3aa3e3bf40280fac9cc141ddb4aa69866396e39e3f7423472b342d4c62aef",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21183035802096128,
+ "representatives_sha256": "8b939ab388426fe7cdd6a6cd2e1869f3f58ad8a2a20b6f35de3f206a8ffbeb98",
+ "structurally_merged_rows": 5663,
+ "backend_artifact_bytes": 1976847,
+ "backend_with_key_registry_bytes": 2180795,
+ "backend_with_datrie_key_index_bytes": 2085439,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9701754657329307,
+ "mean_cosine": 0.9979934603727224,
+ "p05_cosine": 0.9905709174976577,
+ "relative_l2_mean": 0.03788685498270028,
+ "relative_l2_p95": 0.13755797514434065,
+ "relative_l2_max": 0.2435025113904524,
+ "weighted_relative_l2_mean": 0.021085238308215817,
+ "weighted_relative_frobenius": 0.011236771970957654,
+ "weighted_mean_cosine": 0.9989444328806925,
+ "max_relative_norm_change": 0.010144415505645151
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9701766525044704,
+ "mean_cosine": 0.9979934598733242,
+ "p05_cosine": 0.9905711754792335,
+ "relative_l2_mean": 0.03776768677619807,
+ "relative_l2_p95": 0.13755447281450678,
+ "relative_l2_max": 0.2435007667911337,
+ "weighted_relative_l2_mean": 0.02093211167713961,
+ "weighted_relative_frobenius": 0.011235152675622983,
+ "weighted_mean_cosine": 0.9989444405035449,
+ "max_relative_norm_change": 0.01009549559175055
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.844988,
+ "warm_p50_ms": 0.22855599999999998,
+ "warm_p95_ms": 0.23528159999999995,
+ "warm_p99_ms": 0.25215681,
+ "samples_ms": [
+ 0.234667,
+ 0.229428,
+ 0.228521,
+ 0.229009,
+ 0.228521,
+ 0.230407,
+ 0.22894,
+ 0.246959,
+ 0.230406,
+ 0.229918,
+ 0.228521,
+ 0.234248,
+ 0.228521,
+ 0.228032,
+ 0.228521,
+ 0.228032,
+ 0.228032,
+ 0.229499,
+ 0.228521,
+ 0.228591,
+ 0.226636,
+ 0.22901,
+ 0.227054,
+ 0.22894,
+ 0.25548,
+ 0.231385,
+ 0.228521,
+ 0.22901,
+ 0.228032,
+ 0.227613,
+ 0.228101,
+ 0.22908,
+ 0.22852,
+ 0.228032,
+ 0.228102,
+ 0.228102,
+ 0.22908,
+ 0.228591,
+ 0.228102,
+ 0.22901
+ ],
+ "warm_rows_per_second": 2240151.2102066893
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1972179,
+ "unique_rows": 7748,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 243,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 243,
+ "zlib_blocks": 145,
+ "xor_blocks": 98,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7748
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "63cfcd51abe0eda4ed238fd06eb26647a67587409413bfb7ac7f699c9486f1fa",
+ "table_sha256": "a81a1394f0778cafd5297072928a55e30a1bcefd827ceeeba2d960b9b5431172",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21017833007499576,
+ "representatives_sha256": "c481cae84367913420e59ee097fd0ddf5764e1de3671955bd01d640d0f37577a",
+ "structurally_merged_rows": 5676,
+ "backend_artifact_bytes": 1973385,
+ "backend_with_key_registry_bytes": 2177333,
+ "backend_with_datrie_key_index_bytes": 2081977,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.954181865359993,
+ "mean_cosine": 0.9979243542256204,
+ "p05_cosine": 0.9903884267258661,
+ "relative_l2_mean": 0.03835742068413006,
+ "relative_l2_p95": 0.13892009584453222,
+ "relative_l2_max": 0.3022227389598215,
+ "weighted_relative_l2_mean": 0.021322789291119113,
+ "weighted_relative_frobenius": 0.01192848331866381,
+ "weighted_mean_cosine": 0.9989095748776344,
+ "max_relative_norm_change": 0.010144415505645151
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9541761878334689,
+ "mean_cosine": 0.9979243561361446,
+ "p05_cosine": 0.9903923922398393,
+ "relative_l2_mean": 0.038238458244003334,
+ "relative_l2_p95": 0.1389393631537763,
+ "relative_l2_max": 0.30224075637297,
+ "weighted_relative_l2_mean": 0.021169770600335598,
+ "weighted_relative_frobenius": 0.011926915414780088,
+ "weighted_mean_cosine": 0.998909583596941,
+ "max_relative_norm_change": 0.01009549559175055
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.459324,
+ "warm_p50_ms": 0.228521,
+ "warm_p95_ms": 0.2455865,
+ "warm_p99_ms": 0.24980045,
+ "samples_ms": [
+ 0.245562,
+ 0.230825,
+ 0.22901,
+ 0.22901,
+ 0.22894,
+ 0.230896,
+ 0.229009,
+ 0.228521,
+ 0.228521,
+ 0.228591,
+ 0.228591,
+ 0.227613,
+ 0.227612,
+ 0.227543,
+ 0.227473,
+ 0.228451,
+ 0.227962,
+ 0.252197,
+ 0.232852,
+ 0.229499,
+ 0.229079,
+ 0.228032,
+ 0.229009,
+ 0.229499,
+ 0.228032,
+ 0.227613,
+ 0.227124,
+ 0.229499,
+ 0.226635,
+ 0.228451,
+ 0.227962,
+ 0.22894,
+ 0.227962,
+ 0.228451,
+ 0.227054,
+ 0.246052,
+ 0.228032,
+ 0.228032,
+ 0.228521,
+ 0.22901
+ ],
+ "warm_rows_per_second": 2240494.309056936
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1971971,
+ "unique_rows": 7747,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 243,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 242,
+ "zlib_blocks": 143,
+ "xor_blocks": 99,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7747
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "4c4c6b90e7c737fd6c96215bd44b232c78279c6840afb843c1f8d5faadd708dd",
+ "table_sha256": "591e7973ade21bfc369f95299b23a1cc8d7fab6a0dbcc8f687538101027bbe9c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2120769671164453,
+ "representatives_sha256": "61e1e77f028074e2b570d2862e46e8d8dd4fc04ef591847983ec72799b151b5f",
+ "structurally_merged_rows": 5677,
+ "backend_artifact_bytes": 1973177,
+ "backend_with_key_registry_bytes": 2177125,
+ "backend_with_datrie_key_index_bytes": 2081769,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.901805312261854,
+ "mean_cosine": 0.997847901598986,
+ "p05_cosine": 0.9902505659728731,
+ "relative_l2_mean": 0.03871382576390896,
+ "relative_l2_p95": 0.13970869337341724,
+ "relative_l2_max": 0.4441902891025155,
+ "weighted_relative_l2_mean": 0.02163662011637838,
+ "weighted_relative_frobenius": 0.013741876451980672,
+ "weighted_mean_cosine": 0.9988467197431203,
+ "max_relative_norm_change": 0.010144415505645151
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9017986396335538,
+ "mean_cosine": 0.9978479037967246,
+ "p05_cosine": 0.9902529993504882,
+ "relative_l2_mean": 0.03859487242884875,
+ "relative_l2_p95": 0.1396913321348921,
+ "relative_l2_max": 0.4441979481016298,
+ "weighted_relative_l2_mean": 0.02148367721683346,
+ "weighted_relative_frobenius": 0.01374048403282683,
+ "weighted_mean_cosine": 0.9988467293545563,
+ "max_relative_norm_change": 0.01009549559175055
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.392626,
+ "warm_p50_ms": 0.2289745,
+ "warm_p95_ms": 0.24229995,
+ "warm_p99_ms": 0.26868643999999997,
+ "samples_ms": [
+ 0.233689,
+ 0.230895,
+ 0.229499,
+ 0.230896,
+ 0.234667,
+ 0.22894,
+ 0.229009,
+ 0.22901,
+ 0.229009,
+ 0.228032,
+ 0.229009,
+ 0.22901,
+ 0.234248,
+ 0.285302,
+ 0.227612,
+ 0.228102,
+ 0.242698,
+ 0.232292,
+ 0.226705,
+ 0.228102,
+ 0.22901,
+ 0.227543,
+ 0.228102,
+ 0.227962,
+ 0.226565,
+ 0.232781,
+ 0.22908,
+ 0.227613,
+ 0.228101,
+ 0.231803,
+ 0.22901,
+ 0.228033,
+ 0.228521,
+ 0.242279,
+ 0.228521,
+ 0.228032,
+ 0.228521,
+ 0.228032,
+ 0.228451,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2236056.8534924197
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 132,
+ "storage": {
+ "artifact_bytes": 1771968,
+ "unique_rows": 13424,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "936ada1dd32ce053f2eae8e789655b9755eedb813052ae37096c04d212ea1676",
+ "table_sha256": "2f4c5b511ef141ff93c2e65b8dfe703922f9a8ceb888373c3404c3637462433c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.03555343905463815,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1772818,
+ "backend_with_key_registry_bytes": 1976766,
+ "backend_with_datrie_key_index_bytes": 1881410,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999343570082099,
+ "mean_cosine": 0.9999791548061561,
+ "p05_cosine": 0.9999695758894899,
+ "relative_l2_mean": 0.006434222600084341,
+ "relative_l2_p95": 0.007831241239552863,
+ "relative_l2_max": 0.01149384799161637,
+ "weighted_relative_l2_mean": 0.006429317675511283,
+ "weighted_relative_frobenius": 0.006447832497899354,
+ "weighted_mean_cosine": 0.9999791906665058,
+ "max_relative_norm_change": 0.002474475834241514
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.370927,
+ "warm_p50_ms": 0.121593,
+ "warm_p95_ms": 0.12580234999999995,
+ "warm_p99_ms": 0.15032884999999999,
+ "samples_ms": [
+ 0.123969,
+ 0.122502,
+ 0.121035,
+ 0.121035,
+ 0.120197,
+ 0.143384,
+ 0.122502,
+ 0.121594,
+ 0.121594,
+ 0.121105,
+ 0.121105,
+ 0.122991,
+ 0.120686,
+ 0.120197,
+ 0.121663,
+ 0.121524,
+ 0.121035,
+ 0.120616,
+ 0.121594,
+ 0.121594,
+ 0.121105,
+ 0.121105,
+ 0.121594,
+ 0.121105,
+ 0.121105,
+ 0.121593,
+ 0.121594,
+ 0.121105,
+ 0.121594,
+ 0.121175,
+ 0.121175,
+ 0.122013,
+ 0.121036,
+ 0.122502,
+ 0.122013,
+ 0.121593,
+ 0.122991,
+ 0.154769,
+ 0.124877,
+ 0.121524
+ ],
+ "warm_rows_per_second": 4210768.712014671
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1650858,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 217,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 217,
+ "zlib_blocks": 0,
+ "xor_blocks": 217,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "936ada1dd32ce053f2eae8e789655b9755eedb813052ae37096c04d212ea1676",
+ "table_sha256": "4bc4e16333c3920adf5e486738f0bbec7da147d3979e429464b66012c4b57002",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.20863979496061802,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1652111,
+ "backend_with_key_registry_bytes": 1856059,
+ "backend_with_datrie_key_index_bytes": 1760703,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999343570082099,
+ "mean_cosine": 0.9999791548061561,
+ "p05_cosine": 0.9999695758894899,
+ "relative_l2_mean": 0.006434222600084341,
+ "relative_l2_p95": 0.007831241239552863,
+ "relative_l2_max": 0.01149384799161637,
+ "weighted_relative_l2_mean": 0.006429317675511283,
+ "weighted_relative_frobenius": 0.006447832497899354,
+ "weighted_mean_cosine": 0.9999791906665058,
+ "max_relative_norm_change": 0.002474475834241514
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.087396,
+ "warm_p50_ms": 0.132245,
+ "warm_p95_ms": 0.1377069,
+ "warm_p99_ms": 0.15133132,
+ "samples_ms": [
+ 0.13682,
+ 0.135353,
+ 0.132908,
+ 0.132489,
+ 0.131511,
+ 0.132001,
+ 0.132978,
+ 0.133397,
+ 0.131512,
+ 0.133397,
+ 0.133397,
+ 0.133397,
+ 0.132489,
+ 0.133467,
+ 0.159448,
+ 0.137658,
+ 0.133467,
+ 0.131511,
+ 0.133397,
+ 0.133397,
+ 0.131092,
+ 0.131511,
+ 0.138636,
+ 0.131511,
+ 0.131092,
+ 0.131581,
+ 0.131092,
+ 0.130114,
+ 0.131511,
+ 0.131511,
+ 0.132489,
+ 0.132908,
+ 0.131022,
+ 0.131092,
+ 0.131092,
+ 0.132001,
+ 0.131512,
+ 0.135283,
+ 0.131512,
+ 0.131023
+ ],
+ "warm_rows_per_second": 3871601.9509244207
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1650858,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 217,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 217,
+ "zlib_blocks": 0,
+ "xor_blocks": 217,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "936ada1dd32ce053f2eae8e789655b9755eedb813052ae37096c04d212ea1676",
+ "table_sha256": "4bc4e16333c3920adf5e486738f0bbec7da147d3979e429464b66012c4b57002",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24100860790349543,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1652110,
+ "backend_with_key_registry_bytes": 1856058,
+ "backend_with_datrie_key_index_bytes": 1760702,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999343570082099,
+ "mean_cosine": 0.9999791548061561,
+ "p05_cosine": 0.9999695758894899,
+ "relative_l2_mean": 0.006434222600084341,
+ "relative_l2_p95": 0.007831241239552863,
+ "relative_l2_max": 0.01149384799161637,
+ "weighted_relative_l2_mean": 0.006429317675511283,
+ "weighted_relative_frobenius": 0.006447832497899354,
+ "weighted_mean_cosine": 0.9999791906665058,
+ "max_relative_norm_change": 0.002474475834241514
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.986545,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.18123789999999998,
+ "warm_p99_ms": 0.20799994999999996,
+ "samples_ms": [
+ 0.182566,
+ 0.224261,
+ 0.181168,
+ 0.153231,
+ 0.138076,
+ 0.135282,
+ 0.133466,
+ 0.132978,
+ 0.131511,
+ 0.133397,
+ 0.131511,
+ 0.131511,
+ 0.130185,
+ 0.132001,
+ 0.131023,
+ 0.132489,
+ 0.131511,
+ 0.131581,
+ 0.131023,
+ 0.131512,
+ 0.131023,
+ 0.131511,
+ 0.130533,
+ 0.131581,
+ 0.130533,
+ 0.131512,
+ 0.130185,
+ 0.130115,
+ 0.130115,
+ 0.130534,
+ 0.147505,
+ 0.132,
+ 0.132,
+ 0.129137,
+ 0.130603,
+ 0.132419,
+ 0.131511,
+ 0.131092,
+ 0.128718,
+ 0.129696
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1650858,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 217,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 217,
+ "zlib_blocks": 0,
+ "xor_blocks": 217,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "936ada1dd32ce053f2eae8e789655b9755eedb813052ae37096c04d212ea1676",
+ "table_sha256": "4bc4e16333c3920adf5e486738f0bbec7da147d3979e429464b66012c4b57002",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.20534768304787576,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1652110,
+ "backend_with_key_registry_bytes": 1856058,
+ "backend_with_datrie_key_index_bytes": 1760702,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999343570082099,
+ "mean_cosine": 0.9999791548061561,
+ "p05_cosine": 0.9999695758894899,
+ "relative_l2_mean": 0.006434222600084341,
+ "relative_l2_p95": 0.007831241239552863,
+ "relative_l2_max": 0.01149384799161637,
+ "weighted_relative_l2_mean": 0.006429317675511283,
+ "weighted_relative_frobenius": 0.006447832497899354,
+ "weighted_mean_cosine": 0.9999791906665058,
+ "max_relative_norm_change": 0.002474475834241514
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.005053,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.13691669999999997,
+ "warm_p99_ms": 0.16284735,
+ "samples_ms": [
+ 0.13633,
+ 0.132489,
+ 0.132978,
+ 0.132,
+ 0.132,
+ 0.133397,
+ 0.172299,
+ 0.132419,
+ 0.130114,
+ 0.131092,
+ 0.130044,
+ 0.131511,
+ 0.130604,
+ 0.130115,
+ 0.131511,
+ 0.130604,
+ 0.131023,
+ 0.130604,
+ 0.131092,
+ 0.130603,
+ 0.131511,
+ 0.131022,
+ 0.132909,
+ 0.132489,
+ 0.131022,
+ 0.132001,
+ 0.131581,
+ 0.131022,
+ 0.131511,
+ 0.130603,
+ 0.131092,
+ 0.131022,
+ 0.130603,
+ 0.132489,
+ 0.130604,
+ 0.148064,
+ 0.134375,
+ 0.13249,
+ 0.131511,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1121086,
+ "unique_rows": 8260,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 134,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 134,
+ "zlib_blocks": 1,
+ "xor_blocks": 133,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8260
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "7558a424db6b844640af2a6a58ef15709a4b584e7c283baeed479d4527cbf50c",
+ "table_sha256": "d36e52ecb6d3234742426ee269924770d3b4c6fafcbf202d37619a482b2f5f38",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.15218486497178674,
+ "representatives_sha256": "82be164cdf157468610b6db2e2257f1d76bdca0c4b9d182a4e2251863e483cb8",
+ "structurally_merged_rows": 5164,
+ "backend_artifact_bytes": 1122335,
+ "backend_with_key_registry_bytes": 1326283,
+ "backend_with_datrie_key_index_bytes": 1230927,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9899331977747489,
+ "mean_cosine": 0.9985628673427508,
+ "p05_cosine": 0.9930390796651419,
+ "relative_l2_mean": 0.035194847701975375,
+ "relative_l2_p95": 0.11802303300346849,
+ "relative_l2_max": 0.14274862236861094,
+ "weighted_relative_l2_mean": 0.022594762018323916,
+ "weighted_relative_frobenius": 0.011437149201513133,
+ "weighted_mean_cosine": 0.9992194986549326,
+ "max_relative_norm_change": 0.012096284539880018
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900021916904238,
+ "mean_cosine": 0.9985678118846913,
+ "p05_cosine": 0.9929912011373572,
+ "relative_l2_mean": 0.031355825466672094,
+ "relative_l2_p95": 0.11840375444857451,
+ "relative_l2_max": 0.14238893298141875,
+ "weighted_relative_l2_mean": 0.017772480241204428,
+ "weighted_relative_frobenius": 0.00957913732749358,
+ "weighted_mean_cosine": 0.999230502642842,
+ "max_relative_norm_change": 0.010095610896274757
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.333831,
+ "warm_p50_ms": 0.131092,
+ "warm_p95_ms": 0.13349479999999997,
+ "warm_p99_ms": 0.1431488,
+ "samples_ms": [
+ 0.135353,
+ 0.132489,
+ 0.131511,
+ 0.131022,
+ 0.131581,
+ 0.133397,
+ 0.131511,
+ 0.133397,
+ 0.131023,
+ 0.132,
+ 0.129696,
+ 0.131512,
+ 0.130603,
+ 0.130534,
+ 0.129556,
+ 0.130115,
+ 0.131022,
+ 0.131092,
+ 0.131511,
+ 0.131511,
+ 0.131093,
+ 0.130604,
+ 0.131023,
+ 0.131511,
+ 0.131581,
+ 0.130534,
+ 0.148133,
+ 0.131022,
+ 0.131512,
+ 0.129695,
+ 0.129695,
+ 0.131092,
+ 0.130603,
+ 0.131511,
+ 0.131512,
+ 0.132,
+ 0.131092,
+ 0.131022,
+ 0.129556,
+ 0.130045
+ ],
+ "warm_rows_per_second": 3905654.0444878405
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1069332,
+ "unique_rows": 7792,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 126,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 126,
+ "zlib_blocks": 5,
+ "xor_blocks": 121,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7792
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "d62cc5cf74e9c9ccfd174102ce2f797db91c447b8042585a4516071187695b7b",
+ "table_sha256": "f80324fd8e95234685d7aa8ebd715103196c78f0d609974442d47c6578c302a9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13925827806815505,
+ "representatives_sha256": "40292b9de8f011ee9a8b8e640083fd1644dfcbe48e80d8ab318bede6a89016e1",
+ "structurally_merged_rows": 5632,
+ "backend_artifact_bytes": 1070581,
+ "backend_with_key_registry_bytes": 1274529,
+ "backend_with_datrie_key_index_bytes": 1179173,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.980040324320278,
+ "mean_cosine": 0.9980752713199419,
+ "p05_cosine": 0.99092569986586,
+ "relative_l2_mean": 0.040799656727808845,
+ "relative_l2_p95": 0.13498458310318215,
+ "relative_l2_max": 0.19976252785204607,
+ "weighted_relative_l2_mean": 0.0254422516054999,
+ "weighted_relative_frobenius": 0.012937681127503741,
+ "weighted_mean_cosine": 0.9989719276880858,
+ "max_relative_norm_change": 0.012096284539880018
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.980047817712512,
+ "mean_cosine": 0.9980789571770594,
+ "p05_cosine": 0.9909136406377175,
+ "relative_l2_mean": 0.037184530421651284,
+ "relative_l2_p95": 0.13508435168474658,
+ "relative_l2_max": 0.19974615358554437,
+ "weighted_relative_l2_mean": 0.020733881045874192,
+ "weighted_relative_frobenius": 0.011336531353083762,
+ "weighted_mean_cosine": 0.9989822505369915,
+ "max_relative_norm_change": 0.010097003474836263
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.865894,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.1367954,
+ "warm_p99_ms": 0.15384678999999998,
+ "samples_ms": [
+ 0.137658,
+ 0.132978,
+ 0.132,
+ 0.132419,
+ 0.132489,
+ 0.132489,
+ 0.132,
+ 0.13675,
+ 0.132001,
+ 0.131511,
+ 0.132,
+ 0.131022,
+ 0.133397,
+ 0.132001,
+ 0.131511,
+ 0.132489,
+ 0.132,
+ 0.134864,
+ 0.132419,
+ 0.132489,
+ 0.131023,
+ 0.130114,
+ 0.135772,
+ 0.130045,
+ 0.131511,
+ 0.131022,
+ 0.130603,
+ 0.130115,
+ 0.129626,
+ 0.164197,
+ 0.133956,
+ 0.131023,
+ 0.131512,
+ 0.131023,
+ 0.131511,
+ 0.132,
+ 0.130114,
+ 0.129625,
+ 0.130603,
+ 0.129696
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1062058,
+ "unique_rows": 7730,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 125,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 125,
+ "zlib_blocks": 6,
+ "xor_blocks": 119,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7730
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "92ab7de0cff1ab49c6ea19a747212090b80dcecab1ad6f1f938946ee156a61a0",
+ "table_sha256": "482ddd7ed4605d56276772585d97c49ab8bd69597c2b341199ad38d55afc0ee1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1396121650468558,
+ "representatives_sha256": "a557e466cb97ea8c833dc4804e7e0d3aca9eba6052a9c3fc074c3ee0f0407a67",
+ "structurally_merged_rows": 5694,
+ "backend_artifact_bytes": 1063307,
+ "backend_with_key_registry_bytes": 1267255,
+ "backend_with_datrie_key_index_bytes": 1171899,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9710328465263668,
+ "mean_cosine": 0.997964249255204,
+ "p05_cosine": 0.9904656762202363,
+ "relative_l2_mean": 0.04178345412339522,
+ "relative_l2_p95": 0.13827300739605883,
+ "relative_l2_max": 0.24130476390305955,
+ "weighted_relative_l2_mean": 0.025917083383386188,
+ "weighted_relative_frobenius": 0.012971974081592072,
+ "weighted_mean_cosine": 0.9989183426667994,
+ "max_relative_norm_change": 0.012096284539880018
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9708969263450682,
+ "mean_cosine": 0.9979676724494986,
+ "p05_cosine": 0.9904378721927583,
+ "relative_l2_mean": 0.03819809588165435,
+ "relative_l2_p95": 0.13873252170292694,
+ "relative_l2_max": 0.24185314366647992,
+ "weighted_relative_l2_mean": 0.021223080439505985,
+ "weighted_relative_frobenius": 0.011375745984991156,
+ "weighted_mean_cosine": 0.9989285387409689,
+ "max_relative_norm_change": 0.010097003474836263
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.825735,
+ "warm_p50_ms": 0.13130199999999997,
+ "warm_p95_ms": 0.14043964999999997,
+ "warm_p99_ms": 0.16221437,
+ "samples_ms": [
+ 0.156165,
+ 0.132978,
+ 0.132419,
+ 0.131022,
+ 0.130533,
+ 0.131512,
+ 0.135283,
+ 0.139612,
+ 0.131092,
+ 0.132,
+ 0.130114,
+ 0.131511,
+ 0.132419,
+ 0.129207,
+ 0.130114,
+ 0.130114,
+ 0.131581,
+ 0.131022,
+ 0.130044,
+ 0.132489,
+ 0.131512,
+ 0.131023,
+ 0.130115,
+ 0.130534,
+ 0.131511,
+ 0.131511,
+ 0.131022,
+ 0.130603,
+ 0.130533,
+ 0.166082,
+ 0.135772,
+ 0.132489,
+ 0.132489,
+ 0.131092,
+ 0.132489,
+ 0.131093,
+ 0.131092,
+ 0.135353,
+ 0.128718,
+ 0.128717
+ ],
+ "warm_rows_per_second": 3899407.472848853
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1060585,
+ "unique_rows": 7717,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 125,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 125,
+ "zlib_blocks": 3,
+ "xor_blocks": 122,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7717
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "200c592a890fd684e07b59c2c9a0be13bb0f9807ab7a1905f4e9ced4dd7086fa",
+ "table_sha256": "059214151aadee08a67e05a4ad3d998b746e1c7974053c942998857019117e0a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13991869892925024,
+ "representatives_sha256": "16b9a587b67664f168f9a062331119e5f8260ff15708ccef58ddafbab9fc0666",
+ "structurally_merged_rows": 5707,
+ "backend_artifact_bytes": 1061834,
+ "backend_with_key_registry_bytes": 1265782,
+ "backend_with_datrie_key_index_bytes": 1170426,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9540329222020274,
+ "mean_cosine": 0.9978945228003459,
+ "p05_cosine": 0.9902540884725967,
+ "relative_l2_mean": 0.042250876563899045,
+ "relative_l2_p95": 0.13958307447372006,
+ "relative_l2_max": 0.3026913308363412,
+ "weighted_relative_l2_mean": 0.026152889877607343,
+ "weighted_relative_frobenius": 0.013574974339733748,
+ "weighted_mean_cosine": 0.9988831872819481,
+ "max_relative_norm_change": 0.012096284539880018
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9542004255230913,
+ "mean_cosine": 0.9978979280638607,
+ "p05_cosine": 0.990223481380006,
+ "relative_l2_mean": 0.03867101870432088,
+ "relative_l2_p95": 0.1398390536487522,
+ "relative_l2_max": 0.3021072149731349,
+ "weighted_relative_l2_mean": 0.021461805067035492,
+ "weighted_relative_frobenius": 0.012059187754506493,
+ "weighted_mean_cosine": 0.9988933665180764,
+ "max_relative_norm_change": 0.010097003474836263
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.945303,
+ "warm_p50_ms": 0.1327335,
+ "warm_p95_ms": 0.15203059999999996,
+ "warm_p99_ms": 0.17440129999999998,
+ "samples_ms": [
+ 0.163708,
+ 0.137238,
+ 0.134375,
+ 0.132978,
+ 0.131511,
+ 0.133397,
+ 0.133397,
+ 0.132,
+ 0.131511,
+ 0.132419,
+ 0.131512,
+ 0.130604,
+ 0.130185,
+ 0.131512,
+ 0.134375,
+ 0.133886,
+ 0.130534,
+ 0.131511,
+ 0.131022,
+ 0.132,
+ 0.130045,
+ 0.130114,
+ 0.13626,
+ 0.131022,
+ 0.131022,
+ 0.129556,
+ 0.135212,
+ 0.133467,
+ 0.13668,
+ 0.181238,
+ 0.151416,
+ 0.135283,
+ 0.133956,
+ 0.132489,
+ 0.132978,
+ 0.131511,
+ 0.135352,
+ 0.134794,
+ 0.133467,
+ 0.132
+ ],
+ "warm_rows_per_second": 3857353.26801448
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1060362,
+ "unique_rows": 7715,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 125,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 125,
+ "zlib_blocks": 2,
+ "xor_blocks": 123,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7715
+ },
+ "source_float32_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477",
+ "decoded_float32_sha256": "2aa75a12dbb5de141c033222ba01e64cbe02620fadea61b0c4eadfa20ca5d643",
+ "table_sha256": "8aa52b52fa923d903a16b186271dd820d75606a7d15a775bdea4630567d2d584",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13519413978792727,
+ "representatives_sha256": "211862e944b12fc74dd3ca1ef04f19d8d4160da3fcd52ea80b8f46235282d3df",
+ "structurally_merged_rows": 5709,
+ "backend_artifact_bytes": 1061611,
+ "backend_with_key_registry_bytes": 1265559,
+ "backend_with_datrie_key_index_bytes": 1170203,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9006502866041701,
+ "mean_cosine": 0.9978134194705304,
+ "p05_cosine": 0.9901880738674116,
+ "relative_l2_mean": 0.04262239373268259,
+ "relative_l2_p95": 0.13999205434236361,
+ "relative_l2_max": 0.44807027362231283,
+ "weighted_relative_l2_mean": 0.026471557711032132,
+ "weighted_relative_frobenius": 0.015273659929497222,
+ "weighted_mean_cosine": 0.9988181324204763,
+ "max_relative_norm_change": 0.012096284539880018
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9002741030671103,
+ "mean_cosine": 0.9978167763871402,
+ "p05_cosine": 0.9901294460632106,
+ "relative_l2_mean": 0.03904303577942174,
+ "relative_l2_p95": 0.14046688671681148,
+ "relative_l2_max": 0.4488711689341506,
+ "weighted_relative_l2_mean": 0.021783149111407735,
+ "weighted_relative_frobenius": 0.01394742787711847,
+ "weighted_mean_cosine": 0.9988282245388411,
+ "max_relative_norm_change": 0.010097003474836263
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.930706,
+ "warm_p50_ms": 0.131581,
+ "warm_p95_ms": 0.13542614999999997,
+ "warm_p99_ms": 0.14474985999999998,
+ "samples_ms": [
+ 0.135283,
+ 0.133397,
+ 0.132978,
+ 0.132419,
+ 0.131092,
+ 0.132978,
+ 0.132,
+ 0.132,
+ 0.131511,
+ 0.130534,
+ 0.13249,
+ 0.131581,
+ 0.134375,
+ 0.148972,
+ 0.132978,
+ 0.130533,
+ 0.129626,
+ 0.131093,
+ 0.130114,
+ 0.131022,
+ 0.130044,
+ 0.131581,
+ 0.131511,
+ 0.131093,
+ 0.130604,
+ 0.129137,
+ 0.131581,
+ 0.131511,
+ 0.132489,
+ 0.131092,
+ 0.130604,
+ 0.131093,
+ 0.132978,
+ 0.131511,
+ 0.129695,
+ 0.131581,
+ 0.132489,
+ 0.132489,
+ 0.138146,
+ 0.132977
+ ],
+ "warm_rows_per_second": 3891139.298226948
+ }
+ }
+ ],
+ "baseline_quality": {},
+ "training_row_sha256": "00284f971f805571271cde5339ddc057c00eb8e6117ce037be95c658a9f03477"
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 11,
+ "rows": 13424,
+ "dim": 128,
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "source_float32_bytes": 6873088,
+ "key_registry_bytes": 203948,
+ "unchanged_backbone_artifact_bytes": 0,
+ "alternative_datrie_key_index_bytes": 108592,
+ "query_rows": 512,
+ "query_ids_sha256": "80ab83de1ecc982405671e980d9e9a66070e51aead43121126faabb6b44a50cd",
+ "frequency_trace_sha256": "1a0389197253d049e85973cd9812529ea0dc3eae69dcdb62085a34b49c0d90b8",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 512,
+ "storage": {
+ "artifact_bytes": 6873088,
+ "unique_rows": 13424,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "table_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.037072138860821724,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6873893,
+ "backend_with_key_registry_bytes": 7077841,
+ "backend_with_datrie_key_index_bytes": 6982485,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.386572,
+ "warm_p50_ms": 0.06289249999999999,
+ "warm_p95_ms": 0.06544169999999999,
+ "warm_p99_ms": 0.06813809,
+ "samples_ms": [
+ 0.065302,
+ 0.063905,
+ 0.061949,
+ 0.061949,
+ 0.062857,
+ 0.063346,
+ 0.068096,
+ 0.062997,
+ 0.062927,
+ 0.062928,
+ 0.063346,
+ 0.063416,
+ 0.062858,
+ 0.062368,
+ 0.062857,
+ 0.063416,
+ 0.06153,
+ 0.062997,
+ 0.061041,
+ 0.062368,
+ 0.062928,
+ 0.068165,
+ 0.063486,
+ 0.062508,
+ 0.063416,
+ 0.062438,
+ 0.06146,
+ 0.063346,
+ 0.061112,
+ 0.063416,
+ 0.061041,
+ 0.060971,
+ 0.061111,
+ 0.061041,
+ 0.062997,
+ 0.063416,
+ 0.062019,
+ 0.063276,
+ 0.062858,
+ 0.06146
+ ],
+ "warm_rows_per_second": 8140875.303096555
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 6483016,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 839,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 839,
+ "zlib_blocks": 49,
+ "xor_blocks": 790,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "table_sha256": "155a7eeaf0a3e1ce31e77710c5f910993a527b89c27a8e96445f611cb49c6061",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5773331308737397,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6484225,
+ "backend_with_key_registry_bytes": 6688173,
+ "backend_with_datrie_key_index_bytes": 6592817,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.384338,
+ "warm_p50_ms": 0.071937,
+ "warm_p95_ms": 0.07710485,
+ "warm_p99_ms": 0.08512064999999999,
+ "samples_ms": [
+ 0.078432,
+ 0.077035,
+ 0.074242,
+ 0.076127,
+ 0.07473,
+ 0.075219,
+ 0.07522,
+ 0.075219,
+ 0.071937,
+ 0.072356,
+ 0.071937,
+ 0.072425,
+ 0.07054,
+ 0.071378,
+ 0.072355,
+ 0.072355,
+ 0.071867,
+ 0.071937,
+ 0.071517,
+ 0.071378,
+ 0.089397,
+ 0.071447,
+ 0.071518,
+ 0.071448,
+ 0.072844,
+ 0.071028,
+ 0.071448,
+ 0.070889,
+ 0.071867,
+ 0.070889,
+ 0.071377,
+ 0.072845,
+ 0.071867,
+ 0.072286,
+ 0.070959,
+ 0.072356,
+ 0.071867,
+ 0.072356,
+ 0.072355,
+ 0.071867
+ ],
+ "warm_rows_per_second": 7117338.7825458385
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 6483016,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 839,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 839,
+ "zlib_blocks": 49,
+ "xor_blocks": 790,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "table_sha256": "155a7eeaf0a3e1ce31e77710c5f910993a527b89c27a8e96445f611cb49c6061",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.633567641954869,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6484224,
+ "backend_with_key_registry_bytes": 6688172,
+ "backend_with_datrie_key_index_bytes": 6592816,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.382941,
+ "warm_p50_ms": 0.072426,
+ "warm_p95_ms": 0.07584769999999999,
+ "warm_p99_ms": 0.07965185,
+ "samples_ms": [
+ 0.080387,
+ 0.078502,
+ 0.075708,
+ 0.0748,
+ 0.073334,
+ 0.073334,
+ 0.073822,
+ 0.073822,
+ 0.073752,
+ 0.073333,
+ 0.072845,
+ 0.071867,
+ 0.074242,
+ 0.071448,
+ 0.072425,
+ 0.072355,
+ 0.071518,
+ 0.072915,
+ 0.072426,
+ 0.071517,
+ 0.072006,
+ 0.073334,
+ 0.072356,
+ 0.071937,
+ 0.071936,
+ 0.074241,
+ 0.071866,
+ 0.073403,
+ 0.072007,
+ 0.071029,
+ 0.072006,
+ 0.071518,
+ 0.072426,
+ 0.071518,
+ 0.071517,
+ 0.071937,
+ 0.073823,
+ 0.071448,
+ 0.074241,
+ 0.072844
+ ],
+ "warm_rows_per_second": 7069284.511087178
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 6483016,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 839,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 839,
+ "zlib_blocks": 49,
+ "xor_blocks": 790,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "table_sha256": "155a7eeaf0a3e1ce31e77710c5f910993a527b89c27a8e96445f611cb49c6061",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5860995480325073,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6484224,
+ "backend_with_key_registry_bytes": 6688172,
+ "backend_with_datrie_key_index_bytes": 6592816,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.446357,
+ "warm_p50_ms": 0.072146,
+ "warm_p95_ms": 0.0762005,
+ "warm_p99_ms": 0.09407533999999998,
+ "samples_ms": [
+ 0.076127,
+ 0.075708,
+ 0.075219,
+ 0.076267,
+ 0.076127,
+ 0.073752,
+ 0.073892,
+ 0.076197,
+ 0.075708,
+ 0.074241,
+ 0.075219,
+ 0.073822,
+ 0.074311,
+ 0.073822,
+ 0.071028,
+ 0.073333,
+ 0.072355,
+ 0.073334,
+ 0.071937,
+ 0.071867,
+ 0.071447,
+ 0.070959,
+ 0.071378,
+ 0.071378,
+ 0.071867,
+ 0.105461,
+ 0.075289,
+ 0.071029,
+ 0.0704,
+ 0.069981,
+ 0.07047,
+ 0.072356,
+ 0.071937,
+ 0.070958,
+ 0.071448,
+ 0.071448,
+ 0.071447,
+ 0.071518,
+ 0.071447,
+ 0.071936
+ ],
+ "warm_rows_per_second": 7096720.538907216
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 4100169,
+ "unique_rows": 8312,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 520,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 520,
+ "zlib_blocks": 366,
+ "xor_blocks": 154,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8312
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "c4a16f41fcd1cfbf9ebcd3c7266344da7538a4ea7d782e8a916adae61b1781a8",
+ "table_sha256": "0dfd077b4a4269f591f06ad12cd2b429d287f401b93893aa11816fdce8750a62",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.38464223803021014,
+ "representatives_sha256": "18ab6ecdeda144e906690d1f311cd7a0dd27c6372b04725bcf51dc68e687c52a",
+ "structurally_merged_rows": 5112,
+ "backend_artifact_bytes": 4101376,
+ "backend_with_key_registry_bytes": 4305324,
+ "backend_with_datrie_key_index_bytes": 4209968,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900007999373839,
+ "mean_cosine": 0.9985520656727839,
+ "p05_cosine": 0.9927879363822772,
+ "relative_l2_mean": 0.03130694531921343,
+ "relative_l2_p95": 0.12030419446706714,
+ "relative_l2_max": 0.14199628030247896,
+ "weighted_relative_l2_mean": 0.017763284845687673,
+ "weighted_relative_frobenius": 0.009517378602255747,
+ "weighted_mean_cosine": 0.99922030683484,
+ "max_relative_norm_change": 0.010085036488932409
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900007999373839,
+ "mean_cosine": 0.9985520656727839,
+ "p05_cosine": 0.9927879363822772,
+ "relative_l2_mean": 0.03130694531921343,
+ "relative_l2_p95": 0.12030419446706714,
+ "relative_l2_max": 0.14199628030247896,
+ "weighted_relative_l2_mean": 0.017763284845687673,
+ "weighted_relative_frobenius": 0.009517378602255747,
+ "weighted_mean_cosine": 0.99922030683484,
+ "max_relative_norm_change": 0.010085036488932409
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.42459,
+ "warm_p50_ms": 0.0710285,
+ "warm_p95_ms": 0.07528979999999999,
+ "warm_p99_ms": 0.09046238999999999,
+ "samples_ms": [
+ 0.076616,
+ 0.07522,
+ 0.074242,
+ 0.073334,
+ 0.069143,
+ 0.07047,
+ 0.072426,
+ 0.071028,
+ 0.071867,
+ 0.070959,
+ 0.071378,
+ 0.07047,
+ 0.07005,
+ 0.069632,
+ 0.071936,
+ 0.071518,
+ 0.07054,
+ 0.070889,
+ 0.07047,
+ 0.070958,
+ 0.07047,
+ 0.073822,
+ 0.072844,
+ 0.071866,
+ 0.07047,
+ 0.07054,
+ 0.070121,
+ 0.069561,
+ 0.069911,
+ 0.069981,
+ 0.099315,
+ 0.071378,
+ 0.069492,
+ 0.071029,
+ 0.072425,
+ 0.071867,
+ 0.071518,
+ 0.071937,
+ 0.073333,
+ 0.070051
+ ],
+ "warm_rows_per_second": 7208374.10335288
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3867621,
+ "unique_rows": 7825,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 490,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 489,
+ "zlib_blocks": 381,
+ "xor_blocks": 108,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7825
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "600628340ac03ca044a2356eeb0939c2e1e2e8e18a03812c73b9ee26160903ac",
+ "table_sha256": "de8f481b515cd6aa80d1cf98f94d892281aa416f689c956e4b4bc6effd266071",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3573845538776368,
+ "representatives_sha256": "271a9a65c6aaaa61087b3113578bbd5b1bad294312ebcea64e5ebf5639335631",
+ "structurally_merged_rows": 5599,
+ "backend_artifact_bytes": 3868828,
+ "backend_with_key_registry_bytes": 4072776,
+ "backend_with_datrie_key_index_bytes": 3977420,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.980013420759822,
+ "mean_cosine": 0.9980376825820849,
+ "p05_cosine": 0.9905574866144373,
+ "relative_l2_mean": 0.03740989477480981,
+ "relative_l2_p95": 0.1374463580229017,
+ "relative_l2_max": 0.2005315757131508,
+ "weighted_relative_l2_mean": 0.020754561070420903,
+ "weighted_relative_frobenius": 0.009974038075667295,
+ "weighted_mean_cosine": 0.9989679669083437,
+ "max_relative_norm_change": 0.010095772471192354
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.980013420759822,
+ "mean_cosine": 0.9980376825820849,
+ "p05_cosine": 0.9905574866144373,
+ "relative_l2_mean": 0.03740989477480981,
+ "relative_l2_p95": 0.1374463580229017,
+ "relative_l2_max": 0.2005315757131508,
+ "weighted_relative_l2_mean": 0.020754561070420903,
+ "weighted_relative_frobenius": 0.009974038075667295,
+ "weighted_mean_cosine": 0.9989679669083437,
+ "max_relative_norm_change": 0.010095772471192354
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.673505,
+ "warm_p50_ms": 0.07190150000000001,
+ "warm_p95_ms": 0.079081,
+ "warm_p99_ms": 0.09818259999999998,
+ "samples_ms": [
+ 0.077105,
+ 0.073752,
+ 0.07906,
+ 0.076267,
+ 0.073823,
+ 0.073823,
+ 0.071448,
+ 0.071447,
+ 0.071867,
+ 0.073823,
+ 0.074311,
+ 0.073822,
+ 0.073333,
+ 0.071448,
+ 0.070959,
+ 0.071936,
+ 0.069981,
+ 0.070051,
+ 0.070959,
+ 0.070959,
+ 0.07047,
+ 0.071378,
+ 0.069981,
+ 0.071029,
+ 0.071028,
+ 0.073334,
+ 0.073822,
+ 0.072356,
+ 0.11014,
+ 0.07948,
+ 0.071937,
+ 0.070469,
+ 0.070121,
+ 0.072495,
+ 0.072844,
+ 0.071937,
+ 0.071448,
+ 0.071029,
+ 0.069562,
+ 0.07047
+ ],
+ "warm_rows_per_second": 7120852.833390123
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3831029,
+ "unique_rows": 7749,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 485,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 485,
+ "zlib_blocks": 390,
+ "xor_blocks": 95,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7749
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "c105950686a590d7be61e232a22b493cf8e565d079b82606451c2b81261a94e4",
+ "table_sha256": "2caa89e573b6c7149199b782790adfbdf00a028785c9e0f0c809f10e7db1ec57",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.35861606686376035,
+ "representatives_sha256": "91ff7c6d1164808f5722c3a96b485407dddc8d89a3b412521267c4a84fe61875",
+ "structurally_merged_rows": 5675,
+ "backend_artifact_bytes": 3832235,
+ "backend_with_key_registry_bytes": 4036183,
+ "backend_with_datrie_key_index_bytes": 3940827,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9700304236899572,
+ "mean_cosine": 0.9978949631529862,
+ "p05_cosine": 0.990118823150752,
+ "relative_l2_mean": 0.038680935063680276,
+ "relative_l2_p95": 0.1406791364692246,
+ "relative_l2_max": 0.24400340561278033,
+ "weighted_relative_l2_mean": 0.02137579350598353,
+ "weighted_relative_frobenius": 0.010029759713921182,
+ "weighted_mean_cosine": 0.9988982596965421,
+ "max_relative_norm_change": 0.010095772471192354
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9700304236899572,
+ "mean_cosine": 0.9978949631529862,
+ "p05_cosine": 0.990118823150752,
+ "relative_l2_mean": 0.038680935063680276,
+ "relative_l2_p95": 0.1406791364692246,
+ "relative_l2_max": 0.24400340561278033,
+ "weighted_relative_l2_mean": 0.02137579350598353,
+ "weighted_relative_frobenius": 0.010029759713921182,
+ "weighted_mean_cosine": 0.9988982596965421,
+ "max_relative_norm_change": 0.010095772471192354
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.488564,
+ "warm_p50_ms": 0.07047,
+ "warm_p95_ms": 0.07344925,
+ "warm_p99_ms": 0.08952686999999998,
+ "samples_ms": [
+ 0.075639,
+ 0.072845,
+ 0.071867,
+ 0.072914,
+ 0.071378,
+ 0.072915,
+ 0.073334,
+ 0.072356,
+ 0.072356,
+ 0.072425,
+ 0.071867,
+ 0.070959,
+ 0.098406,
+ 0.07047,
+ 0.07047,
+ 0.068584,
+ 0.069631,
+ 0.070051,
+ 0.069493,
+ 0.068654,
+ 0.069073,
+ 0.068514,
+ 0.070539,
+ 0.070051,
+ 0.069981,
+ 0.068584,
+ 0.068654,
+ 0.067606,
+ 0.070051,
+ 0.068654,
+ 0.068584,
+ 0.073333,
+ 0.071867,
+ 0.070959,
+ 0.07054,
+ 0.07054,
+ 0.070051,
+ 0.069981,
+ 0.069492,
+ 0.069631
+ ],
+ "warm_rows_per_second": 7265503.050943663
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3826245,
+ "unique_rows": 7739,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 484,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 484,
+ "zlib_blocks": 380,
+ "xor_blocks": 104,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7739
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "b797fb274cf6bb97c40b3d056b8f4d9ffb1c804a020fa1bea81a8646f81fef31",
+ "table_sha256": "019ffd8ea9f10628cf8512c123802dd2e7c6719a4633964c982b5a9993c51fd5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3633699589408934,
+ "representatives_sha256": "8de86cd8bad44752d2a480c6a19c4591783910fde579a1fc75a12ab6053a1d99",
+ "structurally_merged_rows": 5685,
+ "backend_artifact_bytes": 3827452,
+ "backend_with_key_registry_bytes": 4031400,
+ "backend_with_datrie_key_index_bytes": 3936044,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9554024545296951,
+ "mean_cosine": 0.9978414687987155,
+ "p05_cosine": 0.9900514261641457,
+ "relative_l2_mean": 0.039044707758597454,
+ "relative_l2_p95": 0.14131631695138144,
+ "relative_l2_max": 0.2996513239956172,
+ "weighted_relative_l2_mean": 0.021809966406148944,
+ "weighted_relative_frobenius": 0.011785431782800239,
+ "weighted_mean_cosine": 0.9988329443899867,
+ "max_relative_norm_change": 0.010095772471192354
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9554024545296951,
+ "mean_cosine": 0.9978414687987155,
+ "p05_cosine": 0.9900514261641457,
+ "relative_l2_mean": 0.039044707758597454,
+ "relative_l2_p95": 0.14131631695138144,
+ "relative_l2_max": 0.2996513239956172,
+ "weighted_relative_l2_mean": 0.021809966406148944,
+ "weighted_relative_frobenius": 0.011785431782800239,
+ "weighted_mean_cosine": 0.9988329443899867,
+ "max_relative_norm_change": 0.010095772471192354
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.746838,
+ "warm_p50_ms": 0.070051,
+ "warm_p95_ms": 0.077105,
+ "warm_p99_ms": 0.07829816,
+ "samples_ms": [
+ 0.079061,
+ 0.077105,
+ 0.077105,
+ 0.07473,
+ 0.074241,
+ 0.073822,
+ 0.073892,
+ 0.072425,
+ 0.069981,
+ 0.07012,
+ 0.069981,
+ 0.07047,
+ 0.069562,
+ 0.070051,
+ 0.069982,
+ 0.069911,
+ 0.07047,
+ 0.069562,
+ 0.07054,
+ 0.069981,
+ 0.069981,
+ 0.069562,
+ 0.070051,
+ 0.070051,
+ 0.069911,
+ 0.070051,
+ 0.069632,
+ 0.070051,
+ 0.069492,
+ 0.0704,
+ 0.069493,
+ 0.070121,
+ 0.069981,
+ 0.069492,
+ 0.070051,
+ 0.07005,
+ 0.069981,
+ 0.069492,
+ 0.07047,
+ 0.070121
+ ],
+ "warm_rows_per_second": 7308960.614409502
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 6873088,
+ "artifact_bytes": 3818870,
+ "unique_rows": 7724,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 483,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 483,
+ "zlib_blocks": 375,
+ "xor_blocks": 108,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7724
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "e77e143a51f97e5caed5f01557dcf297c0531b10cde88113dd8211d7f26a9943",
+ "table_sha256": "a2f43e4ff393c1da070ca13dabe4e22ee3ca4ac1b52af5014e2c20b5c28f1ddf",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.39088514586910605,
+ "representatives_sha256": "4d3e104b8903632adcc1ede238ca5305d344b0dfbebf0ee95a8e0b4b9c962cb4",
+ "structurally_merged_rows": 5700,
+ "backend_artifact_bytes": 3820077,
+ "backend_with_key_registry_bytes": 4024025,
+ "backend_with_datrie_key_index_bytes": 3928669,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9044079024947834,
+ "mean_cosine": 0.997713361814159,
+ "p05_cosine": 0.9898971286352017,
+ "relative_l2_mean": 0.03969745170571765,
+ "relative_l2_p95": 0.14202274798903222,
+ "relative_l2_max": 0.4376969589349435,
+ "weighted_relative_l2_mean": 0.022414481830018047,
+ "weighted_relative_frobenius": 0.014658876756557073,
+ "weighted_mean_cosine": 0.9987163314265392,
+ "max_relative_norm_change": 0.010095772471192354
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9044079024947834,
+ "mean_cosine": 0.997713361814159,
+ "p05_cosine": 0.9898971286352017,
+ "relative_l2_mean": 0.03969745170571765,
+ "relative_l2_p95": 0.14202274798903222,
+ "relative_l2_max": 0.4376969589349435,
+ "weighted_relative_l2_mean": 0.022414481830018047,
+ "weighted_relative_frobenius": 0.014658876756557073,
+ "weighted_mean_cosine": 0.9987163314265392,
+ "max_relative_norm_change": 0.010095772471192354
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.396304,
+ "warm_p50_ms": 0.0707845,
+ "warm_p95_ms": 0.0743564,
+ "warm_p99_ms": 0.07747722,
+ "samples_ms": [
+ 0.078921,
+ 0.075219,
+ 0.073752,
+ 0.072355,
+ 0.071029,
+ 0.072915,
+ 0.071867,
+ 0.071448,
+ 0.071937,
+ 0.069981,
+ 0.072425,
+ 0.071029,
+ 0.069562,
+ 0.071448,
+ 0.071797,
+ 0.071378,
+ 0.069562,
+ 0.069073,
+ 0.069562,
+ 0.069562,
+ 0.069562,
+ 0.069143,
+ 0.07054,
+ 0.069493,
+ 0.069562,
+ 0.069073,
+ 0.069492,
+ 0.069073,
+ 0.069073,
+ 0.069562,
+ 0.069981,
+ 0.074311,
+ 0.074311,
+ 0.072006,
+ 0.071518,
+ 0.071937,
+ 0.071517,
+ 0.069981,
+ 0.07047,
+ 0.069632
+ ],
+ "warm_rows_per_second": 7233221.962435279
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 3436544,
+ "unique_rows": 13424,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "745262c69f54fbc70dbe6d692411472e46846937961b94d94883afac06df3fa3",
+ "table_sha256": "7181664941ce04822304ee6afc0b67f281d64784883f2bd11a965520e8851501",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.043027720879763365,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3437350,
+ "backend_with_key_registry_bytes": 3641298,
+ "backend_with_datrie_key_index_bytes": 3545942,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999612548297,
+ "mean_cosine": 0.9999999790165792,
+ "p05_cosine": 0.9999999730537448,
+ "relative_l2_mean": 0.00020669401864291777,
+ "relative_l2_p95": 0.00023473970210358965,
+ "relative_l2_max": 0.00027860911915048036,
+ "weighted_relative_l2_mean": 0.00020586591458364586,
+ "weighted_relative_frobenius": 0.00020520736570042996,
+ "weighted_mean_cosine": 0.9999999791988329,
+ "max_relative_norm_change": 0.0001331667263129161
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.481696,
+ "warm_p50_ms": 0.220559,
+ "warm_p95_ms": 0.22721464999999993,
+ "warm_p99_ms": 0.26642889,
+ "samples_ms": [
+ 0.221886,
+ 0.221048,
+ 0.220489,
+ 0.220978,
+ 0.221398,
+ 0.220978,
+ 0.220908,
+ 0.221397,
+ 0.25548,
+ 0.223771,
+ 0.220489,
+ 0.220489,
+ 0.220559,
+ 0.220978,
+ 0.219581,
+ 0.220559,
+ 0.220978,
+ 0.219162,
+ 0.219581,
+ 0.219092,
+ 0.220559,
+ 0.22,
+ 0.22049,
+ 0.220978,
+ 0.219511,
+ 0.220978,
+ 0.273429,
+ 0.225727,
+ 0.221956,
+ 0.220908,
+ 0.220489,
+ 0.220489,
+ 0.219022,
+ 0.221397,
+ 0.219023,
+ 0.219512,
+ 0.219581,
+ 0.22007,
+ 0.219092,
+ 0.219022
+ ],
+ "warm_rows_per_second": 2321374.3261440247
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 3249576,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 420,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 420,
+ "zlib_blocks": 0,
+ "xor_blocks": 420,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "745262c69f54fbc70dbe6d692411472e46846937961b94d94883afac06df3fa3",
+ "table_sha256": "a8c6941fc7172d12eb0e5479d247f94e791b50cd69d34b5ed0f9432d026bd57d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.32230694498866796,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3250785,
+ "backend_with_key_registry_bytes": 3454733,
+ "backend_with_datrie_key_index_bytes": 3359377,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999612548297,
+ "mean_cosine": 0.9999999790165792,
+ "p05_cosine": 0.9999999730537448,
+ "relative_l2_mean": 0.00020669401864291777,
+ "relative_l2_p95": 0.00023473970210358965,
+ "relative_l2_max": 0.00027860911915048036,
+ "weighted_relative_l2_mean": 0.00020586591458364586,
+ "weighted_relative_frobenius": 0.00020520736570042996,
+ "weighted_mean_cosine": 0.9999999791988329,
+ "max_relative_norm_change": 0.0001331667263129161
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.125798,
+ "warm_p50_ms": 0.229499,
+ "warm_p95_ms": 0.28088375,
+ "warm_p99_ms": 0.29063838000000003,
+ "samples_ms": [
+ 0.292845,
+ 0.230337,
+ 0.230896,
+ 0.232292,
+ 0.231384,
+ 0.22901,
+ 0.22901,
+ 0.229428,
+ 0.229988,
+ 0.230895,
+ 0.229428,
+ 0.230336,
+ 0.229988,
+ 0.22908,
+ 0.229429,
+ 0.229499,
+ 0.280552,
+ 0.232292,
+ 0.22901,
+ 0.229918,
+ 0.229917,
+ 0.22901,
+ 0.230476,
+ 0.229917,
+ 0.229499,
+ 0.22901,
+ 0.22901,
+ 0.228591,
+ 0.228102,
+ 0.229499,
+ 0.229499,
+ 0.22901,
+ 0.229428,
+ 0.24235,
+ 0.230407,
+ 0.22894,
+ 0.22901,
+ 0.22852,
+ 0.228521,
+ 0.287187
+ ],
+ "warm_rows_per_second": 2230946.54007207
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 3249576,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 420,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 420,
+ "zlib_blocks": 0,
+ "xor_blocks": 420,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "745262c69f54fbc70dbe6d692411472e46846937961b94d94883afac06df3fa3",
+ "table_sha256": "a8c6941fc7172d12eb0e5479d247f94e791b50cd69d34b5ed0f9432d026bd57d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3722030739299953,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3250784,
+ "backend_with_key_registry_bytes": 3454732,
+ "backend_with_datrie_key_index_bytes": 3359376,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999612548297,
+ "mean_cosine": 0.9999999790165792,
+ "p05_cosine": 0.9999999730537448,
+ "relative_l2_mean": 0.00020669401864291777,
+ "relative_l2_p95": 0.00023473970210358965,
+ "relative_l2_max": 0.00027860911915048036,
+ "weighted_relative_l2_mean": 0.00020586591458364586,
+ "weighted_relative_frobenius": 0.00020520736570042996,
+ "weighted_mean_cosine": 0.9999999791988329,
+ "max_relative_norm_change": 0.0001331667263129161
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.138648,
+ "warm_p50_ms": 0.230406,
+ "warm_p95_ms": 0.28818695,
+ "warm_p99_ms": 0.28952196,
+ "samples_ms": [
+ 0.231315,
+ 0.231803,
+ 0.229429,
+ 0.230896,
+ 0.254991,
+ 0.234667,
+ 0.229918,
+ 0.230476,
+ 0.230826,
+ 0.230407,
+ 0.290121,
+ 0.288166,
+ 0.288585,
+ 0.288166,
+ 0.229918,
+ 0.229009,
+ 0.228591,
+ 0.231314,
+ 0.22901,
+ 0.230896,
+ 0.241721,
+ 0.230406,
+ 0.230476,
+ 0.22894,
+ 0.229498,
+ 0.230407,
+ 0.229918,
+ 0.229917,
+ 0.228521,
+ 0.228521,
+ 0.229918,
+ 0.229359,
+ 0.230406,
+ 0.229498,
+ 0.229499,
+ 0.229428,
+ 0.229429,
+ 0.230407,
+ 0.241721,
+ 0.229988
+ ],
+ "warm_rows_per_second": 2222164.3533588536
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 3249576,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 420,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 420,
+ "zlib_blocks": 0,
+ "xor_blocks": 420,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "745262c69f54fbc70dbe6d692411472e46846937961b94d94883afac06df3fa3",
+ "table_sha256": "a8c6941fc7172d12eb0e5479d247f94e791b50cd69d34b5ed0f9432d026bd57d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3346426740754396,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3250784,
+ "backend_with_key_registry_bytes": 3454732,
+ "backend_with_datrie_key_index_bytes": 3359376,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999612548297,
+ "mean_cosine": 0.9999999790165792,
+ "p05_cosine": 0.9999999730537448,
+ "relative_l2_mean": 0.00020669401864291777,
+ "relative_l2_p95": 0.00023473970210358965,
+ "relative_l2_max": 0.00027860911915048036,
+ "weighted_relative_l2_mean": 0.00020586591458364586,
+ "weighted_relative_frobenius": 0.00020520736570042996,
+ "weighted_mean_cosine": 0.9999999791988329,
+ "max_relative_norm_change": 0.0001331667263129161
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.183556,
+ "warm_p50_ms": 0.229499,
+ "warm_p95_ms": 0.2674156499999999,
+ "warm_p99_ms": 0.28808287,
+ "samples_ms": [
+ 0.231315,
+ 0.229429,
+ 0.25548,
+ 0.232711,
+ 0.230337,
+ 0.229917,
+ 0.229429,
+ 0.22901,
+ 0.22901,
+ 0.230896,
+ 0.229918,
+ 0.229917,
+ 0.229359,
+ 0.229848,
+ 0.22901,
+ 0.22908,
+ 0.229079,
+ 0.229988,
+ 0.22859,
+ 0.242699,
+ 0.288655,
+ 0.287188,
+ 0.229429,
+ 0.229429,
+ 0.229499,
+ 0.266375,
+ 0.236134,
+ 0.230825,
+ 0.229429,
+ 0.228102,
+ 0.228032,
+ 0.229498,
+ 0.22894,
+ 0.227613,
+ 0.229918,
+ 0.228521,
+ 0.239905,
+ 0.229498,
+ 0.229499,
+ 0.229499
+ ],
+ "warm_rows_per_second": 2230946.54007207
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 2105028,
+ "unique_rows": 8311,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 260,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 260,
+ "zlib_blocks": 119,
+ "xor_blocks": 141,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8311
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "faca6fff0c110b03e63edd9d50b71c9d1c0c7d2b0742217c219c983386c75403",
+ "table_sha256": "b6745745bf9452dcedf4cc9a07775cec18e50b9734f765b69278759ec885d9aa",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2781077749095857,
+ "representatives_sha256": "0b04f4fca40a83c1d082ca4b51cd2ff6ce3a581e6c4a9cbcca8e95362052745e",
+ "structurally_merged_rows": 5113,
+ "backend_artifact_bytes": 2106235,
+ "backend_with_key_registry_bytes": 2310183,
+ "backend_with_datrie_key_index_bytes": 2214827,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9900077149362516,
+ "mean_cosine": 0.9985521257011059,
+ "p05_cosine": 0.9927844616386878,
+ "relative_l2_mean": 0.031438301030621084,
+ "relative_l2_p95": 0.12031554395198002,
+ "relative_l2_max": 0.14199922735932338,
+ "weighted_relative_l2_mean": 0.01792222597194396,
+ "weighted_relative_frobenius": 0.009520497179207372,
+ "weighted_mean_cosine": 0.9992203252964822,
+ "max_relative_norm_change": 0.010125723311944002
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.990009379823296,
+ "mean_cosine": 0.9985521112039947,
+ "p05_cosine": 0.9927839291169913,
+ "relative_l2_mean": 0.03131066276960376,
+ "relative_l2_p95": 0.12030417355944241,
+ "relative_l2_max": 0.14197100159414283,
+ "weighted_relative_l2_mean": 0.017765175325278264,
+ "weighted_relative_frobenius": 0.009518618838387841,
+ "weighted_mean_cosine": 0.9992203250344446,
+ "max_relative_norm_change": 0.01008796306532617
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.850156,
+ "warm_p50_ms": 0.22859049999999997,
+ "warm_p95_ms": 0.23193654999999996,
+ "warm_p99_ms": 0.24919940000000002,
+ "samples_ms": [
+ 0.231315,
+ 0.229988,
+ 0.229079,
+ 0.228591,
+ 0.229079,
+ 0.228102,
+ 0.22859,
+ 0.228591,
+ 0.228591,
+ 0.229499,
+ 0.228101,
+ 0.228102,
+ 0.228591,
+ 0.252686,
+ 0.230825,
+ 0.22901,
+ 0.227543,
+ 0.229918,
+ 0.227543,
+ 0.230407,
+ 0.229429,
+ 0.229009,
+ 0.228521,
+ 0.228521,
+ 0.227543,
+ 0.227962,
+ 0.227055,
+ 0.228521,
+ 0.228032,
+ 0.227543,
+ 0.243746,
+ 0.230407,
+ 0.227124,
+ 0.227543,
+ 0.22852,
+ 0.229429,
+ 0.229429,
+ 0.226635,
+ 0.227613,
+ 0.228102
+ ],
+ "warm_rows_per_second": 2239813.115593168
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1990400,
+ "unique_rows": 7824,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 245,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 245,
+ "zlib_blocks": 149,
+ "xor_blocks": 96,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7824
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "5df2d348bb831b82bc6213c26b6396afefa01092432d062ce8f75b2d020e60a3",
+ "table_sha256": "87fb565c05c180538a76ae6409591c2958ff83e7ff6d842c533c5585d7498b20",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21020766301080585,
+ "representatives_sha256": "4a440ed6fd489e7e7935c570841a7840ed7290964d5fcf2f4465e22a1e48c597",
+ "structurally_merged_rows": 5600,
+ "backend_artifact_bytes": 1991606,
+ "backend_with_key_registry_bytes": 2195554,
+ "backend_with_datrie_key_index_bytes": 2100198,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9800137839059521,
+ "mean_cosine": 0.9980377116486364,
+ "p05_cosine": 0.9905560009410573,
+ "relative_l2_mean": 0.03753393260000457,
+ "relative_l2_p95": 0.13745740963301878,
+ "relative_l2_max": 0.20008525457545137,
+ "weighted_relative_l2_mean": 0.020909921191685353,
+ "weighted_relative_frobenius": 0.009976854081682007,
+ "weighted_mean_cosine": 0.998967970622337,
+ "max_relative_norm_change": 0.010125723311944002
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9800163940067987,
+ "mean_cosine": 0.998037706994905,
+ "p05_cosine": 0.990556319900482,
+ "relative_l2_mean": 0.037413786078987626,
+ "relative_l2_p95": 0.13743524070760002,
+ "relative_l2_max": 0.2000961318971666,
+ "weighted_relative_l2_mean": 0.020756536232808258,
+ "weighted_relative_frobenius": 0.009975070344183962,
+ "weighted_mean_cosine": 0.998967974737777,
+ "max_relative_norm_change": 0.01008796306532617
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.306931,
+ "warm_p50_ms": 0.22901,
+ "warm_p95_ms": 0.28265199999999996,
+ "warm_p99_ms": 0.28695459,
+ "samples_ms": [
+ 0.232781,
+ 0.265886,
+ 0.232293,
+ 0.229918,
+ 0.22894,
+ 0.229429,
+ 0.229917,
+ 0.228521,
+ 0.22901,
+ 0.22901,
+ 0.22901,
+ 0.228521,
+ 0.226636,
+ 0.228381,
+ 0.228032,
+ 0.228451,
+ 0.228451,
+ 0.22894,
+ 0.241791,
+ 0.231385,
+ 0.228521,
+ 0.228032,
+ 0.22894,
+ 0.233619,
+ 0.22901,
+ 0.22901,
+ 0.229987,
+ 0.286699,
+ 0.287118,
+ 0.233201,
+ 0.227613,
+ 0.227613,
+ 0.227194,
+ 0.228032,
+ 0.228102,
+ 0.229918,
+ 0.242279,
+ 0.282439,
+ 0.240813,
+ 0.230896
+ ],
+ "warm_rows_per_second": 2235710.2309942795
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1972568,
+ "unique_rows": 7748,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 243,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 243,
+ "zlib_blocks": 155,
+ "xor_blocks": 88,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7748
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "18dd991f3b1c8f2173d81198257f2501d390838c64e6578da9e5909796f1e886",
+ "table_sha256": "7c27ff22a9f60768c990701a6ae4aebbd3b3f61b3bc153ce9ca4ab12865aec7b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.23467504396103323,
+ "representatives_sha256": "6028a6047a0ea9de9c4b2b7403abbc402a30d37ff4299deace29d310622f18a2",
+ "structurally_merged_rows": 5676,
+ "backend_artifact_bytes": 1973774,
+ "backend_with_key_registry_bytes": 2177722,
+ "backend_with_datrie_key_index_bytes": 2082366,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9700314079851591,
+ "mean_cosine": 0.9978949938076623,
+ "p05_cosine": 0.9901183891582513,
+ "relative_l2_mean": 0.038803787130537434,
+ "relative_l2_p95": 0.14068049744242087,
+ "relative_l2_max": 0.2439915857988374,
+ "weighted_relative_l2_mean": 0.02153057473686633,
+ "weighted_relative_frobenius": 0.010032559648566038,
+ "weighted_mean_cosine": 0.9988982640629619,
+ "max_relative_norm_change": 0.010125723311944002
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9700309212986293,
+ "mean_cosine": 0.9978949885233197,
+ "p05_cosine": 0.990119777173615,
+ "relative_l2_mean": 0.038684829313671856,
+ "relative_l2_p95": 0.1406630418933251,
+ "relative_l2_max": 0.24399958599152996,
+ "weighted_relative_l2_mean": 0.021377770680033402,
+ "weighted_relative_frobenius": 0.010030786695295522,
+ "weighted_mean_cosine": 0.9988982678307401,
+ "max_relative_norm_change": 0.01008796306532617
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.688823,
+ "warm_p50_ms": 0.2289745,
+ "warm_p95_ms": 0.2428109,
+ "warm_p99_ms": 0.25174378999999997,
+ "samples_ms": [
+ 0.230896,
+ 0.229499,
+ 0.22901,
+ 0.243607,
+ 0.229498,
+ 0.22894,
+ 0.227543,
+ 0.228032,
+ 0.227543,
+ 0.228451,
+ 0.228032,
+ 0.22894,
+ 0.228521,
+ 0.227963,
+ 0.226635,
+ 0.236134,
+ 0.24221,
+ 0.230407,
+ 0.229009,
+ 0.227194,
+ 0.240813,
+ 0.229499,
+ 0.227962,
+ 0.228521,
+ 0.227962,
+ 0.226705,
+ 0.22859,
+ 0.229429,
+ 0.228521,
+ 0.233829,
+ 0.230895,
+ 0.231873,
+ 0.242769,
+ 0.229988,
+ 0.22901,
+ 0.228521,
+ 0.227544,
+ 0.256946,
+ 0.231315,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2236056.8534924197
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1969970,
+ "unique_rows": 7738,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 242,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 242,
+ "zlib_blocks": 151,
+ "xor_blocks": 91,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7738
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "206eb3c30678730cd57a9b65ae4f9c502f6deb02acc5e5ad5cabc6613dadcb55",
+ "table_sha256": "4ed3d64b6c20f54fa506a70a344028c235d62354f6854b3658ba556c8eb15344",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21138525917194784,
+ "representatives_sha256": "a277c4ddcc764481205e42ec0dc39d62704a14287db2966b15c819a1145c7898",
+ "structurally_merged_rows": 5686,
+ "backend_artifact_bytes": 1971176,
+ "backend_with_key_registry_bytes": 2175124,
+ "backend_with_datrie_key_index_bytes": 2079768,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9554112304718978,
+ "mean_cosine": 0.997841501751493,
+ "p05_cosine": 0.9900522530297954,
+ "relative_l2_mean": 0.039167393131070684,
+ "relative_l2_p95": 0.14129570340293618,
+ "relative_l2_max": 0.29962859056383023,
+ "weighted_relative_l2_mean": 0.021964581487790142,
+ "weighted_relative_frobenius": 0.011787515632833112,
+ "weighted_mean_cosine": 0.998832956953083,
+ "max_relative_norm_change": 0.010125723311944002
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9554183376699096,
+ "mean_cosine": 0.9978414956615338,
+ "p05_cosine": 0.9900537656880667,
+ "relative_l2_mean": 0.03904859538817834,
+ "relative_l2_p95": 0.14128987318352873,
+ "relative_l2_max": 0.2996102536492556,
+ "weighted_relative_l2_mean": 0.02181190485627813,
+ "weighted_relative_frobenius": 0.011786082613841144,
+ "weighted_mean_cosine": 0.9988329595639267,
+ "max_relative_norm_change": 0.01008796306532617
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.690639,
+ "warm_p50_ms": 0.22928900000000002,
+ "warm_p95_ms": 0.28870035,
+ "warm_p99_ms": 0.30425995,
+ "samples_ms": [
+ 0.231803,
+ 0.235575,
+ 0.23327,
+ 0.231873,
+ 0.229917,
+ 0.289562,
+ 0.287188,
+ 0.288655,
+ 0.287188,
+ 0.287606,
+ 0.313657,
+ 0.232292,
+ 0.22901,
+ 0.230407,
+ 0.230406,
+ 0.229987,
+ 0.229499,
+ 0.229498,
+ 0.228591,
+ 0.227613,
+ 0.227194,
+ 0.229079,
+ 0.22908,
+ 0.287187,
+ 0.227613,
+ 0.227613,
+ 0.228521,
+ 0.241232,
+ 0.228032,
+ 0.227613,
+ 0.227124,
+ 0.229009,
+ 0.228591,
+ 0.22859,
+ 0.287188,
+ 0.228521,
+ 0.226635,
+ 0.227054,
+ 0.228939,
+ 0.22894
+ ],
+ "warm_rows_per_second": 2232989.807622694
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3436544,
+ "artifact_bytes": 1966232,
+ "unique_rows": 7723,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 242,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 242,
+ "zlib_blocks": 150,
+ "xor_blocks": 92,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7723
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "5a274c0aa20e3c70495c29c4432d3ad33f4fbfa7a7cf7a415418ffb07642bc37",
+ "table_sha256": "923968a09633745982eb0a1e9950eb882e896a9cc32404ebe57aaae23db8421b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2159291368443519,
+ "representatives_sha256": "57b2b6c1ab7ed4773e1775a97b909dbfacb6519463048d213d565efb187c80c9",
+ "structurally_merged_rows": 5701,
+ "backend_artifact_bytes": 1967438,
+ "backend_with_key_registry_bytes": 2171386,
+ "backend_with_datrie_key_index_bytes": 2076030,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.904412617018626,
+ "mean_cosine": 0.9977202774797137,
+ "p05_cosine": 0.9898974661899778,
+ "relative_l2_mean": 0.03979067732544779,
+ "relative_l2_p95": 0.14200537379054382,
+ "relative_l2_max": 0.4376919153864184,
+ "weighted_relative_l2_mean": 0.022554738165779152,
+ "weighted_relative_frobenius": 0.014593319661202971,
+ "weighted_mean_cosine": 0.9987196662683778,
+ "max_relative_norm_change": 0.010125723311944002
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9044076837305283,
+ "mean_cosine": 0.997720269099691,
+ "p05_cosine": 0.9899000698085584,
+ "relative_l2_mean": 0.03967211665200063,
+ "relative_l2_p95": 0.142004485756289,
+ "relative_l2_max": 0.4377167106420038,
+ "weighted_relative_l2_mean": 0.022402312863921987,
+ "weighted_relative_frobenius": 0.014592223061960206,
+ "weighted_mean_cosine": 0.9987196673359467,
+ "max_relative_norm_change": 0.01008796306532617
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.49208,
+ "warm_p50_ms": 0.2285555,
+ "warm_p95_ms": 0.2530909,
+ "warm_p99_ms": 0.26441492,
+ "samples_ms": [
+ 0.231385,
+ 0.229499,
+ 0.229917,
+ 0.22901,
+ 0.227544,
+ 0.228591,
+ 0.229009,
+ 0.227613,
+ 0.227124,
+ 0.252616,
+ 0.232362,
+ 0.228591,
+ 0.22859,
+ 0.228591,
+ 0.227124,
+ 0.265886,
+ 0.234667,
+ 0.22859,
+ 0.228591,
+ 0.228101,
+ 0.227124,
+ 0.227613,
+ 0.228521,
+ 0.227054,
+ 0.227962,
+ 0.228032,
+ 0.241372,
+ 0.229498,
+ 0.227613,
+ 0.227543,
+ 0.227544,
+ 0.227054,
+ 0.227473,
+ 0.227055,
+ 0.227613,
+ 0.226705,
+ 0.231314,
+ 0.262114,
+ 0.231803,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2240156.110878977
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 132,
+ "storage": {
+ "artifact_bytes": 1771968,
+ "unique_rows": 13424,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "ed5fe2646195c36093dae73b4e27ecb710aaf4493c4643375246e406ec0f1dfd",
+ "table_sha256": "145179f5873a93b683607b1475e5ff91182ff4d09cf914774cef2ef32c002ed1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.03684794786386192,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1772818,
+ "backend_with_key_registry_bytes": 1976766,
+ "backend_with_datrie_key_index_bytes": 1881410,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999429184170107,
+ "mean_cosine": 0.9999772888320042,
+ "p05_cosine": 0.9999629801925154,
+ "relative_l2_mean": 0.006690927009116968,
+ "relative_l2_p95": 0.00863155067483264,
+ "relative_l2_max": 0.010732325702084305,
+ "weighted_relative_l2_mean": 0.006846066546862484,
+ "weighted_relative_frobenius": 0.007427315453576613,
+ "weighted_mean_cosine": 0.9999762114726573,
+ "max_relative_norm_change": 0.002874768094226643
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.394604,
+ "warm_p50_ms": 0.122502,
+ "warm_p95_ms": 0.12723345,
+ "warm_p99_ms": 0.13708587,
+ "samples_ms": [
+ 0.124458,
+ 0.122642,
+ 0.122083,
+ 0.12306,
+ 0.122991,
+ 0.121594,
+ 0.124457,
+ 0.12306,
+ 0.121663,
+ 0.122991,
+ 0.122083,
+ 0.121594,
+ 0.122013,
+ 0.137658,
+ 0.122502,
+ 0.122991,
+ 0.122501,
+ 0.122502,
+ 0.122083,
+ 0.122083,
+ 0.122083,
+ 0.122571,
+ 0.126762,
+ 0.122572,
+ 0.121593,
+ 0.122083,
+ 0.122502,
+ 0.122572,
+ 0.122152,
+ 0.12348,
+ 0.121664,
+ 0.122502,
+ 0.122572,
+ 0.136191,
+ 0.123969,
+ 0.122572,
+ 0.120197,
+ 0.121594,
+ 0.121036,
+ 0.122013
+ ],
+ "warm_rows_per_second": 4179523.5996147003
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1644481,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 217,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 217,
+ "zlib_blocks": 0,
+ "xor_blocks": 217,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "ed5fe2646195c36093dae73b4e27ecb710aaf4493c4643375246e406ec0f1dfd",
+ "table_sha256": "976970440e84e9144ffbc9b2e88d574df0db406d5d107b5c4899f9635b081a4e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.19422474992461503,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1645734,
+ "backend_with_key_registry_bytes": 1849682,
+ "backend_with_datrie_key_index_bytes": 1754326,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999429184170107,
+ "mean_cosine": 0.9999772888320042,
+ "p05_cosine": 0.9999629801925154,
+ "relative_l2_mean": 0.006690927009116968,
+ "relative_l2_p95": 0.00863155067483264,
+ "relative_l2_max": 0.010732325702084305,
+ "weighted_relative_l2_mean": 0.006846066546862484,
+ "weighted_relative_frobenius": 0.007427315453576613,
+ "weighted_mean_cosine": 0.9999762114726573,
+ "max_relative_norm_change": 0.002874768094226643
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.106742,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.13870819999999998,
+ "warm_p99_ms": 0.14407029999999998,
+ "samples_ms": [
+ 0.135283,
+ 0.135423,
+ 0.13242,
+ 0.131581,
+ 0.138565,
+ 0.131511,
+ 0.134304,
+ 0.132,
+ 0.131511,
+ 0.132978,
+ 0.131581,
+ 0.131511,
+ 0.131511,
+ 0.132,
+ 0.132,
+ 0.132,
+ 0.131511,
+ 0.141429,
+ 0.138076,
+ 0.132489,
+ 0.131092,
+ 0.132978,
+ 0.130534,
+ 0.132,
+ 0.131092,
+ 0.145759,
+ 0.131512,
+ 0.132001,
+ 0.130534,
+ 0.133397,
+ 0.135841,
+ 0.133886,
+ 0.133467,
+ 0.132,
+ 0.131511,
+ 0.131023,
+ 0.134794,
+ 0.133467,
+ 0.132001,
+ 0.131092
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1644481,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 217,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 217,
+ "zlib_blocks": 0,
+ "xor_blocks": 217,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "ed5fe2646195c36093dae73b4e27ecb710aaf4493c4643375246e406ec0f1dfd",
+ "table_sha256": "976970440e84e9144ffbc9b2e88d574df0db406d5d107b5c4899f9635b081a4e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24700916791334748,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1645733,
+ "backend_with_key_registry_bytes": 1849681,
+ "backend_with_datrie_key_index_bytes": 1754325,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999429184170107,
+ "mean_cosine": 0.9999772888320042,
+ "p05_cosine": 0.9999629801925154,
+ "relative_l2_mean": 0.006690927009116968,
+ "relative_l2_p95": 0.00863155067483264,
+ "relative_l2_max": 0.010732325702084305,
+ "weighted_relative_l2_mean": 0.006846066546862484,
+ "weighted_relative_frobenius": 0.007427315453576613,
+ "weighted_mean_cosine": 0.9999762114726573,
+ "max_relative_norm_change": 0.002874768094226643
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.020209,
+ "warm_p50_ms": 0.1315465,
+ "warm_p95_ms": 0.13810509999999998,
+ "warm_p99_ms": 0.15621889,
+ "samples_ms": [
+ 0.135283,
+ 0.133397,
+ 0.137727,
+ 0.132978,
+ 0.132977,
+ 0.132978,
+ 0.133467,
+ 0.130604,
+ 0.163219,
+ 0.134864,
+ 0.133397,
+ 0.131022,
+ 0.131022,
+ 0.131022,
+ 0.130603,
+ 0.131511,
+ 0.132489,
+ 0.131023,
+ 0.131511,
+ 0.132489,
+ 0.132,
+ 0.131022,
+ 0.131511,
+ 0.131023,
+ 0.137728,
+ 0.131093,
+ 0.131512,
+ 0.132489,
+ 0.131511,
+ 0.132978,
+ 0.132,
+ 0.131511,
+ 0.131581,
+ 0.130534,
+ 0.131511,
+ 0.131022,
+ 0.130603,
+ 0.14527,
+ 0.133397,
+ 0.131511
+ ],
+ "warm_rows_per_second": 3892159.8066083095
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1644481,
+ "unique_rows": 13424,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 217,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 217,
+ "zlib_blocks": 0,
+ "xor_blocks": 217,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 13424
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "ed5fe2646195c36093dae73b4e27ecb710aaf4493c4643375246e406ec0f1dfd",
+ "table_sha256": "976970440e84e9144ffbc9b2e88d574df0db406d5d107b5c4899f9635b081a4e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1931517068296671,
+ "representatives_sha256": "35888549e73e40bba1ad2e09fd530343fb49575b458c9b6a5a8768f1715dc97d",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1645733,
+ "backend_with_key_registry_bytes": 1849681,
+ "backend_with_datrie_key_index_bytes": 1754325,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999429184170107,
+ "mean_cosine": 0.9999772888320042,
+ "p05_cosine": 0.9999629801925154,
+ "relative_l2_mean": 0.006690927009116968,
+ "relative_l2_p95": 0.00863155067483264,
+ "relative_l2_max": 0.010732325702084305,
+ "weighted_relative_l2_mean": 0.006846066546862484,
+ "weighted_relative_frobenius": 0.007427315453576613,
+ "weighted_mean_cosine": 0.9999762114726573,
+ "max_relative_norm_change": 0.002874768094226643
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.021187,
+ "warm_p50_ms": 0.132489,
+ "warm_p95_ms": 0.13897454999999995,
+ "warm_p99_ms": 0.16134456,
+ "samples_ms": [
+ 0.154698,
+ 0.138146,
+ 0.138147,
+ 0.134375,
+ 0.134933,
+ 0.132908,
+ 0.135771,
+ 0.134864,
+ 0.132489,
+ 0.132978,
+ 0.132,
+ 0.132,
+ 0.132,
+ 0.133467,
+ 0.132,
+ 0.132,
+ 0.132489,
+ 0.131512,
+ 0.131511,
+ 0.132489,
+ 0.131511,
+ 0.132489,
+ 0.132,
+ 0.131511,
+ 0.13633,
+ 0.134305,
+ 0.132489,
+ 0.131511,
+ 0.132489,
+ 0.165594,
+ 0.137728,
+ 0.132489,
+ 0.131022,
+ 0.133466,
+ 0.132489,
+ 0.131092,
+ 0.132909,
+ 0.132489,
+ 0.132419,
+ 0.131582
+ ],
+ "warm_rows_per_second": 3864471.7674674885
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1125117,
+ "unique_rows": 8330,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 135,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 135,
+ "zlib_blocks": 2,
+ "xor_blocks": 133,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8330
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "293b129e64a0605a67cbee86bf34bb8c806922eec637179b615339f130b75f4c",
+ "table_sha256": "7e051a24117fe0bcb7af7955f511dd060d4492a461d42a2dbd060d0f0b8cddbe",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1461592318955809,
+ "representatives_sha256": "ef38bd0185746b6eb3c08beb73f5bf4a7f67137414b83e830db9c591f11444b1",
+ "structurally_merged_rows": 5094,
+ "backend_artifact_bytes": 1126366,
+ "backend_with_key_registry_bytes": 1330314,
+ "backend_with_datrie_key_index_bytes": 1234958,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.989883376131355,
+ "mean_cosine": 0.9985366108160204,
+ "p05_cosine": 0.9927796835663705,
+ "relative_l2_mean": 0.03544035891840036,
+ "relative_l2_p95": 0.12031219140066435,
+ "relative_l2_max": 0.14283478277746048,
+ "weighted_relative_l2_mean": 0.02304568387445217,
+ "weighted_relative_frobenius": 0.012055645188533535,
+ "weighted_mean_cosine": 0.9992004261084678,
+ "max_relative_norm_change": 0.0111298361935299
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.990003104072174,
+ "mean_cosine": 0.998541844251837,
+ "p05_cosine": 0.9927626218423953,
+ "relative_l2_mean": 0.03141726644040223,
+ "relative_l2_p95": 0.120413714532134,
+ "relative_l2_max": 0.14209593221613123,
+ "weighted_relative_l2_mean": 0.01785312916105629,
+ "weighted_relative_frobenius": 0.00963143524074153,
+ "weighted_mean_cosine": 0.9992135642014484,
+ "max_relative_norm_change": 0.010090093026292535
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.325799,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.1375416499999999,
+ "warm_p99_ms": 0.22353399999999995,
+ "samples_ms": [
+ 0.135841,
+ 0.135282,
+ 0.132978,
+ 0.132,
+ 0.132908,
+ 0.132,
+ 0.169854,
+ 0.132489,
+ 0.131093,
+ 0.132489,
+ 0.132489,
+ 0.131581,
+ 0.130114,
+ 0.131511,
+ 0.131092,
+ 0.131023,
+ 0.131093,
+ 0.130604,
+ 0.131023,
+ 0.131023,
+ 0.132,
+ 0.131511,
+ 0.131511,
+ 0.130114,
+ 0.129206,
+ 0.131581,
+ 0.130533,
+ 0.131023,
+ 0.130604,
+ 0.129695,
+ 0.131092,
+ 0.130603,
+ 0.132,
+ 0.131023,
+ 0.132,
+ 0.257854,
+ 0.135282,
+ 0.131022,
+ 0.130533,
+ 0.131581
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1070533,
+ "unique_rows": 7840,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 127,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 127,
+ "zlib_blocks": 4,
+ "xor_blocks": 123,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7840
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "4cc2cacc4cffc29cd28bfffb6a2063fb80cd75890cf3332cef620e01f9109e23",
+ "table_sha256": "ddada4dbadb5027d32efee0d25463c834ce4d3f5aaa51b44bd467ad2128fa03d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14123639510944486,
+ "representatives_sha256": "a85e9a37532ae7d133b949e4603535ed993b0ecc14edcc8aedca7bc771fe2cc7",
+ "structurally_merged_rows": 5584,
+ "backend_artifact_bytes": 1071782,
+ "backend_with_key_registry_bytes": 1275730,
+ "backend_with_datrie_key_index_bytes": 1180374,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9799888216619596,
+ "mean_cosine": 0.9980203268379337,
+ "p05_cosine": 0.990537724069153,
+ "relative_l2_mean": 0.04134452123811339,
+ "relative_l2_p95": 0.13755965810914922,
+ "relative_l2_max": 0.20010886707252595,
+ "weighted_relative_l2_mean": 0.025939585992463807,
+ "weighted_relative_frobenius": 0.012420949562250233,
+ "weighted_mean_cosine": 0.9989471847191407,
+ "max_relative_norm_change": 0.0111298361935299
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9800645740292964,
+ "mean_cosine": 0.9980241146613112,
+ "p05_cosine": 0.9905119624083044,
+ "relative_l2_mean": 0.03756072155684393,
+ "relative_l2_p95": 0.13782818640777994,
+ "relative_l2_max": 0.19984363334867963,
+ "weighted_relative_l2_mean": 0.02086386226008907,
+ "weighted_relative_frobenius": 0.010086130604784648,
+ "weighted_mean_cosine": 0.9989596191879863,
+ "max_relative_norm_change": 0.010096948723578298
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.875463,
+ "warm_p50_ms": 0.13130199999999997,
+ "warm_p95_ms": 0.1354228,
+ "warm_p99_ms": 0.14160643,
+ "samples_ms": [
+ 0.136749,
+ 0.133955,
+ 0.135353,
+ 0.134863,
+ 0.132978,
+ 0.131023,
+ 0.132,
+ 0.131511,
+ 0.131511,
+ 0.132,
+ 0.131022,
+ 0.131511,
+ 0.132,
+ 0.129207,
+ 0.131092,
+ 0.130045,
+ 0.144712,
+ 0.130533,
+ 0.131581,
+ 0.131092,
+ 0.130603,
+ 0.131022,
+ 0.131511,
+ 0.131023,
+ 0.131093,
+ 0.129626,
+ 0.131023,
+ 0.131023,
+ 0.131511,
+ 0.131092,
+ 0.131092,
+ 0.131581,
+ 0.130603,
+ 0.131511,
+ 0.131582,
+ 0.131023,
+ 0.130604,
+ 0.129626,
+ 0.131581,
+ 0.131512
+ ],
+ "warm_rows_per_second": 3899407.472848853
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1061596,
+ "unique_rows": 7764,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 126,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 126,
+ "zlib_blocks": 7,
+ "xor_blocks": 119,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7764
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "85acf75e956064d2922765527a96fd0f836c41a857cd6e98ebb3e438ca2d458e",
+ "table_sha256": "e3edb7720c73fa1f64d600770d4b6d66c9cfad69023e468c5c40378163a8fc5a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13075865502469242,
+ "representatives_sha256": "599101fdb6931287ef5564ea38c95caea41b60d7d1d429146ae354c4a7952ecc",
+ "structurally_merged_rows": 5660,
+ "backend_artifact_bytes": 1062845,
+ "backend_with_key_registry_bytes": 1266793,
+ "backend_with_datrie_key_index_bytes": 1171437,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9700998667021034,
+ "mean_cosine": 0.9978769785585703,
+ "p05_cosine": 0.9901246605818729,
+ "relative_l2_mean": 0.04258538715461437,
+ "relative_l2_p95": 0.14088813718565277,
+ "relative_l2_max": 0.24497508276344912,
+ "weighted_relative_l2_mean": 0.026546119474474904,
+ "weighted_relative_frobenius": 0.01246643812514673,
+ "weighted_mean_cosine": 0.9988771576024272,
+ "max_relative_norm_change": 0.0111298361935299
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9701794907300444,
+ "mean_cosine": 0.9978806276363834,
+ "p05_cosine": 0.9901037209797523,
+ "relative_l2_mean": 0.038836266683763034,
+ "relative_l2_p95": 0.14105104167974752,
+ "relative_l2_max": 0.24412266774317654,
+ "weighted_relative_l2_mean": 0.021487302568305788,
+ "weighted_relative_frobenius": 0.010142155276340896,
+ "weighted_mean_cosine": 0.9988895351383342,
+ "max_relative_norm_change": 0.010096948723578298
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.029113,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.13725895,
+ "warm_p99_ms": 0.14545402,
+ "samples_ms": [
+ 0.137657,
+ 0.134794,
+ 0.137238,
+ 0.134863,
+ 0.133886,
+ 0.133467,
+ 0.132,
+ 0.131022,
+ 0.131511,
+ 0.13249,
+ 0.132001,
+ 0.131581,
+ 0.132489,
+ 0.150439,
+ 0.132978,
+ 0.129626,
+ 0.131511,
+ 0.132908,
+ 0.129626,
+ 0.132001,
+ 0.131023,
+ 0.131511,
+ 0.132419,
+ 0.131022,
+ 0.132489,
+ 0.132419,
+ 0.131092,
+ 0.131511,
+ 0.130533,
+ 0.131022,
+ 0.130603,
+ 0.132908,
+ 0.132,
+ 0.130533,
+ 0.132419,
+ 0.131511,
+ 0.132,
+ 0.130534,
+ 0.130604,
+ 0.132908
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1060741,
+ "unique_rows": 7755,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 126,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 125,
+ "zlib_blocks": 7,
+ "xor_blocks": 118,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7755
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "e9a7984a19c4b069f4a9392f4ec1ae2fa47e05bd66dccd4a007f06f7de49f95e",
+ "table_sha256": "50f729d92051e4fe5eb1dcfd3110097fcd3ca04b2166d6fbc721ae9860db91f4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13268425199203193,
+ "representatives_sha256": "ab965745e290e733ba2e090edcc95679f8c068f19cf0fd16716985e6c65e2e04",
+ "structurally_merged_rows": 5669,
+ "backend_artifact_bytes": 1061990,
+ "backend_with_key_registry_bytes": 1265938,
+ "backend_with_datrie_key_index_bytes": 1170582,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9557065263986815,
+ "mean_cosine": 0.9978268548710147,
+ "p05_cosine": 0.9900065748645731,
+ "relative_l2_mean": 0.04292268929545207,
+ "relative_l2_p95": 0.1414940391005668,
+ "relative_l2_max": 0.2965704847402198,
+ "weighted_relative_l2_mean": 0.026965828093723367,
+ "weighted_relative_frobenius": 0.013918167697519409,
+ "weighted_mean_cosine": 0.9988134998170167,
+ "max_relative_norm_change": 0.0111298361935299
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9557439075746106,
+ "mean_cosine": 0.9978304961514793,
+ "p05_cosine": 0.9899764394161308,
+ "relative_l2_mean": 0.039177046495271935,
+ "relative_l2_p95": 0.1415856539324187,
+ "relative_l2_max": 0.2964602215300839,
+ "weighted_relative_l2_mean": 0.021909437042092,
+ "weighted_relative_frobenius": 0.011874855595628387,
+ "weighted_mean_cosine": 0.9988260130853636,
+ "max_relative_norm_change": 0.010096948723578298
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.154479,
+ "warm_p50_ms": 0.13008,
+ "warm_p95_ms": 0.14809529999999999,
+ "warm_p99_ms": 0.17825984,
+ "samples_ms": [
+ 0.134863,
+ 0.131511,
+ 0.133397,
+ 0.132489,
+ 0.131022,
+ 0.132419,
+ 0.132001,
+ 0.131022,
+ 0.129695,
+ 0.130044,
+ 0.157981,
+ 0.191225,
+ 0.135282,
+ 0.130045,
+ 0.134375,
+ 0.130115,
+ 0.132489,
+ 0.132908,
+ 0.130603,
+ 0.129625,
+ 0.147575,
+ 0.130534,
+ 0.130045,
+ 0.129556,
+ 0.129067,
+ 0.128648,
+ 0.128718,
+ 0.129626,
+ 0.12774,
+ 0.130603,
+ 0.12781,
+ 0.128718,
+ 0.129695,
+ 0.129695,
+ 0.131581,
+ 0.128717,
+ 0.128299,
+ 0.12781,
+ 0.129206,
+ 0.128648
+ ],
+ "warm_rows_per_second": 3936039.360393604
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 13424,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1771968,
+ "artifact_bytes": 1059135,
+ "unique_rows": 7742,
+ "row_map_bytes": 107392,
+ "n_rows": 13424,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 125,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 125,
+ "zlib_blocks": 6,
+ "xor_blocks": 119,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 7742
+ },
+ "source_float32_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05",
+ "decoded_float32_sha256": "0605382b46ed5748b8909f990c0419e8f45a4efa95ef7efebd85e0421392f8c7",
+ "table_sha256": "e1819522091ce34dc9daea36f0974bad0d1515c2e17b2a7f05325010841aa629",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13126444700174034,
+ "representatives_sha256": "f5f06c36d66192dec9d070f1b57787a8e3bd3df3a24b66630891d204b692d048",
+ "structurally_merged_rows": 5682,
+ "backend_artifact_bytes": 1060384,
+ "backend_with_key_registry_bytes": 1264332,
+ "backend_with_datrie_key_index_bytes": 1168976,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 13424,
+ "minimum_cosine": 0.9046011955085531,
+ "mean_cosine": 0.99771373339742,
+ "p05_cosine": 0.9898863383879815,
+ "relative_l2_mean": 0.043498349017067534,
+ "relative_l2_p95": 0.14221283141432034,
+ "relative_l2_max": 0.4373025064573066,
+ "weighted_relative_l2_mean": 0.02747763035423054,
+ "weighted_relative_frobenius": 0.016280209324312445,
+ "weighted_mean_cosine": 0.9987135768531176,
+ "max_relative_norm_change": 0.0111298361935299
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 13424,
+ "minimum_cosine": 0.9049461656915113,
+ "mean_cosine": 0.9977174368184777,
+ "p05_cosine": 0.9898406733669267,
+ "relative_l2_mean": 0.039758301746644985,
+ "relative_l2_p95": 0.1425129598986701,
+ "relative_l2_max": 0.4364542837019542,
+ "weighted_relative_l2_mean": 0.02242761496627819,
+ "weighted_relative_frobenius": 0.014570493521539054,
+ "weighted_mean_cosine": 0.9987261497009806,
+ "max_relative_norm_change": 0.010096948723578298
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.086383,
+ "warm_p50_ms": 0.1313015,
+ "warm_p95_ms": 0.13734249999999995,
+ "warm_p99_ms": 0.15659556,
+ "samples_ms": [
+ 0.149949,
+ 0.134375,
+ 0.133885,
+ 0.132489,
+ 0.132489,
+ 0.133466,
+ 0.132419,
+ 0.130603,
+ 0.132908,
+ 0.131511,
+ 0.132489,
+ 0.132489,
+ 0.132,
+ 0.131581,
+ 0.131511,
+ 0.130533,
+ 0.136679,
+ 0.130533,
+ 0.131092,
+ 0.129206,
+ 0.130673,
+ 0.130603,
+ 0.130114,
+ 0.130533,
+ 0.134445,
+ 0.129207,
+ 0.130604,
+ 0.130115,
+ 0.132,
+ 0.160845,
+ 0.134934,
+ 0.132489,
+ 0.130604,
+ 0.130115,
+ 0.131023,
+ 0.129626,
+ 0.130115,
+ 0.128648,
+ 0.128159,
+ 0.130534
+ ],
+ "warm_rows_per_second": 3899422.3219079752
+ }
+ }
+ ],
+ "baseline_quality": {},
+ "training_row_sha256": "2c882b270b22669d0de174b3941a1445e2f0dada1c2c346d1e5e0ee8d8063d05"
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 3,
+ "rows": 14132,
+ "dim": 128,
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "source_float32_bytes": 7235584,
+ "key_registry_bytes": 218478,
+ "unchanged_backbone_artifact_bytes": 0,
+ "alternative_datrie_key_index_bytes": 113712,
+ "query_rows": 512,
+ "query_ids_sha256": "48cdbdeffaad5c50239fec7524368c85bd12fc21eefeb42effaa6c7c5ff0c323",
+ "frequency_trace_sha256": "fc8afe9c1f9a53dfd826267bdad324151b311d81a4be4e4834bd129ccf363886",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 512,
+ "storage": {
+ "artifact_bytes": 7235584,
+ "unique_rows": 14132,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "table_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.03851464297622442,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 7236389,
+ "backend_with_key_registry_bytes": 7454867,
+ "backend_with_datrie_key_index_bytes": 7350101,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.377073,
+ "warm_p50_ms": 0.0603425,
+ "warm_p95_ms": 0.061981249999999995,
+ "warm_p99_ms": 0.06475656,
+ "samples_ms": [
+ 0.06188,
+ 0.060972,
+ 0.061531,
+ 0.059086,
+ 0.06146,
+ 0.059086,
+ 0.060552,
+ 0.059644,
+ 0.060552,
+ 0.06146,
+ 0.060133,
+ 0.060552,
+ 0.060623,
+ 0.0616,
+ 0.059505,
+ 0.059715,
+ 0.059574,
+ 0.065301,
+ 0.059575,
+ 0.060553,
+ 0.060972,
+ 0.061042,
+ 0.060064,
+ 0.059086,
+ 0.059644,
+ 0.060971,
+ 0.059156,
+ 0.059645,
+ 0.060552,
+ 0.059644,
+ 0.059644,
+ 0.060064,
+ 0.061111,
+ 0.063905,
+ 0.061111,
+ 0.059993,
+ 0.061111,
+ 0.059156,
+ 0.058667,
+ 0.059085
+ ],
+ "warm_rows_per_second": 8484898.703235697
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 6810278,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 884,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 884,
+ "zlib_blocks": 57,
+ "xor_blocks": 827,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "table_sha256": "f7683077057a2241d3358ae7b97510fae0590048137d3eadaf5ffc35c2e690ea",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6136417649686337,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6811487,
+ "backend_with_key_registry_bytes": 7029965,
+ "backend_with_datrie_key_index_bytes": 6925199,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 13.077025,
+ "warm_p50_ms": 0.07239100000000001,
+ "warm_p95_ms": 0.07666835,
+ "warm_p99_ms": 0.08426685999999999,
+ "samples_ms": [
+ 0.077663,
+ 0.076616,
+ 0.076547,
+ 0.075708,
+ 0.074311,
+ 0.0748,
+ 0.073822,
+ 0.075289,
+ 0.074242,
+ 0.074311,
+ 0.07466,
+ 0.074311,
+ 0.075149,
+ 0.072844,
+ 0.071377,
+ 0.070959,
+ 0.071378,
+ 0.071378,
+ 0.071937,
+ 0.071936,
+ 0.071447,
+ 0.070959,
+ 0.070959,
+ 0.070959,
+ 0.072426,
+ 0.071937,
+ 0.074311,
+ 0.071867,
+ 0.070051,
+ 0.069632,
+ 0.071936,
+ 0.072007,
+ 0.072426,
+ 0.071936,
+ 0.072356,
+ 0.072426,
+ 0.071937,
+ 0.088489,
+ 0.073334,
+ 0.072356
+ ],
+ "warm_rows_per_second": 7072702.40775787
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 6810278,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 884,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 884,
+ "zlib_blocks": 57,
+ "xor_blocks": 827,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "table_sha256": "f7683077057a2241d3358ae7b97510fae0590048137d3eadaf5ffc35c2e690ea",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6651246889960021,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6811486,
+ "backend_with_key_registry_bytes": 7029964,
+ "backend_with_datrie_key_index_bytes": 6925198,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 13.019336,
+ "warm_p50_ms": 0.07190150000000001,
+ "warm_p95_ms": 0.07951124999999999,
+ "warm_p99_ms": 0.08629181999999999,
+ "samples_ms": [
+ 0.07941,
+ 0.076616,
+ 0.075708,
+ 0.075219,
+ 0.075708,
+ 0.076616,
+ 0.081435,
+ 0.074242,
+ 0.072356,
+ 0.071867,
+ 0.071866,
+ 0.07047,
+ 0.071378,
+ 0.071447,
+ 0.071937,
+ 0.07047,
+ 0.070121,
+ 0.070121,
+ 0.070959,
+ 0.07047,
+ 0.076686,
+ 0.071029,
+ 0.070539,
+ 0.071518,
+ 0.074731,
+ 0.072426,
+ 0.071866,
+ 0.071936,
+ 0.070959,
+ 0.071517,
+ 0.089397,
+ 0.072425,
+ 0.073752,
+ 0.071866,
+ 0.072915,
+ 0.073334,
+ 0.071867,
+ 0.072355,
+ 0.071377,
+ 0.070889
+ ],
+ "warm_rows_per_second": 7120852.833390123
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 6810278,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 884,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 884,
+ "zlib_blocks": 57,
+ "xor_blocks": 827,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "table_sha256": "f7683077057a2241d3358ae7b97510fae0590048137d3eadaf5ffc35c2e690ea",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6406189650297165,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6811486,
+ "backend_with_key_registry_bytes": 7029964,
+ "backend_with_datrie_key_index_bytes": 6925198,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.988117,
+ "warm_p50_ms": 0.071448,
+ "warm_p95_ms": 0.07633319999999999,
+ "warm_p99_ms": 0.08820885999999999,
+ "samples_ms": [
+ 0.078921,
+ 0.094147,
+ 0.076197,
+ 0.075219,
+ 0.0748,
+ 0.072426,
+ 0.072355,
+ 0.072285,
+ 0.071867,
+ 0.071378,
+ 0.069912,
+ 0.070051,
+ 0.070051,
+ 0.07047,
+ 0.070889,
+ 0.071867,
+ 0.069912,
+ 0.072426,
+ 0.07047,
+ 0.071448,
+ 0.070959,
+ 0.071448,
+ 0.071937,
+ 0.071518,
+ 0.070121,
+ 0.072425,
+ 0.070051,
+ 0.071866,
+ 0.071378,
+ 0.071448,
+ 0.071866,
+ 0.070889,
+ 0.07047,
+ 0.071029,
+ 0.071937,
+ 0.072355,
+ 0.07054,
+ 0.070959,
+ 0.071936,
+ 0.070051
+ ],
+ "warm_rows_per_second": 7166050.834173105
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4307441,
+ "unique_rows": 8746,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 547,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 547,
+ "zlib_blocks": 326,
+ "xor_blocks": 221,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8746
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "667b8a6710c6beda63d49ba2caf0940536a5ad9e4edfa886a69c48c44c1b5a70",
+ "table_sha256": "153c0767c1982d9c87738e408eaf929f362d1fa24add1856f83418b62c849e68",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.40754929301328957,
+ "representatives_sha256": "bd1c7e95e0c55d05a3530176d1829d47bb1dda86fb4b52fa4bad88bcee35ef87",
+ "structurally_merged_rows": 5386,
+ "backend_artifact_bytes": 4308648,
+ "backend_with_key_registry_bytes": 4527126,
+ "backend_with_datrie_key_index_bytes": 4422360,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900015337324976,
+ "mean_cosine": 0.9985769339166507,
+ "p05_cosine": 0.9927441288359597,
+ "relative_l2_mean": 0.031006401826595625,
+ "relative_l2_p95": 0.12056503533434654,
+ "relative_l2_max": 0.14213808847272216,
+ "weighted_relative_l2_mean": 0.0178832102708785,
+ "weighted_relative_frobenius": 0.009597568466419546,
+ "weighted_mean_cosine": 0.9992198977556092,
+ "max_relative_norm_change": 0.010083232692545538
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900015337324976,
+ "mean_cosine": 0.9985769339166507,
+ "p05_cosine": 0.9927441288359597,
+ "relative_l2_mean": 0.031006401826595625,
+ "relative_l2_p95": 0.12056503533434654,
+ "relative_l2_max": 0.14213808847272216,
+ "weighted_relative_l2_mean": 0.0178832102708785,
+ "weighted_relative_frobenius": 0.009597568466419546,
+ "weighted_mean_cosine": 0.9992198977556092,
+ "max_relative_norm_change": 0.010083232692545538
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.767232,
+ "warm_p50_ms": 0.070889,
+ "warm_p95_ms": 0.0752225,
+ "warm_p99_ms": 0.07899596999999998,
+ "samples_ms": [
+ 0.081366,
+ 0.073822,
+ 0.072845,
+ 0.07054,
+ 0.070051,
+ 0.069493,
+ 0.070051,
+ 0.07473,
+ 0.073683,
+ 0.071378,
+ 0.070889,
+ 0.071378,
+ 0.07047,
+ 0.07054,
+ 0.07005,
+ 0.07054,
+ 0.075289,
+ 0.071029,
+ 0.073333,
+ 0.072844,
+ 0.072356,
+ 0.070889,
+ 0.07012,
+ 0.070051,
+ 0.068584,
+ 0.069981,
+ 0.069562,
+ 0.068584,
+ 0.0704,
+ 0.071867,
+ 0.072356,
+ 0.070889,
+ 0.073263,
+ 0.075219,
+ 0.071936,
+ 0.071029,
+ 0.070959,
+ 0.069073,
+ 0.068654,
+ 0.069073
+ ],
+ "warm_rows_per_second": 7222559.212289637
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4080103,
+ "unique_rows": 8266,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 517,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 517,
+ "zlib_blocks": 347,
+ "xor_blocks": 170,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8266
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "502e9617f8fa2ab8d43108fe433e21c326a22ebd963d22937cce03503b847a1e",
+ "table_sha256": "eda23ee1c368d0df35235856b12dcc5e4821ea26f6db92e1186b0cb901d0ffbd",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.37877186201512814,
+ "representatives_sha256": "a7907d8deeaa847a4f7a0cca2eaaf5fd223458d3742a31ada2d2ba00926153c6",
+ "structurally_merged_rows": 5866,
+ "backend_artifact_bytes": 4081310,
+ "backend_with_key_registry_bytes": 4299788,
+ "backend_with_datrie_key_index_bytes": 4195022,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800370872506456,
+ "mean_cosine": 0.9981046556900883,
+ "p05_cosine": 0.9907898956835881,
+ "relative_l2_mean": 0.03665354715979584,
+ "relative_l2_p95": 0.1358853294819434,
+ "relative_l2_max": 0.20056834471285476,
+ "weighted_relative_l2_mean": 0.02082327130666535,
+ "weighted_relative_frobenius": 0.011449487051204693,
+ "weighted_mean_cosine": 0.9989732798840621,
+ "max_relative_norm_change": 0.010097242954185409
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800370872506456,
+ "mean_cosine": 0.9981046556900883,
+ "p05_cosine": 0.9907898956835881,
+ "relative_l2_mean": 0.03665354715979584,
+ "relative_l2_p95": 0.1358853294819434,
+ "relative_l2_max": 0.20056834471285476,
+ "weighted_relative_l2_mean": 0.02082327130666535,
+ "weighted_relative_frobenius": 0.011449487051204693,
+ "weighted_mean_cosine": 0.9989732798840621,
+ "max_relative_norm_change": 0.010097242954185409
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.947074,
+ "warm_p50_ms": 0.0707145,
+ "warm_p95_ms": 0.0771574,
+ "warm_p99_ms": 0.09250995999999999,
+ "samples_ms": [
+ 0.078153,
+ 0.077105,
+ 0.07473,
+ 0.073892,
+ 0.072914,
+ 0.072844,
+ 0.072356,
+ 0.073334,
+ 0.072356,
+ 0.071378,
+ 0.072844,
+ 0.071866,
+ 0.070959,
+ 0.073334,
+ 0.069981,
+ 0.070889,
+ 0.069142,
+ 0.07054,
+ 0.069073,
+ 0.069492,
+ 0.073822,
+ 0.073822,
+ 0.101689,
+ 0.072356,
+ 0.070958,
+ 0.07047,
+ 0.07054,
+ 0.069143,
+ 0.070051,
+ 0.068515,
+ 0.069632,
+ 0.069073,
+ 0.069493,
+ 0.069073,
+ 0.069492,
+ 0.069981,
+ 0.069632,
+ 0.069073,
+ 0.07047,
+ 0.069003
+ ],
+ "warm_rows_per_second": 7240382.099852223
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4045652,
+ "unique_rows": 8196,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 513,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 513,
+ "zlib_blocks": 351,
+ "xor_blocks": 162,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8196
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "6f80a541f40d8ddb31b97b8b5e34e996b3e5047745dbf2e053efb92dacebcbf2",
+ "table_sha256": "627595a6e7ded85684f853b79726edbe749601f5c0b17efd194fdddd1cd62e02",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4442353490740061,
+ "representatives_sha256": "e1a1ae7ddfddf044ffde352196fc3347d86678c07064681ca6a053413eb30d0c",
+ "structurally_merged_rows": 5936,
+ "backend_artifact_bytes": 4046859,
+ "backend_with_key_registry_bytes": 4265337,
+ "backend_with_datrie_key_index_bytes": 4160571,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9705909970636375,
+ "mean_cosine": 0.9979711391605847,
+ "p05_cosine": 0.9903802210783794,
+ "relative_l2_mean": 0.03784439763998729,
+ "relative_l2_p95": 0.13865994162687642,
+ "relative_l2_max": 0.24349894495101002,
+ "weighted_relative_l2_mean": 0.021413248484090965,
+ "weighted_relative_frobenius": 0.011589934801130889,
+ "weighted_mean_cosine": 0.9989071324486565,
+ "max_relative_norm_change": 0.010097242954185409
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9705909970636375,
+ "mean_cosine": 0.9979711391605847,
+ "p05_cosine": 0.9903802210783794,
+ "relative_l2_mean": 0.03784439763998729,
+ "relative_l2_p95": 0.13865994162687642,
+ "relative_l2_max": 0.24349894495101002,
+ "weighted_relative_l2_mean": 0.021413248484090965,
+ "weighted_relative_frobenius": 0.011589934801130889,
+ "weighted_mean_cosine": 0.9989071324486565,
+ "max_relative_norm_change": 0.010097242954185409
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.804038,
+ "warm_p50_ms": 0.0709585,
+ "warm_p95_ms": 0.0790389,
+ "warm_p99_ms": 0.08312109,
+ "samples_ms": [
+ 0.079968,
+ 0.076267,
+ 0.074312,
+ 0.073334,
+ 0.071867,
+ 0.07899,
+ 0.07054,
+ 0.07054,
+ 0.069003,
+ 0.069074,
+ 0.069143,
+ 0.07054,
+ 0.068585,
+ 0.069562,
+ 0.069143,
+ 0.069982,
+ 0.069981,
+ 0.070051,
+ 0.069003,
+ 0.069492,
+ 0.070051,
+ 0.072845,
+ 0.072426,
+ 0.071448,
+ 0.070959,
+ 0.071378,
+ 0.075289,
+ 0.073823,
+ 0.085137,
+ 0.072845,
+ 0.070958,
+ 0.070959,
+ 0.07047,
+ 0.070051,
+ 0.071518,
+ 0.071448,
+ 0.072495,
+ 0.071518,
+ 0.069981,
+ 0.069073
+ ],
+ "warm_rows_per_second": 7215485.107492409
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4035247,
+ "unique_rows": 8175,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 511,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 511,
+ "zlib_blocks": 354,
+ "xor_blocks": 157,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8175
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "9950c39558a2e16ecd8020319f22b7f21053f15dfb4ac94a30135d35b19449a3",
+ "table_sha256": "f88164237644c70a9fb4052c30ede88c597be801cc1bc4bdf7aa70cc73749f44",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.38341121398843825,
+ "representatives_sha256": "49ba1ef3554d37429d76dea99b128bc135acf26008bcac1c3810e3a3bbacc4f0",
+ "structurally_merged_rows": 5957,
+ "backend_artifact_bytes": 4036454,
+ "backend_with_key_registry_bytes": 4254932,
+ "backend_with_datrie_key_index_bytes": 4150166,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9519201373230705,
+ "mean_cosine": 0.9979023065652721,
+ "p05_cosine": 0.9902684251190589,
+ "relative_l2_mean": 0.03833110127692099,
+ "relative_l2_p95": 0.1396269426984356,
+ "relative_l2_max": 0.3099892446519552,
+ "weighted_relative_l2_mean": 0.021686112527658816,
+ "weighted_relative_frobenius": 0.011889347328295592,
+ "weighted_mean_cosine": 0.9988684579740494,
+ "max_relative_norm_change": 0.010097242954185409
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9519201373230705,
+ "mean_cosine": 0.9979023065652721,
+ "p05_cosine": 0.9902684251190589,
+ "relative_l2_mean": 0.03833110127692099,
+ "relative_l2_p95": 0.1396269426984356,
+ "relative_l2_max": 0.3099892446519552,
+ "weighted_relative_l2_mean": 0.021686112527658816,
+ "weighted_relative_frobenius": 0.011889347328295592,
+ "weighted_mean_cosine": 0.9988684579740494,
+ "max_relative_norm_change": 0.010097242954185409
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.631879,
+ "warm_p50_ms": 0.07050500000000001,
+ "warm_p95_ms": 0.0775975,
+ "warm_p99_ms": 0.09001833,
+ "samples_ms": [
+ 0.076197,
+ 0.077594,
+ 0.077664,
+ 0.073822,
+ 0.073822,
+ 0.074241,
+ 0.071866,
+ 0.071866,
+ 0.070959,
+ 0.071867,
+ 0.0704,
+ 0.070959,
+ 0.069143,
+ 0.07047,
+ 0.069073,
+ 0.07047,
+ 0.097917,
+ 0.073333,
+ 0.069073,
+ 0.069562,
+ 0.069632,
+ 0.07047,
+ 0.070469,
+ 0.071378,
+ 0.069003,
+ 0.071029,
+ 0.07054,
+ 0.069073,
+ 0.069074,
+ 0.069073,
+ 0.069562,
+ 0.070958,
+ 0.071378,
+ 0.070959,
+ 0.07047,
+ 0.069143,
+ 0.069073,
+ 0.069004,
+ 0.070959,
+ 0.070121
+ ],
+ "warm_rows_per_second": 7261896.319409969
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4027967,
+ "unique_rows": 8159,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 510,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 510,
+ "zlib_blocks": 348,
+ "xor_blocks": 162,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8159
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "6034934eff3b90eca02d80e44a5d9cd2e9e5c08b5c10e88e8912fdd4cb496cb0",
+ "table_sha256": "93d8683f589c380bf7e4c6361dceec90d113c9aec7cdab841b2721d4ed7640b7",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4235941560473293,
+ "representatives_sha256": "24ea241cf3e96037e0e4faf4f1b2273c0896e9a063a205b714b24f3ec65497cf",
+ "structurally_merged_rows": 5973,
+ "backend_artifact_bytes": 4029174,
+ "backend_with_key_registry_bytes": 4247652,
+ "backend_with_datrie_key_index_bytes": 4142886,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9069112665063119,
+ "mean_cosine": 0.9977733189647556,
+ "p05_cosine": 0.9900972903136396,
+ "relative_l2_mean": 0.03896459512890429,
+ "relative_l2_p95": 0.14063370589982319,
+ "relative_l2_max": 0.4315442500650112,
+ "weighted_relative_l2_mean": 0.02236025039160826,
+ "weighted_relative_frobenius": 0.015609319162613444,
+ "weighted_mean_cosine": 0.9987317415365643,
+ "max_relative_norm_change": 0.010097242954185409
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9069112665063119,
+ "mean_cosine": 0.9977733189647556,
+ "p05_cosine": 0.9900972903136396,
+ "relative_l2_mean": 0.03896459512890429,
+ "relative_l2_p95": 0.14063370589982319,
+ "relative_l2_max": 0.4315442500650112,
+ "weighted_relative_l2_mean": 0.02236025039160826,
+ "weighted_relative_frobenius": 0.015609319162613444,
+ "weighted_mean_cosine": 0.9987317415365643,
+ "max_relative_norm_change": 0.010097242954185409
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.623009,
+ "warm_p50_ms": 0.07047,
+ "warm_p95_ms": 0.07622145,
+ "warm_p99_ms": 0.08299096,
+ "samples_ms": [
+ 0.076686,
+ 0.072426,
+ 0.074311,
+ 0.073823,
+ 0.072356,
+ 0.073333,
+ 0.072356,
+ 0.076127,
+ 0.076197,
+ 0.072914,
+ 0.071518,
+ 0.069981,
+ 0.069492,
+ 0.070051,
+ 0.070121,
+ 0.070051,
+ 0.069493,
+ 0.070051,
+ 0.072355,
+ 0.072356,
+ 0.071937,
+ 0.070051,
+ 0.069562,
+ 0.072425,
+ 0.071517,
+ 0.07047,
+ 0.07047,
+ 0.068165,
+ 0.068584,
+ 0.070051,
+ 0.070051,
+ 0.069073,
+ 0.087022,
+ 0.072007,
+ 0.071937,
+ 0.069492,
+ 0.068096,
+ 0.070121,
+ 0.069981,
+ 0.069492
+ ],
+ "warm_rows_per_second": 7265503.050943663
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 3617792,
+ "unique_rows": 14132,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "965e67427fd2f63b0379f2c343356af421d0b17cd7971cffda2236625ac6d2b2",
+ "table_sha256": "f398ae975042657c209fa7ae6041b7bf9b9ad537f36366d974ec5992fba17f9b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.045753000071272254,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3618598,
+ "backend_with_key_registry_bytes": 3837076,
+ "backend_with_datrie_key_index_bytes": 3732310,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999614942229,
+ "mean_cosine": 0.9999999789582108,
+ "p05_cosine": 0.9999999733717188,
+ "relative_l2_mean": 0.00020684625216337265,
+ "relative_l2_p95": 0.00023325418518764534,
+ "relative_l2_max": 0.0002805592168150135,
+ "weighted_relative_l2_mean": 0.0002068157332193522,
+ "weighted_relative_frobenius": 0.00020571840501630532,
+ "weighted_mean_cosine": 0.9999999789659619,
+ "max_relative_norm_change": 0.00012469551075047655
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.545042,
+ "warm_p50_ms": 0.22007,
+ "warm_p95_ms": 0.22245169999999997,
+ "warm_p99_ms": 0.23490434,
+ "samples_ms": [
+ 0.221886,
+ 0.221886,
+ 0.220978,
+ 0.22007,
+ 0.22049,
+ 0.22007,
+ 0.220978,
+ 0.219511,
+ 0.219511,
+ 0.219022,
+ 0.22,
+ 0.220978,
+ 0.221398,
+ 0.235994,
+ 0.221397,
+ 0.219512,
+ 0.220489,
+ 0.219022,
+ 0.220908,
+ 0.218603,
+ 0.219511,
+ 0.219092,
+ 0.221886,
+ 0.219581,
+ 0.219581,
+ 0.22049,
+ 0.22007,
+ 0.219092,
+ 0.220419,
+ 0.218952,
+ 0.220419,
+ 0.2332,
+ 0.219093,
+ 0.219581,
+ 0.219581,
+ 0.220489,
+ 0.220908,
+ 0.219931,
+ 0.219931,
+ 0.22
+ ],
+ "warm_rows_per_second": 2326532.4669423364
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 3406763,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 442,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 442,
+ "zlib_blocks": 0,
+ "xor_blocks": 442,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "965e67427fd2f63b0379f2c343356af421d0b17cd7971cffda2236625ac6d2b2",
+ "table_sha256": "e7718e7ec3faaa2b3b85f43fd8f38dd8e523f5dbb78b0d66b41877199263ae44",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.34386564302258193,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3407972,
+ "backend_with_key_registry_bytes": 3626450,
+ "backend_with_datrie_key_index_bytes": 3521684,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999614942229,
+ "mean_cosine": 0.9999999789582108,
+ "p05_cosine": 0.9999999733717188,
+ "relative_l2_mean": 0.00020684625216337265,
+ "relative_l2_p95": 0.00023325418518764534,
+ "relative_l2_max": 0.0002805592168150135,
+ "weighted_relative_l2_mean": 0.0002068157332193522,
+ "weighted_relative_frobenius": 0.00020571840501630532,
+ "weighted_mean_cosine": 0.9999999789659619,
+ "max_relative_norm_change": 0.00012469551075047655
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.826585,
+ "warm_p50_ms": 0.2285555,
+ "warm_p95_ms": 0.23616084999999995,
+ "warm_p99_ms": 0.25177478,
+ "samples_ms": [
+ 0.233759,
+ 0.230407,
+ 0.229429,
+ 0.227613,
+ 0.235644,
+ 0.229429,
+ 0.230406,
+ 0.228032,
+ 0.229918,
+ 0.229847,
+ 0.22901,
+ 0.22901,
+ 0.255479,
+ 0.229429,
+ 0.227613,
+ 0.228521,
+ 0.228102,
+ 0.228032,
+ 0.227124,
+ 0.227543,
+ 0.227613,
+ 0.22859,
+ 0.229429,
+ 0.228101,
+ 0.227054,
+ 0.228451,
+ 0.22887,
+ 0.228032,
+ 0.22894,
+ 0.245981,
+ 0.230895,
+ 0.22901,
+ 0.227962,
+ 0.227124,
+ 0.22901,
+ 0.228102,
+ 0.227613,
+ 0.228521,
+ 0.228032,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2240156.110878977
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 3406763,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 442,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 442,
+ "zlib_blocks": 0,
+ "xor_blocks": 442,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "965e67427fd2f63b0379f2c343356af421d0b17cd7971cffda2236625ac6d2b2",
+ "table_sha256": "e7718e7ec3faaa2b3b85f43fd8f38dd8e523f5dbb78b0d66b41877199263ae44",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.44691041205078363,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3407971,
+ "backend_with_key_registry_bytes": 3626449,
+ "backend_with_datrie_key_index_bytes": 3521683,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999614942229,
+ "mean_cosine": 0.9999999789582108,
+ "p05_cosine": 0.9999999733717188,
+ "relative_l2_mean": 0.00020684625216337265,
+ "relative_l2_p95": 0.00023325418518764534,
+ "relative_l2_max": 0.0002805592168150135,
+ "weighted_relative_l2_mean": 0.0002068157332193522,
+ "weighted_relative_frobenius": 0.00020571840501630532,
+ "weighted_mean_cosine": 0.9999999789659619,
+ "max_relative_norm_change": 0.00012469551075047655
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.836992,
+ "warm_p50_ms": 0.22897499999999998,
+ "warm_p95_ms": 0.23543884999999998,
+ "warm_p99_ms": 0.24771515,
+ "samples_ms": [
+ 0.234178,
+ 0.230896,
+ 0.22894,
+ 0.22901,
+ 0.22901,
+ 0.229917,
+ 0.228521,
+ 0.22901,
+ 0.228032,
+ 0.228521,
+ 0.231804,
+ 0.228521,
+ 0.22901,
+ 0.252128,
+ 0.235156,
+ 0.229848,
+ 0.22894,
+ 0.228451,
+ 0.228521,
+ 0.22894,
+ 0.228032,
+ 0.228451,
+ 0.22894,
+ 0.22894,
+ 0.228521,
+ 0.22901,
+ 0.22901,
+ 0.22894,
+ 0.229079,
+ 0.22901,
+ 0.240813,
+ 0.230826,
+ 0.22901,
+ 0.227613,
+ 0.22901,
+ 0.229079,
+ 0.227613,
+ 0.227613,
+ 0.22859,
+ 0.228032
+ ],
+ "warm_rows_per_second": 2236051.970739164
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 3406763,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 442,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 442,
+ "zlib_blocks": 0,
+ "xor_blocks": 442,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "965e67427fd2f63b0379f2c343356af421d0b17cd7971cffda2236625ac6d2b2",
+ "table_sha256": "e7718e7ec3faaa2b3b85f43fd8f38dd8e523f5dbb78b0d66b41877199263ae44",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.34564261697232723,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3407971,
+ "backend_with_key_registry_bytes": 3626449,
+ "backend_with_datrie_key_index_bytes": 3521683,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999614942229,
+ "mean_cosine": 0.9999999789582108,
+ "p05_cosine": 0.9999999733717188,
+ "relative_l2_mean": 0.00020684625216337265,
+ "relative_l2_p95": 0.00023325418518764534,
+ "relative_l2_max": 0.0002805592168150135,
+ "weighted_relative_l2_mean": 0.0002068157332193522,
+ "weighted_relative_frobenius": 0.00020571840501630532,
+ "weighted_mean_cosine": 0.9999999789659619,
+ "max_relative_norm_change": 0.00012469551075047655
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.869678,
+ "warm_p50_ms": 0.22900949999999998,
+ "warm_p95_ms": 0.23279234999999998,
+ "warm_p99_ms": 0.25237672,
+ "samples_ms": [
+ 0.232293,
+ 0.230337,
+ 0.230406,
+ 0.22901,
+ 0.22908,
+ 0.229499,
+ 0.22894,
+ 0.22901,
+ 0.227683,
+ 0.229429,
+ 0.22901,
+ 0.229009,
+ 0.24228,
+ 0.231314,
+ 0.229428,
+ 0.230337,
+ 0.228451,
+ 0.228521,
+ 0.22894,
+ 0.22894,
+ 0.228032,
+ 0.228032,
+ 0.228451,
+ 0.227962,
+ 0.229009,
+ 0.22901,
+ 0.228451,
+ 0.228381,
+ 0.22894,
+ 0.258832,
+ 0.232292,
+ 0.22894,
+ 0.229429,
+ 0.228451,
+ 0.228521,
+ 0.229498,
+ 0.229988,
+ 0.229499,
+ 0.228521,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2235715.1122551686
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2208336,
+ "unique_rows": 8746,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 274,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 274,
+ "zlib_blocks": 89,
+ "xor_blocks": 185,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8746
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "6b6c28446000c3faa7f71c67269342ba9a31273cb6cca044dce1932a6c008f60",
+ "table_sha256": "7a840443c32a6e758971b6a8e21ad75ea5a2829494c699881e228f20d756fa13",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2349200490862131,
+ "representatives_sha256": "388497d9f767962dcf6fcfa9c748f8aaf8235df91625d576f5c28cc15aa8d59d",
+ "structurally_merged_rows": 5386,
+ "backend_artifact_bytes": 2209542,
+ "backend_with_key_registry_bytes": 2428020,
+ "backend_with_datrie_key_index_bytes": 2323254,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900081908684233,
+ "mean_cosine": 0.9985776215363579,
+ "p05_cosine": 0.9927595161626462,
+ "relative_l2_mean": 0.031127708198948993,
+ "relative_l2_p95": 0.12041610625178735,
+ "relative_l2_max": 0.14217809684457094,
+ "weighted_relative_l2_mean": 0.018037095789352302,
+ "weighted_relative_frobenius": 0.009600233959965716,
+ "weighted_mean_cosine": 0.9992202466134478,
+ "max_relative_norm_change": 0.01008429606092528
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900063056169309,
+ "mean_cosine": 0.9985776202179737,
+ "p05_cosine": 0.9927590619696117,
+ "relative_l2_mean": 0.030999892838835974,
+ "relative_l2_p95": 0.12044711814496074,
+ "relative_l2_max": 0.1421703389032944,
+ "weighted_relative_l2_mean": 0.017879802244459758,
+ "weighted_relative_frobenius": 0.009598270767519452,
+ "weighted_mean_cosine": 0.9992202531533373,
+ "max_relative_norm_change": 0.010089724453680103
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.133572,
+ "warm_p50_ms": 0.2280665,
+ "warm_p95_ms": 0.23551899999999998,
+ "warm_p99_ms": 0.25090313,
+ "samples_ms": [
+ 0.235086,
+ 0.229988,
+ 0.228101,
+ 0.232293,
+ 0.230477,
+ 0.231385,
+ 0.231315,
+ 0.228032,
+ 0.22908,
+ 0.228102,
+ 0.227124,
+ 0.230896,
+ 0.229499,
+ 0.231384,
+ 0.23362,
+ 0.243746,
+ 0.232292,
+ 0.231384,
+ 0.228032,
+ 0.227962,
+ 0.227054,
+ 0.227124,
+ 0.227543,
+ 0.227613,
+ 0.227613,
+ 0.227544,
+ 0.227543,
+ 0.225657,
+ 0.227194,
+ 0.228102,
+ 0.227543,
+ 0.226635,
+ 0.226076,
+ 0.255479,
+ 0.230896,
+ 0.229918,
+ 0.228032,
+ 0.227054,
+ 0.227543,
+ 0.227054
+ ],
+ "warm_rows_per_second": 2244959.255304922
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2096994,
+ "unique_rows": 8265,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 259,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 259,
+ "zlib_blocks": 127,
+ "xor_blocks": 132,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8265
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "3b0e0f6ec11e3df889b0bc8e1af1d5360039f6d5417d8c252e86d654544c5c58",
+ "table_sha256": "6a2b12f7c64f809e1fe69791247c692f725f61a10222e638bf8b2b0fafdfe0bd",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22290237713605165,
+ "representatives_sha256": "feffcce9b3c698f4da826db68af8e6ef29a638f659a8b3af5889bbb80fa528eb",
+ "structurally_merged_rows": 5867,
+ "backend_artifact_bytes": 2098201,
+ "backend_with_key_registry_bytes": 2316679,
+ "backend_with_datrie_key_index_bytes": 2211913,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800421398034684,
+ "mean_cosine": 0.9981047704915305,
+ "p05_cosine": 0.9907985880247167,
+ "relative_l2_mean": 0.03677686713143057,
+ "relative_l2_p95": 0.13584521041237388,
+ "relative_l2_max": 0.20054694321241362,
+ "weighted_relative_l2_mean": 0.02097795549079116,
+ "weighted_relative_frobenius": 0.011451663323190498,
+ "weighted_mean_cosine": 0.998973349064148,
+ "max_relative_norm_change": 0.010093588270012993
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800401520236893,
+ "mean_cosine": 0.9981047640986421,
+ "p05_cosine": 0.9907991999638034,
+ "relative_l2_mean": 0.03665613588618342,
+ "relative_l2_p95": 0.13583220356511932,
+ "relative_l2_max": 0.20055616801789974,
+ "weighted_relative_l2_mean": 0.02082433902623526,
+ "weighted_relative_frobenius": 0.01144982086855333,
+ "weighted_mean_cosine": 0.9989733547391864,
+ "max_relative_norm_change": 0.010100504012249704
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.084404,
+ "warm_p50_ms": 0.22897499999999998,
+ "warm_p95_ms": 0.24273029999999998,
+ "warm_p99_ms": 0.25551289,
+ "samples_ms": [
+ 0.232851,
+ 0.230826,
+ 0.229918,
+ 0.252616,
+ 0.236623,
+ 0.234736,
+ 0.231804,
+ 0.230407,
+ 0.229429,
+ 0.228939,
+ 0.228451,
+ 0.232292,
+ 0.231384,
+ 0.230406,
+ 0.230895,
+ 0.229499,
+ 0.228521,
+ 0.227613,
+ 0.229498,
+ 0.227124,
+ 0.257365,
+ 0.231803,
+ 0.22901,
+ 0.228521,
+ 0.227543,
+ 0.22894,
+ 0.227124,
+ 0.227054,
+ 0.22852,
+ 0.228521,
+ 0.228521,
+ 0.228521,
+ 0.226216,
+ 0.227682,
+ 0.226636,
+ 0.227962,
+ 0.227543,
+ 0.24221,
+ 0.229429,
+ 0.228102
+ ],
+ "warm_rows_per_second": 2236051.970739164
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2079846,
+ "unique_rows": 8195,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 257,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 256,
+ "zlib_blocks": 118,
+ "xor_blocks": 138,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8195
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "1cbd7860e9ec4d28e8fb5160419d8b7999362e0852d68370160995df52a01449",
+ "table_sha256": "f350383ebdb9829a2c1a8d98d0044bf47932367442eeb900c412e56ea04da56f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22908319695852697,
+ "representatives_sha256": "e0ebbe2f8b9ec11acf0083be0cadc18da4efa8c4bd62ac90eda390a69035149b",
+ "structurally_merged_rows": 5937,
+ "backend_artifact_bytes": 2081053,
+ "backend_with_key_registry_bytes": 2299531,
+ "backend_with_datrie_key_index_bytes": 2194765,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.970585578118649,
+ "mean_cosine": 0.9979712542989172,
+ "p05_cosine": 0.990390049273828,
+ "relative_l2_mean": 0.037966665140841765,
+ "relative_l2_p95": 0.13859286993259123,
+ "relative_l2_max": 0.24352684348208944,
+ "weighted_relative_l2_mean": 0.02156741124781128,
+ "weighted_relative_frobenius": 0.011592084095370938,
+ "weighted_mean_cosine": 0.9989072017956463,
+ "max_relative_norm_change": 0.010093588270012993
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9705884310415499,
+ "mean_cosine": 0.9979712447439355,
+ "p05_cosine": 0.990390953577697,
+ "relative_l2_mean": 0.03784700510746191,
+ "relative_l2_p95": 0.1385812581912458,
+ "relative_l2_max": 0.24350950083384507,
+ "weighted_relative_l2_mean": 0.021414325488465155,
+ "weighted_relative_frobenius": 0.011590271896844764,
+ "weighted_mean_cosine": 0.9989072059041043,
+ "max_relative_norm_change": 0.010100504012249704
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.150963,
+ "warm_p50_ms": 0.22901,
+ "warm_p95_ms": 0.28583990000000004,
+ "warm_p99_ms": 0.28847273,
+ "samples_ms": [
+ 0.231873,
+ 0.229009,
+ 0.229429,
+ 0.229429,
+ 0.229429,
+ 0.229918,
+ 0.229498,
+ 0.240324,
+ 0.289562,
+ 0.286769,
+ 0.285791,
+ 0.228032,
+ 0.22901,
+ 0.228451,
+ 0.22901,
+ 0.228521,
+ 0.228032,
+ 0.232223,
+ 0.228521,
+ 0.228102,
+ 0.229079,
+ 0.233689,
+ 0.228032,
+ 0.227543,
+ 0.263022,
+ 0.231314,
+ 0.230896,
+ 0.228521,
+ 0.227474,
+ 0.227543,
+ 0.228032,
+ 0.22901,
+ 0.226635,
+ 0.228032,
+ 0.230407,
+ 0.229428,
+ 0.228451,
+ 0.228032,
+ 0.227124,
+ 0.227613
+ ],
+ "warm_rows_per_second": 2235710.2309942795
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2074878,
+ "unique_rows": 8174,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 256,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 256,
+ "zlib_blocks": 124,
+ "xor_blocks": 132,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8174
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "8c15ce75b7083992f68e04ddec4b68dc6c61e1013303ddaad47dda9656d689d3",
+ "table_sha256": "33c103d9a26b56b7945dcb400146290041cadc977a92c452ccb373a89f89a409",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22787759499624372,
+ "representatives_sha256": "117f503a7b9ba36569800ce516775efd7241f0e7d79707430cb64300952ac9f8",
+ "structurally_merged_rows": 5958,
+ "backend_artifact_bytes": 2076085,
+ "backend_with_key_registry_bytes": 2294563,
+ "backend_with_datrie_key_index_bytes": 2189797,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9519140491379375,
+ "mean_cosine": 0.9979024186868628,
+ "p05_cosine": 0.9902829300193834,
+ "relative_l2_mean": 0.0384530659329802,
+ "relative_l2_p95": 0.1395751234873996,
+ "relative_l2_max": 0.31000986075714415,
+ "weighted_relative_l2_mean": 0.021840116887272,
+ "weighted_relative_frobenius": 0.011891428298117412,
+ "weighted_mean_cosine": 0.9988685261740616,
+ "max_relative_norm_change": 0.010093588270012993
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9519135406106861,
+ "mean_cosine": 0.9979024112977994,
+ "p05_cosine": 0.9902785476081027,
+ "relative_l2_mean": 0.03833370890301321,
+ "relative_l2_p95": 0.13956606440344563,
+ "relative_l2_max": 0.31001166950115155,
+ "weighted_relative_l2_mean": 0.02168718814151804,
+ "weighted_relative_frobenius": 0.011889664115582324,
+ "weighted_mean_cosine": 0.9988685313492044,
+ "max_relative_norm_change": 0.010100504012249704
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.922093,
+ "warm_p50_ms": 0.2292545,
+ "warm_p95_ms": 0.23645164999999996,
+ "warm_p99_ms": 0.25948901,
+ "samples_ms": [
+ 0.235645,
+ 0.234178,
+ 0.264419,
+ 0.232292,
+ 0.229499,
+ 0.231385,
+ 0.231804,
+ 0.231804,
+ 0.231804,
+ 0.230896,
+ 0.228521,
+ 0.228521,
+ 0.23327,
+ 0.232781,
+ 0.22894,
+ 0.228101,
+ 0.22894,
+ 0.229917,
+ 0.226705,
+ 0.251778,
+ 0.229499,
+ 0.227613,
+ 0.227543,
+ 0.22901,
+ 0.226635,
+ 0.228871,
+ 0.22894,
+ 0.229009,
+ 0.229987,
+ 0.228032,
+ 0.229429,
+ 0.230407,
+ 0.227613,
+ 0.226635,
+ 0.228102,
+ 0.228591,
+ 0.22908,
+ 0.229429,
+ 0.229499,
+ 0.228102
+ ],
+ "warm_rows_per_second": 2233325.8452942036
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2070983,
+ "unique_rows": 8158,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 255,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 255,
+ "zlib_blocks": 124,
+ "xor_blocks": 131,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8158
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "dc2312b22728a2665d920a5ec2a3e1acbb0d31cdf8a55690286ff5fa8920e668",
+ "table_sha256": "034383ffcb0d1a5e2a49f2486b82698ed00ff6d58c53f602d2c123f020ecc832",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22084003197960556,
+ "representatives_sha256": "f381bf6fdd18467318b6364e1ef6cbb5bc5dfb715055742acff2694626a03b18",
+ "structurally_merged_rows": 5974,
+ "backend_artifact_bytes": 2072190,
+ "backend_with_key_registry_bytes": 2290668,
+ "backend_with_datrie_key_index_bytes": 2185902,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9069161678324782,
+ "mean_cosine": 0.9977734249342971,
+ "p05_cosine": 0.9901206520914708,
+ "relative_l2_mean": 0.03908632558061861,
+ "relative_l2_p95": 0.14059008562937259,
+ "relative_l2_max": 0.4315264679104603,
+ "weighted_relative_l2_mean": 0.02251396925269012,
+ "weighted_relative_frobenius": 0.015610929943765554,
+ "weighted_mean_cosine": 0.9987318073693344,
+ "max_relative_norm_change": 0.010093588270012993
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9069171665107606,
+ "mean_cosine": 0.997773420022632,
+ "p05_cosine": 0.9901177184637575,
+ "relative_l2_mean": 0.03896720916489434,
+ "relative_l2_p95": 0.1405727746847956,
+ "relative_l2_max": 0.43152819830688083,
+ "weighted_relative_l2_mean": 0.022361326989332794,
+ "weighted_relative_frobenius": 0.015609563474743491,
+ "weighted_mean_cosine": 0.9987318146454679,
+ "max_relative_norm_change": 0.010100504012249704
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.005414,
+ "warm_p50_ms": 0.229429,
+ "warm_p95_ms": 0.2876559,
+ "warm_p99_ms": 0.29062972,
+ "samples_ms": [
+ 0.232851,
+ 0.230895,
+ 0.230895,
+ 0.229429,
+ 0.22901,
+ 0.22894,
+ 0.229429,
+ 0.229429,
+ 0.228451,
+ 0.228521,
+ 0.228451,
+ 0.228032,
+ 0.231315,
+ 0.232293,
+ 0.231804,
+ 0.260648,
+ 0.291937,
+ 0.288585,
+ 0.287187,
+ 0.28621,
+ 0.287607,
+ 0.22894,
+ 0.231384,
+ 0.230895,
+ 0.230896,
+ 0.230896,
+ 0.227613,
+ 0.229429,
+ 0.22901,
+ 0.228591,
+ 0.226636,
+ 0.250172,
+ 0.229499,
+ 0.226565,
+ 0.226705,
+ 0.228521,
+ 0.228521,
+ 0.228102,
+ 0.227613,
+ 0.228451
+ ],
+ "warm_rows_per_second": 2231627.2136477954
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 132,
+ "storage": {
+ "artifact_bytes": 1865424,
+ "unique_rows": 14132,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "d37754bf9356e1c82806b733d2f09d6fcd41d5b2725b70c407328b1a5c7fe149",
+ "table_sha256": "4bdc5394076e8f3096248d2b1217897e6d6f7716cf114cd367d91f7100a27c2d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.03702052589505911,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1866274,
+ "backend_with_key_registry_bytes": 2084752,
+ "backend_with_datrie_key_index_bytes": 1979986,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999509431920363,
+ "mean_cosine": 0.9999804481517602,
+ "p05_cosine": 0.9999710753229143,
+ "relative_l2_mean": 0.006229599497413254,
+ "relative_l2_p95": 0.0076318861968173845,
+ "relative_l2_max": 0.009914014821093008,
+ "weighted_relative_l2_mean": 0.006165011200244576,
+ "weighted_relative_frobenius": 0.005943315937829566,
+ "weighted_mean_cosine": 0.9999808428943063,
+ "max_relative_norm_change": 0.002319510056391344
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.373232,
+ "warm_p50_ms": 0.1215585,
+ "warm_p95_ms": 0.12658379999999997,
+ "warm_p99_ms": 0.13713232,
+ "samples_ms": [
+ 0.124876,
+ 0.122502,
+ 0.122013,
+ 0.121524,
+ 0.121593,
+ 0.121175,
+ 0.132489,
+ 0.121105,
+ 0.121593,
+ 0.121594,
+ 0.121663,
+ 0.121594,
+ 0.122013,
+ 0.121663,
+ 0.121594,
+ 0.121593,
+ 0.122013,
+ 0.140101,
+ 0.122013,
+ 0.121524,
+ 0.121524,
+ 0.121524,
+ 0.121524,
+ 0.120616,
+ 0.122082,
+ 0.121594,
+ 0.121105,
+ 0.121105,
+ 0.121524,
+ 0.121035,
+ 0.121524,
+ 0.120127,
+ 0.121524,
+ 0.120057,
+ 0.120685,
+ 0.122083,
+ 0.126273,
+ 0.119778,
+ 0.121174,
+ 0.121105
+ ],
+ "warm_rows_per_second": 4211963.786983222
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1733494,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 228,
+ "zlib_blocks": 0,
+ "xor_blocks": 228,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "d37754bf9356e1c82806b733d2f09d6fcd41d5b2725b70c407328b1a5c7fe149",
+ "table_sha256": "c0da2a861a0f3abfe145d9554c6788224135e25d085a74fa3b44f9ae9c3956c4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2033181639853865,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1734747,
+ "backend_with_key_registry_bytes": 1953225,
+ "backend_with_datrie_key_index_bytes": 1848459,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999509431920363,
+ "mean_cosine": 0.9999804481517602,
+ "p05_cosine": 0.9999710753229143,
+ "relative_l2_mean": 0.006229599497413254,
+ "relative_l2_p95": 0.0076318861968173845,
+ "relative_l2_max": 0.009914014821093008,
+ "weighted_relative_l2_mean": 0.006165011200244576,
+ "weighted_relative_frobenius": 0.005943315937829566,
+ "weighted_mean_cosine": 0.9999808428943063,
+ "max_relative_norm_change": 0.002319510056391344
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.004983,
+ "warm_p50_ms": 0.131512,
+ "warm_p95_ms": 0.13451814999999998,
+ "warm_p99_ms": 0.14703703999999998,
+ "samples_ms": [
+ 0.137238,
+ 0.134375,
+ 0.133816,
+ 0.131511,
+ 0.132,
+ 0.132419,
+ 0.132909,
+ 0.132489,
+ 0.131092,
+ 0.131511,
+ 0.130603,
+ 0.130603,
+ 0.132,
+ 0.131512,
+ 0.131512,
+ 0.131511,
+ 0.153302,
+ 0.132978,
+ 0.132,
+ 0.131511,
+ 0.131511,
+ 0.131512,
+ 0.132,
+ 0.131511,
+ 0.131581,
+ 0.131511,
+ 0.132,
+ 0.131512,
+ 0.131023,
+ 0.132,
+ 0.131511,
+ 0.131092,
+ 0.131581,
+ 0.131023,
+ 0.131512,
+ 0.130604,
+ 0.130604,
+ 0.132001,
+ 0.132489,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3893180.850416692
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1733494,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 228,
+ "zlib_blocks": 0,
+ "xor_blocks": 228,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "d37754bf9356e1c82806b733d2f09d6fcd41d5b2725b70c407328b1a5c7fe149",
+ "table_sha256": "c0da2a861a0f3abfe145d9554c6788224135e25d085a74fa3b44f9ae9c3956c4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2549184209201485,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1734746,
+ "backend_with_key_registry_bytes": 1953224,
+ "backend_with_datrie_key_index_bytes": 1848458,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999509431920363,
+ "mean_cosine": 0.9999804481517602,
+ "p05_cosine": 0.9999710753229143,
+ "relative_l2_mean": 0.006229599497413254,
+ "relative_l2_p95": 0.0076318861968173845,
+ "relative_l2_max": 0.009914014821093008,
+ "weighted_relative_l2_mean": 0.006165011200244576,
+ "weighted_relative_frobenius": 0.005943315937829566,
+ "weighted_mean_cosine": 0.9999808428943063,
+ "max_relative_norm_change": 0.002319510056391344
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.943523,
+ "warm_p50_ms": 0.130603,
+ "warm_p95_ms": 0.13261119999999998,
+ "warm_p99_ms": 0.14869398999999997,
+ "samples_ms": [
+ 0.134933,
+ 0.131022,
+ 0.131511,
+ 0.132489,
+ 0.130603,
+ 0.132419,
+ 0.132489,
+ 0.130604,
+ 0.128788,
+ 0.129626,
+ 0.131022,
+ 0.130114,
+ 0.132489,
+ 0.131511,
+ 0.130603,
+ 0.131092,
+ 0.130533,
+ 0.131022,
+ 0.131022,
+ 0.129136,
+ 0.129625,
+ 0.130603,
+ 0.131092,
+ 0.130603,
+ 0.128299,
+ 0.131512,
+ 0.130603,
+ 0.129137,
+ 0.131023,
+ 0.157492,
+ 0.131511,
+ 0.130184,
+ 0.128718,
+ 0.130533,
+ 0.130533,
+ 0.129066,
+ 0.129067,
+ 0.130045,
+ 0.130534,
+ 0.130604
+ ],
+ "warm_rows_per_second": 3920277.4821405327
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1733494,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 228,
+ "zlib_blocks": 0,
+ "xor_blocks": 228,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "d37754bf9356e1c82806b733d2f09d6fcd41d5b2725b70c407328b1a5c7fe149",
+ "table_sha256": "c0da2a861a0f3abfe145d9554c6788224135e25d085a74fa3b44f9ae9c3956c4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22678576596081257,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1734746,
+ "backend_with_key_registry_bytes": 1953224,
+ "backend_with_datrie_key_index_bytes": 1848458,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999509431920363,
+ "mean_cosine": 0.9999804481517602,
+ "p05_cosine": 0.9999710753229143,
+ "relative_l2_mean": 0.006229599497413254,
+ "relative_l2_p95": 0.0076318861968173845,
+ "relative_l2_max": 0.009914014821093008,
+ "weighted_relative_l2_mean": 0.006165011200244576,
+ "weighted_relative_frobenius": 0.005943315937829566,
+ "weighted_mean_cosine": 0.9999808428943063,
+ "max_relative_norm_change": 0.002319510056391344
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.908043,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.1340536,
+ "warm_p99_ms": 0.15713437,
+ "samples_ms": [
+ 0.137238,
+ 0.133886,
+ 0.131023,
+ 0.131092,
+ 0.130673,
+ 0.130603,
+ 0.132489,
+ 0.131581,
+ 0.131023,
+ 0.132,
+ 0.131022,
+ 0.131092,
+ 0.131511,
+ 0.131022,
+ 0.131092,
+ 0.131581,
+ 0.169855,
+ 0.132908,
+ 0.131511,
+ 0.131023,
+ 0.131511,
+ 0.130184,
+ 0.130604,
+ 0.131511,
+ 0.130533,
+ 0.132489,
+ 0.130114,
+ 0.130115,
+ 0.131581,
+ 0.131511,
+ 0.130534,
+ 0.132978,
+ 0.131511,
+ 0.132489,
+ 0.132909,
+ 0.132908,
+ 0.132489,
+ 0.133397,
+ 0.131511,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1185695,
+ "unique_rows": 8773,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 142,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 142,
+ "zlib_blocks": 1,
+ "xor_blocks": 141,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8773
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "e3e907c384b71e729a0aebd53e8aa9533b143e6799234253559671bd06448b98",
+ "table_sha256": "5415c8d5a573c4e941775705e851a7fda8a767b0dc3ec02bbf486f86878b986e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.15289068198762834,
+ "representatives_sha256": "3f27486158efe3a45eeba191fb9fb8aa19da7dd6c9f103e6cb9c27f8bdd505d3",
+ "structurally_merged_rows": 5359,
+ "backend_artifact_bytes": 1186944,
+ "backend_with_key_registry_bytes": 1405422,
+ "backend_with_datrie_key_index_bytes": 1300656,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9898991758915479,
+ "mean_cosine": 0.9985624306039453,
+ "p05_cosine": 0.9927117436485267,
+ "relative_l2_mean": 0.034843087947229484,
+ "relative_l2_p95": 0.1208930519539888,
+ "relative_l2_max": 0.14267755652829675,
+ "weighted_relative_l2_mean": 0.022564555339120232,
+ "weighted_relative_frobenius": 0.011273324565210031,
+ "weighted_mean_cosine": 0.9992034365758994,
+ "max_relative_norm_change": 0.011073758177466862
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900036368177192,
+ "mean_cosine": 0.9985675330928376,
+ "p05_cosine": 0.9926988551751524,
+ "relative_l2_mean": 0.031079395593628233,
+ "relative_l2_p95": 0.12095334734357205,
+ "relative_l2_max": 0.14178901201605482,
+ "weighted_relative_l2_mean": 0.017956041919640616,
+ "weighted_relative_frobenius": 0.009708818886474999,
+ "weighted_mean_cosine": 0.9992134804590077,
+ "max_relative_norm_change": 0.010098292410539783
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.498517,
+ "warm_p50_ms": 0.1310225,
+ "warm_p95_ms": 0.13397345,
+ "warm_p99_ms": 0.1611877,
+ "samples_ms": [
+ 0.133956,
+ 0.131022,
+ 0.132489,
+ 0.132,
+ 0.131092,
+ 0.131022,
+ 0.130603,
+ 0.131022,
+ 0.131582,
+ 0.130604,
+ 0.132001,
+ 0.131022,
+ 0.131511,
+ 0.131022,
+ 0.130603,
+ 0.131022,
+ 0.131022,
+ 0.131022,
+ 0.131023,
+ 0.131023,
+ 0.131092,
+ 0.131023,
+ 0.130115,
+ 0.130184,
+ 0.131092,
+ 0.131512,
+ 0.130115,
+ 0.178375,
+ 0.134305,
+ 0.131093,
+ 0.130604,
+ 0.128718,
+ 0.131092,
+ 0.131023,
+ 0.131093,
+ 0.131582,
+ 0.130534,
+ 0.129626,
+ 0.130115,
+ 0.129626
+ ],
+ "warm_rows_per_second": 3907725.772291019
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1132000,
+ "unique_rows": 8284,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 134,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 134,
+ "zlib_blocks": 2,
+ "xor_blocks": 132,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8284
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "39abe9e020e0fa36b42308058b67c647fcb32b21d834148ef6c13acdb7c9e16a",
+ "table_sha256": "f5f5f8f557e088dbbaf135b2e41557587a0dc9ec9b112622b3781dd14c48e1de",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.16384774213656783,
+ "representatives_sha256": "a53571cebfc88234b2ccaf0e36f5d95fe432e78310acd13043d4bbb7b0a87159",
+ "structurally_merged_rows": 5848,
+ "backend_artifact_bytes": 1133249,
+ "backend_with_key_registry_bytes": 1351727,
+ "backend_with_datrie_key_index_bytes": 1246961,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800609363681603,
+ "mean_cosine": 0.9980835864851856,
+ "p05_cosine": 0.9907192079250814,
+ "relative_l2_mean": 0.040370998729445585,
+ "relative_l2_p95": 0.13648493463487885,
+ "relative_l2_max": 0.20040302515574998,
+ "weighted_relative_l2_mean": 0.025439388941910904,
+ "weighted_relative_frobenius": 0.012882211183388767,
+ "weighted_mean_cosine": 0.9989536429345286,
+ "max_relative_norm_change": 0.011073758177466862
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9801845112059807,
+ "mean_cosine": 0.9980874938018807,
+ "p05_cosine": 0.9906959908080297,
+ "relative_l2_mean": 0.036826332511338385,
+ "relative_l2_p95": 0.13654300595165328,
+ "relative_l2_max": 0.19979866428782744,
+ "weighted_relative_l2_mean": 0.020946080485034813,
+ "weighted_relative_frobenius": 0.011561476750624822,
+ "weighted_mean_cosine": 0.9989629229526767,
+ "max_relative_norm_change": 0.010098292410539783
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.51451,
+ "warm_p50_ms": 0.1315115,
+ "warm_p95_ms": 0.134794,
+ "warm_p99_ms": 0.15733105999999997,
+ "samples_ms": [
+ 0.134794,
+ 0.134794,
+ 0.132978,
+ 0.131022,
+ 0.132419,
+ 0.131581,
+ 0.130604,
+ 0.132001,
+ 0.132978,
+ 0.131022,
+ 0.132489,
+ 0.131093,
+ 0.132489,
+ 0.132,
+ 0.130603,
+ 0.130603,
+ 0.130114,
+ 0.132,
+ 0.132001,
+ 0.131093,
+ 0.130115,
+ 0.131093,
+ 0.132,
+ 0.132908,
+ 0.131022,
+ 0.17174,
+ 0.131511,
+ 0.132,
+ 0.130604,
+ 0.131512,
+ 0.131023,
+ 0.130534,
+ 0.129626,
+ 0.131511,
+ 0.131022,
+ 0.132419,
+ 0.130603,
+ 0.132,
+ 0.132001,
+ 0.130534
+ ],
+ "warm_rows_per_second": 3893195.6520912615
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1123444,
+ "unique_rows": 8212,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 133,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 133,
+ "zlib_blocks": 2,
+ "xor_blocks": 131,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8212
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "3b62dd395678f17f8a18c39bbcc08a79b7f779d435de694eb18ed31f8a82b284",
+ "table_sha256": "60cfc36ef9a1234b18bd362da8210f476392bbb9a93d239adcbd71adf575f34f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14480039989575744,
+ "representatives_sha256": "2056b5595db4696ab7f357cc3c183ee0db15737e8ff61829f7970ec4b4b78509",
+ "structurally_merged_rows": 5920,
+ "backend_artifact_bytes": 1124693,
+ "backend_with_key_registry_bytes": 1343171,
+ "backend_with_datrie_key_index_bytes": 1238405,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.970533398984063,
+ "mean_cosine": 0.9979517868279669,
+ "p05_cosine": 0.9903524458799131,
+ "relative_l2_mean": 0.04152405653797273,
+ "relative_l2_p95": 0.1389526598971901,
+ "relative_l2_max": 0.24293060424031895,
+ "weighted_relative_l2_mean": 0.02601064268249304,
+ "weighted_relative_frobenius": 0.01299657603660586,
+ "weighted_mean_cosine": 0.9988883460806876,
+ "max_relative_norm_change": 0.011073758177466862
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9706767662198073,
+ "mean_cosine": 0.9979554209545061,
+ "p05_cosine": 0.9903252160808494,
+ "relative_l2_mean": 0.03801150163211964,
+ "relative_l2_p95": 0.1392648222596999,
+ "relative_l2_max": 0.24320645540282654,
+ "weighted_relative_l2_mean": 0.02153324297460118,
+ "weighted_relative_frobenius": 0.01168888666741782,
+ "weighted_mean_cosine": 0.9988974907535848,
+ "max_relative_norm_change": 0.010098292410539783
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.394802,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.137595,
+ "warm_p99_ms": 0.16243227999999998,
+ "samples_ms": [
+ 0.135702,
+ 0.133885,
+ 0.173137,
+ 0.137169,
+ 0.132489,
+ 0.131162,
+ 0.131511,
+ 0.132,
+ 0.132,
+ 0.131092,
+ 0.136749,
+ 0.132489,
+ 0.132,
+ 0.131512,
+ 0.132978,
+ 0.131512,
+ 0.131092,
+ 0.132489,
+ 0.132,
+ 0.131582,
+ 0.131022,
+ 0.132,
+ 0.130603,
+ 0.130603,
+ 0.132001,
+ 0.132489,
+ 0.132908,
+ 0.130604,
+ 0.130534,
+ 0.131511,
+ 0.131511,
+ 0.132,
+ 0.145689,
+ 0.131022,
+ 0.132,
+ 0.130114,
+ 0.131022,
+ 0.132489,
+ 0.131092,
+ 0.133467
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1121274,
+ "unique_rows": 8192,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 133,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 132,
+ "zlib_blocks": 3,
+ "xor_blocks": 129,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8192
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "4347d36d220d562717012c8dc4185b21cde760ad2e4dceb8e1551a527efb99fd",
+ "table_sha256": "b407aa718cdc94b129f0ed927230512af321dc1c7441189ef0fcff6e9852e3f2",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14451132505200803,
+ "representatives_sha256": "4295e5aa8792a9f176f49ca4cc2d2bd875799074b16a719944247ad4144fbf98",
+ "structurally_merged_rows": 5940,
+ "backend_artifact_bytes": 1122523,
+ "backend_with_key_registry_bytes": 1341001,
+ "backend_with_datrie_key_index_bytes": 1236235,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9516396125089167,
+ "mean_cosine": 0.9978860315050953,
+ "p05_cosine": 0.9902062405450521,
+ "relative_l2_mean": 0.0419809939042737,
+ "relative_l2_p95": 0.13987874549353838,
+ "relative_l2_max": 0.31107273743945374,
+ "weighted_relative_l2_mean": 0.026268627807504365,
+ "weighted_relative_frobenius": 0.013264231963570525,
+ "weighted_mean_cosine": 0.9988511739345237,
+ "max_relative_norm_change": 0.011073758177466862
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9515900745539979,
+ "mean_cosine": 0.997889624367305,
+ "p05_cosine": 0.9901928744310519,
+ "relative_l2_mean": 0.038477430801658524,
+ "relative_l2_p95": 0.14004325529443465,
+ "relative_l2_max": 0.31121001199064846,
+ "weighted_relative_l2_mean": 0.021795814636628608,
+ "weighted_relative_frobenius": 0.011985214831303536,
+ "weighted_mean_cosine": 0.9988603101202684,
+ "max_relative_norm_change": 0.010098292410539783
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.26378,
+ "warm_p50_ms": 0.132001,
+ "warm_p95_ms": 0.16930295,
+ "warm_p99_ms": 0.17228918999999998,
+ "samples_ms": [
+ 0.136819,
+ 0.131511,
+ 0.132,
+ 0.132908,
+ 0.132001,
+ 0.138635,
+ 0.132489,
+ 0.132908,
+ 0.13242,
+ 0.132489,
+ 0.131511,
+ 0.131581,
+ 0.132,
+ 0.131023,
+ 0.132489,
+ 0.130603,
+ 0.132,
+ 0.131022,
+ 0.132001,
+ 0.166572,
+ 0.133467,
+ 0.130115,
+ 0.130185,
+ 0.132001,
+ 0.131093,
+ 0.130045,
+ 0.131022,
+ 0.131581,
+ 0.129625,
+ 0.131022,
+ 0.129625,
+ 0.132489,
+ 0.164686,
+ 0.174114,
+ 0.167409,
+ 0.169435,
+ 0.167479,
+ 0.167968,
+ 0.167898,
+ 0.169296
+ ],
+ "warm_rows_per_second": 3878758.4942538314
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1119725,
+ "unique_rows": 8177,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 132,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 132,
+ "zlib_blocks": 3,
+ "xor_blocks": 129,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8177
+ },
+ "source_float32_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127",
+ "decoded_float32_sha256": "afc8defabf0539639b6ab778db34a0733084b808149cf4aad31bb6c119dcb051",
+ "table_sha256": "58f13c4ccf9ee34518a3d1139ffdae723e637f0e3aef368b40acb212c717c130",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.143010085914284,
+ "representatives_sha256": "75f0985092dacb57da3675dd32889a5eafa52457b1b4315ddd5aeffbad786241",
+ "structurally_merged_rows": 5955,
+ "backend_artifact_bytes": 1120974,
+ "backend_with_key_registry_bytes": 1339452,
+ "backend_with_datrie_key_index_bytes": 1234686,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9069200513135023,
+ "mean_cosine": 0.9977608335117717,
+ "p05_cosine": 0.9900603464531534,
+ "relative_l2_mean": 0.042584448685336854,
+ "relative_l2_p95": 0.14098517608415817,
+ "relative_l2_max": 0.4313462450471783,
+ "weighted_relative_l2_mean": 0.026923009724454265,
+ "weighted_relative_frobenius": 0.016679475285300137,
+ "weighted_mean_cosine": 0.9987163534494873,
+ "max_relative_norm_change": 0.011073758177466862
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.906865860939273,
+ "mean_cosine": 0.9977643453812376,
+ "p05_cosine": 0.99007159701529,
+ "relative_l2_mean": 0.039087591362981705,
+ "relative_l2_p95": 0.14104134847131763,
+ "relative_l2_max": 0.43140187157121745,
+ "weighted_relative_l2_mean": 0.022458199434176405,
+ "weighted_relative_frobenius": 0.015683609464981013,
+ "weighted_mean_cosine": 0.9987253930745409,
+ "max_relative_norm_change": 0.010098292410539783
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.134154,
+ "warm_p50_ms": 0.13318750000000001,
+ "warm_p95_ms": 0.1732557,
+ "warm_p99_ms": 0.17811021,
+ "samples_ms": [
+ 0.135283,
+ 0.134375,
+ 0.132908,
+ 0.132001,
+ 0.133397,
+ 0.132978,
+ 0.132,
+ 0.132489,
+ 0.135422,
+ 0.131581,
+ 0.131511,
+ 0.131511,
+ 0.130045,
+ 0.131023,
+ 0.131512,
+ 0.132001,
+ 0.132489,
+ 0.132,
+ 0.150927,
+ 0.130603,
+ 0.130603,
+ 0.131511,
+ 0.130603,
+ 0.132978,
+ 0.131581,
+ 0.172718,
+ 0.179772,
+ 0.161823,
+ 0.175511,
+ 0.168388,
+ 0.158051,
+ 0.169855,
+ 0.162731,
+ 0.173137,
+ 0.161333,
+ 0.159937,
+ 0.160845,
+ 0.168387,
+ 0.170343,
+ 0.170762
+ ],
+ "warm_rows_per_second": 3844204.5987799154
+ }
+ }
+ ],
+ "baseline_quality": {},
+ "training_row_sha256": "f52f97f6be5a5c643586efdc377a6b8e236e245f915abcc4e9dbfbb0558cd127"
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 7,
+ "rows": 14132,
+ "dim": 128,
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "source_float32_bytes": 7235584,
+ "key_registry_bytes": 218478,
+ "unchanged_backbone_artifact_bytes": 0,
+ "alternative_datrie_key_index_bytes": 113712,
+ "query_rows": 512,
+ "query_ids_sha256": "48cdbdeffaad5c50239fec7524368c85bd12fc21eefeb42effaa6c7c5ff0c323",
+ "frequency_trace_sha256": "fc8afe9c1f9a53dfd826267bdad324151b311d81a4be4e4834bd129ccf363886",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 512,
+ "storage": {
+ "artifact_bytes": 7235584,
+ "unique_rows": 14132,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "table_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.041674753185361624,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 7236389,
+ "backend_with_key_registry_bytes": 7454867,
+ "backend_with_datrie_key_index_bytes": 7350101,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.376166,
+ "warm_p50_ms": 0.0600635,
+ "warm_p95_ms": 0.06112845,
+ "warm_p99_ms": 0.0615027,
+ "samples_ms": [
+ 0.061041,
+ 0.06146,
+ 0.06153,
+ 0.060553,
+ 0.059574,
+ 0.061041,
+ 0.060552,
+ 0.060552,
+ 0.059086,
+ 0.060623,
+ 0.058667,
+ 0.059644,
+ 0.060552,
+ 0.060552,
+ 0.060552,
+ 0.058667,
+ 0.060902,
+ 0.058736,
+ 0.059574,
+ 0.059226,
+ 0.060064,
+ 0.059645,
+ 0.059155,
+ 0.058666,
+ 0.060064,
+ 0.059645,
+ 0.061041,
+ 0.059994,
+ 0.060133,
+ 0.060063,
+ 0.059645,
+ 0.059575,
+ 0.061111,
+ 0.060064,
+ 0.060133,
+ 0.059574,
+ 0.059155,
+ 0.059575,
+ 0.058667,
+ 0.060622
+ ],
+ "warm_rows_per_second": 8524311.770043371
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 6811178,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 884,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 884,
+ "zlib_blocks": 50,
+ "xor_blocks": 834,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "table_sha256": "060c766f7a8ee8758960728799dc557117c3f4f33b659a2ea00bdf54a1a81e46",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6181737699080259,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6812387,
+ "backend_with_key_registry_bytes": 7030865,
+ "backend_with_datrie_key_index_bytes": 6926099,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 13.007952,
+ "warm_p50_ms": 0.071937,
+ "warm_p95_ms": 0.07675585,
+ "warm_p99_ms": 0.08362119,
+ "samples_ms": [
+ 0.078083,
+ 0.074801,
+ 0.0748,
+ 0.074311,
+ 0.0748,
+ 0.07473,
+ 0.074241,
+ 0.072426,
+ 0.072425,
+ 0.071866,
+ 0.076686,
+ 0.074312,
+ 0.072845,
+ 0.071936,
+ 0.070958,
+ 0.070959,
+ 0.072355,
+ 0.071517,
+ 0.071448,
+ 0.071447,
+ 0.072425,
+ 0.071937,
+ 0.071448,
+ 0.072844,
+ 0.071028,
+ 0.070889,
+ 0.070959,
+ 0.070539,
+ 0.087162,
+ 0.071448,
+ 0.071518,
+ 0.072426,
+ 0.071866,
+ 0.071448,
+ 0.071937,
+ 0.071378,
+ 0.071448,
+ 0.071448,
+ 0.073264,
+ 0.073822
+ ],
+ "warm_rows_per_second": 7117338.7825458385
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 6811178,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 884,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 884,
+ "zlib_blocks": 50,
+ "xor_blocks": 834,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "table_sha256": "060c766f7a8ee8758960728799dc557117c3f4f33b659a2ea00bdf54a1a81e46",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6556007731705904,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6812386,
+ "backend_with_key_registry_bytes": 7030864,
+ "backend_with_datrie_key_index_bytes": 6926098,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 13.043362,
+ "warm_p50_ms": 0.071448,
+ "warm_p95_ms": 0.07643004999999999,
+ "warm_p99_ms": 0.09962168999999998,
+ "samples_ms": [
+ 0.076196,
+ 0.075708,
+ 0.074311,
+ 0.074311,
+ 0.074311,
+ 0.070469,
+ 0.071378,
+ 0.072356,
+ 0.071937,
+ 0.071378,
+ 0.071448,
+ 0.071448,
+ 0.071867,
+ 0.111606,
+ 0.080877,
+ 0.071937,
+ 0.071866,
+ 0.071936,
+ 0.07054,
+ 0.071377,
+ 0.071378,
+ 0.069981,
+ 0.071378,
+ 0.071937,
+ 0.071517,
+ 0.071029,
+ 0.071448,
+ 0.07054,
+ 0.071029,
+ 0.071029,
+ 0.071517,
+ 0.071448,
+ 0.070469,
+ 0.071378,
+ 0.071378,
+ 0.071378,
+ 0.069492,
+ 0.071447,
+ 0.071448,
+ 0.071029
+ ],
+ "warm_rows_per_second": 7166050.834173105
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 6811178,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 884,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 884,
+ "zlib_blocks": 50,
+ "xor_blocks": 834,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "table_sha256": "060c766f7a8ee8758960728799dc557117c3f4f33b659a2ea00bdf54a1a81e46",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6131828369107097,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6812386,
+ "backend_with_key_registry_bytes": 7030864,
+ "backend_with_datrie_key_index_bytes": 6926098,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.974847,
+ "warm_p50_ms": 0.0720065,
+ "warm_p95_ms": 0.0767549,
+ "warm_p99_ms": 0.09333482999999998,
+ "samples_ms": [
+ 0.076127,
+ 0.075219,
+ 0.076127,
+ 0.072845,
+ 0.073823,
+ 0.073823,
+ 0.078083,
+ 0.076685,
+ 0.076197,
+ 0.075219,
+ 0.075289,
+ 0.0748,
+ 0.073333,
+ 0.071448,
+ 0.071937,
+ 0.071028,
+ 0.071448,
+ 0.07054,
+ 0.072006,
+ 0.071867,
+ 0.103086,
+ 0.071866,
+ 0.071448,
+ 0.069492,
+ 0.073333,
+ 0.073752,
+ 0.073892,
+ 0.072495,
+ 0.071029,
+ 0.072007,
+ 0.071028,
+ 0.071378,
+ 0.070889,
+ 0.071447,
+ 0.0704,
+ 0.071447,
+ 0.071448,
+ 0.071867,
+ 0.073333,
+ 0.071447
+ ],
+ "warm_rows_per_second": 7110469.193753342
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4292179,
+ "unique_rows": 8713,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 545,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 545,
+ "zlib_blocks": 323,
+ "xor_blocks": 222,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8713
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "2d03ab5abc18e43848f97f0c1fae076c613a04b95b258ddfaa1a91582c6e0853",
+ "table_sha256": "23201c1783f77c2e154234ee08dba949bdb115fe2e1e576121c728114f63ef07",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.43180519505403936,
+ "representatives_sha256": "1d7f2f8f770c4e64c61188f000dff74864d006900e9d8b82d4e1fe53ca787c27",
+ "structurally_merged_rows": 5419,
+ "backend_artifact_bytes": 4293386,
+ "backend_with_key_registry_bytes": 4511864,
+ "backend_with_datrie_key_index_bytes": 4407098,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900073902714597,
+ "mean_cosine": 0.9986074490701802,
+ "p05_cosine": 0.9931238826886435,
+ "relative_l2_mean": 0.03078177462642599,
+ "relative_l2_p95": 0.11744627815304662,
+ "relative_l2_max": 0.1422066988799605,
+ "weighted_relative_l2_mean": 0.017772639499589293,
+ "weighted_relative_frobenius": 0.009467831744684692,
+ "weighted_mean_cosine": 0.9992362196570699,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900073902714597,
+ "mean_cosine": 0.9986074490701802,
+ "p05_cosine": 0.9931238826886435,
+ "relative_l2_mean": 0.03078177462642599,
+ "relative_l2_p95": 0.11744627815304662,
+ "relative_l2_max": 0.1422066988799605,
+ "weighted_relative_l2_mean": 0.017772639499589293,
+ "weighted_relative_frobenius": 0.009467831744684692,
+ "weighted_mean_cosine": 0.9992362196570699,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.992051,
+ "warm_p50_ms": 0.070051,
+ "warm_p95_ms": 0.07428989999999999,
+ "warm_p99_ms": 0.08880918999999998,
+ "samples_ms": [
+ 0.075219,
+ 0.072426,
+ 0.071448,
+ 0.073333,
+ 0.071936,
+ 0.073333,
+ 0.071937,
+ 0.069562,
+ 0.070051,
+ 0.069562,
+ 0.069492,
+ 0.068584,
+ 0.069073,
+ 0.069003,
+ 0.097498,
+ 0.072356,
+ 0.069562,
+ 0.073333,
+ 0.070959,
+ 0.07047,
+ 0.070051,
+ 0.070889,
+ 0.069981,
+ 0.069562,
+ 0.070051,
+ 0.07005,
+ 0.069562,
+ 0.073682,
+ 0.074241,
+ 0.071936,
+ 0.070051,
+ 0.069981,
+ 0.069073,
+ 0.069562,
+ 0.071518,
+ 0.072006,
+ 0.07047,
+ 0.069562,
+ 0.069981,
+ 0.069562
+ ],
+ "warm_rows_per_second": 7308960.614409502
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4084875,
+ "unique_rows": 8276,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 518,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 518,
+ "zlib_blocks": 349,
+ "xor_blocks": 169,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8276
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "5916738c9930feda7070e45357cb7ba053200fced3550e1ca30fe371e1c697d4",
+ "table_sha256": "09df40dfc6fbab1b1e4094cf6d88f9cf35be5684fda4745ef9231c6a7c961073",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3878235809970647,
+ "representatives_sha256": "72bd69c0991c8f59ba0fea41d746a21d07d320e00e85dbf4053e970b62972166",
+ "structurally_merged_rows": 5856,
+ "backend_artifact_bytes": 4086082,
+ "backend_with_key_registry_bytes": 4304560,
+ "backend_with_datrie_key_index_bytes": 4199794,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800632074662583,
+ "mean_cosine": 0.9981682497466008,
+ "p05_cosine": 0.9912024672313878,
+ "relative_l2_mean": 0.03598248125019667,
+ "relative_l2_p95": 0.13290198548669502,
+ "relative_l2_max": 0.2008804026085917,
+ "weighted_relative_l2_mean": 0.020372717570488277,
+ "weighted_relative_frobenius": 0.009890047376844057,
+ "weighted_mean_cosine": 0.9990166368590908,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800632074662583,
+ "mean_cosine": 0.9981682497466008,
+ "p05_cosine": 0.9912024672313878,
+ "relative_l2_mean": 0.03598248125019667,
+ "relative_l2_p95": 0.13290198548669502,
+ "relative_l2_max": 0.2008804026085917,
+ "weighted_relative_l2_mean": 0.020372717570488277,
+ "weighted_relative_frobenius": 0.009890047376844057,
+ "weighted_mean_cosine": 0.9990166368590908,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.509378,
+ "warm_p50_ms": 0.071867,
+ "warm_p95_ms": 0.0748908,
+ "warm_p99_ms": 0.09685213999999998,
+ "samples_ms": [
+ 0.0748,
+ 0.072356,
+ 0.076616,
+ 0.073822,
+ 0.073752,
+ 0.073403,
+ 0.072495,
+ 0.071937,
+ 0.072356,
+ 0.074242,
+ 0.070889,
+ 0.070889,
+ 0.070959,
+ 0.071448,
+ 0.071029,
+ 0.068654,
+ 0.072845,
+ 0.10979,
+ 0.070959,
+ 0.069143,
+ 0.071448,
+ 0.072495,
+ 0.070121,
+ 0.072356,
+ 0.073753,
+ 0.073823,
+ 0.071867,
+ 0.07047,
+ 0.072844,
+ 0.069492,
+ 0.07005,
+ 0.073404,
+ 0.069562,
+ 0.07054,
+ 0.073334,
+ 0.071517,
+ 0.071378,
+ 0.071867,
+ 0.068654,
+ 0.070959
+ ],
+ "warm_rows_per_second": 7124271.223231804
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4048616,
+ "unique_rows": 8201,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 513,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 513,
+ "zlib_blocks": 347,
+ "xor_blocks": 166,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8201
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "5f22f7b5a51b64103430e36ba1d0bd54077ea2f16d3d4bb871f51692d993995b",
+ "table_sha256": "30cde4f7eec3681cc425eaa499f2a58de9ef5c1aba06a07878c0e60893e2474a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.36942198988981545,
+ "representatives_sha256": "c3f844016d2aa27522f40276a4692138cf6475e2ba2a582770bdaa8120ca3584",
+ "structurally_merged_rows": 5931,
+ "backend_artifact_bytes": 4049823,
+ "backend_with_key_registry_bytes": 4268301,
+ "backend_with_datrie_key_index_bytes": 4163535,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9702642593028836,
+ "mean_cosine": 0.9980366950229913,
+ "p05_cosine": 0.9907400526112273,
+ "relative_l2_mean": 0.03717392426551632,
+ "relative_l2_p95": 0.13639079380091268,
+ "relative_l2_max": 0.24371533510351767,
+ "weighted_relative_l2_mean": 0.021110074541361344,
+ "weighted_relative_frobenius": 0.011079169105965923,
+ "weighted_mean_cosine": 0.9989359764157608,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9702642593028836,
+ "mean_cosine": 0.9980366950229913,
+ "p05_cosine": 0.9907400526112273,
+ "relative_l2_mean": 0.03717392426551632,
+ "relative_l2_p95": 0.13639079380091268,
+ "relative_l2_max": 0.24371533510351767,
+ "weighted_relative_l2_mean": 0.021110074541361344,
+ "weighted_relative_frobenius": 0.011079169105965923,
+ "weighted_mean_cosine": 0.9989359764157608,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.472502,
+ "warm_p50_ms": 0.071238,
+ "warm_p95_ms": 0.0781109,
+ "warm_p99_ms": 0.09069825999999999,
+ "samples_ms": [
+ 0.077664,
+ 0.07515,
+ 0.073822,
+ 0.073822,
+ 0.074311,
+ 0.074241,
+ 0.078641,
+ 0.078083,
+ 0.076616,
+ 0.075708,
+ 0.073334,
+ 0.073333,
+ 0.073333,
+ 0.073333,
+ 0.070958,
+ 0.070959,
+ 0.069143,
+ 0.071099,
+ 0.070958,
+ 0.069003,
+ 0.071377,
+ 0.0704,
+ 0.070958,
+ 0.069143,
+ 0.069003,
+ 0.069073,
+ 0.069073,
+ 0.069492,
+ 0.070889,
+ 0.069981,
+ 0.07047,
+ 0.068654,
+ 0.071377,
+ 0.071448,
+ 0.098407,
+ 0.076127,
+ 0.071029,
+ 0.071448,
+ 0.069562,
+ 0.07047
+ ],
+ "warm_rows_per_second": 7187175.38392431
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4040206,
+ "unique_rows": 8184,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 512,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 512,
+ "zlib_blocks": 347,
+ "xor_blocks": 165,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8184
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "0a25021ff9755d6ad9e840fc6b37e339f7a014b5c331461e8c1dcf795e1e6faa",
+ "table_sha256": "d72869d2e1b2e4da739889e6b88f84e95469d18ea8870256037ff8ab9036f8c1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3716650858987123,
+ "representatives_sha256": "f3d6c5e68e2b580f8aad1c18eb930cf5258ab8e22f7c62d2c60eef9f4fade95b",
+ "structurally_merged_rows": 5948,
+ "backend_artifact_bytes": 4041413,
+ "backend_with_key_registry_bytes": 4259891,
+ "backend_with_datrie_key_index_bytes": 4155125,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9525588782396796,
+ "mean_cosine": 0.9979672603102632,
+ "p05_cosine": 0.9905287290957443,
+ "relative_l2_mean": 0.03764276506627799,
+ "relative_l2_p95": 0.13752160879166392,
+ "relative_l2_max": 0.30795876807260114,
+ "weighted_relative_l2_mean": 0.02143238962795817,
+ "weighted_relative_frobenius": 0.011613642395987852,
+ "weighted_mean_cosine": 0.998888214039435,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9525588782396796,
+ "mean_cosine": 0.9979672603102632,
+ "p05_cosine": 0.9905287290957443,
+ "relative_l2_mean": 0.03764276506627799,
+ "relative_l2_p95": 0.13752160879166392,
+ "relative_l2_max": 0.30795876807260114,
+ "weighted_relative_l2_mean": 0.02143238962795817,
+ "weighted_relative_frobenius": 0.011613642395987852,
+ "weighted_mean_cosine": 0.998888214039435,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.799847,
+ "warm_p50_ms": 0.069632,
+ "warm_p95_ms": 0.07457274999999998,
+ "warm_p99_ms": 0.09271976,
+ "samples_ms": [
+ 0.080876,
+ 0.074241,
+ 0.073822,
+ 0.073822,
+ 0.071517,
+ 0.071448,
+ 0.071937,
+ 0.071518,
+ 0.071936,
+ 0.071518,
+ 0.071518,
+ 0.073333,
+ 0.071028,
+ 0.071518,
+ 0.068585,
+ 0.068584,
+ 0.070121,
+ 0.068584,
+ 0.069981,
+ 0.069073,
+ 0.069492,
+ 0.100292,
+ 0.070121,
+ 0.069562,
+ 0.067188,
+ 0.069632,
+ 0.071378,
+ 0.068584,
+ 0.069143,
+ 0.067117,
+ 0.069562,
+ 0.069143,
+ 0.068096,
+ 0.069632,
+ 0.069073,
+ 0.068515,
+ 0.068235,
+ 0.067607,
+ 0.069143,
+ 0.068584
+ ],
+ "warm_rows_per_second": 7352941.176470588
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4041132,
+ "unique_rows": 8184,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 512,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 512,
+ "zlib_blocks": 351,
+ "xor_blocks": 161,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8184
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "98ea4d2f169a39f18f9b142b2ee0f1c3d79c8f9e3707ae76cac857f39ff8023b",
+ "table_sha256": "781e5ddb46a92bd848dc486bfb3d94d6eef253f1674f0a87584e114cd9483b9a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3834197339601815,
+ "representatives_sha256": "a9165656a314128de8ba2f8a95e95db533045e89eeda35bd848291a47b1ac5ab",
+ "structurally_merged_rows": 5948,
+ "backend_artifact_bytes": 4042339,
+ "backend_with_key_registry_bytes": 4260817,
+ "backend_with_datrie_key_index_bytes": 4156051,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9013500902532418,
+ "mean_cosine": 0.9978668806757741,
+ "p05_cosine": 0.9904035009451494,
+ "relative_l2_mean": 0.03808442010560562,
+ "relative_l2_p95": 0.1386711610756497,
+ "relative_l2_max": 0.4454053094763166,
+ "weighted_relative_l2_mean": 0.021717454181962366,
+ "weighted_relative_frobenius": 0.012749720608418173,
+ "weighted_mean_cosine": 0.9988267853378323,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9013500902532418,
+ "mean_cosine": 0.9978668806757741,
+ "p05_cosine": 0.9904035009451494,
+ "relative_l2_mean": 0.03808442010560562,
+ "relative_l2_p95": 0.1386711610756497,
+ "relative_l2_max": 0.4454053094763166,
+ "weighted_relative_l2_mean": 0.021717454181962366,
+ "weighted_relative_frobenius": 0.012749720608418173,
+ "weighted_mean_cosine": 0.9988267853378323,
+ "max_relative_norm_change": 0.010097972516154134
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.586971,
+ "warm_p50_ms": 0.0695965,
+ "warm_p95_ms": 0.07293929999999998,
+ "warm_p99_ms": 0.07592355,
+ "samples_ms": [
+ 0.076686,
+ 0.074731,
+ 0.072845,
+ 0.071937,
+ 0.071447,
+ 0.071867,
+ 0.069562,
+ 0.069632,
+ 0.069981,
+ 0.07054,
+ 0.069143,
+ 0.070051,
+ 0.071447,
+ 0.070051,
+ 0.068515,
+ 0.069632,
+ 0.069562,
+ 0.069492,
+ 0.070051,
+ 0.068584,
+ 0.069073,
+ 0.069631,
+ 0.069562,
+ 0.069073,
+ 0.069074,
+ 0.070051,
+ 0.070051,
+ 0.069911,
+ 0.070959,
+ 0.068723,
+ 0.069492,
+ 0.068584,
+ 0.069143,
+ 0.069562,
+ 0.069562,
+ 0.069143,
+ 0.069562,
+ 0.069562,
+ 0.069562,
+ 0.069632
+ ],
+ "warm_rows_per_second": 7356691.787661735
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 3617792,
+ "unique_rows": 14132,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "a6f992eab989830921c022ff3d6f88caceeb1eeebe107b7ccc88caeb09e9ca68",
+ "table_sha256": "c09228c734adcadf26af9e437aa54e9dbbc0d275c3c2a829e871216873acebd6",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.04507679585367441,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3618598,
+ "backend_with_key_registry_bytes": 3837076,
+ "backend_with_datrie_key_index_bytes": 3732310,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999640630156,
+ "mean_cosine": 0.9999999789175185,
+ "p05_cosine": 0.9999999730721181,
+ "relative_l2_mean": 0.0002070676065899127,
+ "relative_l2_p95": 0.00023440921164902042,
+ "relative_l2_max": 0.00027016899099782454,
+ "weighted_relative_l2_mean": 0.0002072185978470986,
+ "weighted_relative_frobenius": 0.00020769691233671656,
+ "weighted_mean_cosine": 0.9999999788781782,
+ "max_relative_norm_change": 0.00012017913339375236
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.50195,
+ "warm_p50_ms": 0.22,
+ "warm_p95_ms": 0.22368454999999998,
+ "warm_p99_ms": 0.24720571999999996,
+ "samples_ms": [
+ 0.223283,
+ 0.222375,
+ 0.22,
+ 0.220489,
+ 0.220908,
+ 0.22049,
+ 0.220978,
+ 0.219511,
+ 0.22,
+ 0.257366,
+ 0.223283,
+ 0.22,
+ 0.219511,
+ 0.22007,
+ 0.219582,
+ 0.221467,
+ 0.219581,
+ 0.220419,
+ 0.219581,
+ 0.219162,
+ 0.219511,
+ 0.218533,
+ 0.219511,
+ 0.219511,
+ 0.219581,
+ 0.219022,
+ 0.231314,
+ 0.21993,
+ 0.221327,
+ 0.22042,
+ 0.21993,
+ 0.21993,
+ 0.221397,
+ 0.220839,
+ 0.220489,
+ 0.219023,
+ 0.220419,
+ 0.219022,
+ 0.220978,
+ 0.219511
+ ],
+ "warm_rows_per_second": 2327272.727272727
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 3406703,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 442,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 442,
+ "zlib_blocks": 0,
+ "xor_blocks": 442,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "a6f992eab989830921c022ff3d6f88caceeb1eeebe107b7ccc88caeb09e9ca68",
+ "table_sha256": "bbaaa5100062ad8592fd8c4f376ec4ec38424253a6c1bee0d0f21c18ec923b04",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.36845126608386636,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3407912,
+ "backend_with_key_registry_bytes": 3626390,
+ "backend_with_datrie_key_index_bytes": 3521624,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999640630156,
+ "mean_cosine": 0.9999999789175185,
+ "p05_cosine": 0.9999999730721181,
+ "relative_l2_mean": 0.0002070676065899127,
+ "relative_l2_p95": 0.00023440921164902042,
+ "relative_l2_max": 0.00027016899099782454,
+ "weighted_relative_l2_mean": 0.0002072185978470986,
+ "weighted_relative_frobenius": 0.00020769691233671656,
+ "weighted_mean_cosine": 0.9999999788781782,
+ "max_relative_norm_change": 0.00012017913339375236
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.788172,
+ "warm_p50_ms": 0.22901,
+ "warm_p95_ms": 0.24352324999999997,
+ "warm_p99_ms": 0.26475615,
+ "samples_ms": [
+ 0.232292,
+ 0.229498,
+ 0.229429,
+ 0.229988,
+ 0.230406,
+ 0.22901,
+ 0.229918,
+ 0.228521,
+ 0.228521,
+ 0.229499,
+ 0.228521,
+ 0.229079,
+ 0.22901,
+ 0.228032,
+ 0.242769,
+ 0.229918,
+ 0.227613,
+ 0.22859,
+ 0.228102,
+ 0.22894,
+ 0.228521,
+ 0.228032,
+ 0.228521,
+ 0.227962,
+ 0.269169,
+ 0.257854,
+ 0.231384,
+ 0.229499,
+ 0.228102,
+ 0.22901,
+ 0.22859,
+ 0.241721,
+ 0.229848,
+ 0.228521,
+ 0.227054,
+ 0.227473,
+ 0.230825,
+ 0.232781,
+ 0.230476,
+ 0.22901
+ ],
+ "warm_rows_per_second": 2235710.2309942795
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 3406703,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 442,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 442,
+ "zlib_blocks": 0,
+ "xor_blocks": 442,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "a6f992eab989830921c022ff3d6f88caceeb1eeebe107b7ccc88caeb09e9ca68",
+ "table_sha256": "bbaaa5100062ad8592fd8c4f376ec4ec38424253a6c1bee0d0f21c18ec923b04",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4048690621275455,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3407911,
+ "backend_with_key_registry_bytes": 3626389,
+ "backend_with_datrie_key_index_bytes": 3521623,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999640630156,
+ "mean_cosine": 0.9999999789175185,
+ "p05_cosine": 0.9999999730721181,
+ "relative_l2_mean": 0.0002070676065899127,
+ "relative_l2_p95": 0.00023440921164902042,
+ "relative_l2_max": 0.00027016899099782454,
+ "weighted_relative_l2_mean": 0.0002072185978470986,
+ "weighted_relative_frobenius": 0.00020769691233671656,
+ "weighted_mean_cosine": 0.9999999788781782,
+ "max_relative_norm_change": 0.00012017913339375236
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.780699,
+ "warm_p50_ms": 0.22901,
+ "warm_p95_ms": 0.23269805,
+ "warm_p99_ms": 0.240394,
+ "samples_ms": [
+ 0.232293,
+ 0.230896,
+ 0.229429,
+ 0.22894,
+ 0.22894,
+ 0.230406,
+ 0.229918,
+ 0.22901,
+ 0.229079,
+ 0.229429,
+ 0.229009,
+ 0.240394,
+ 0.230337,
+ 0.22901,
+ 0.229499,
+ 0.229429,
+ 0.229918,
+ 0.22852,
+ 0.228521,
+ 0.229428,
+ 0.228521,
+ 0.227055,
+ 0.229499,
+ 0.228032,
+ 0.228102,
+ 0.228521,
+ 0.22894,
+ 0.230406,
+ 0.240394,
+ 0.229987,
+ 0.22901,
+ 0.228102,
+ 0.22901,
+ 0.22901,
+ 0.228102,
+ 0.227544,
+ 0.228102,
+ 0.227613,
+ 0.228032,
+ 0.22901
+ ],
+ "warm_rows_per_second": 2235710.2309942795
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 3406703,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 442,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 442,
+ "zlib_blocks": 0,
+ "xor_blocks": 442,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "a6f992eab989830921c022ff3d6f88caceeb1eeebe107b7ccc88caeb09e9ca68",
+ "table_sha256": "bbaaa5100062ad8592fd8c4f376ec4ec38424253a6c1bee0d0f21c18ec923b04",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.355561135802418,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3407911,
+ "backend_with_key_registry_bytes": 3626389,
+ "backend_with_datrie_key_index_bytes": 3521623,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999640630156,
+ "mean_cosine": 0.9999999789175185,
+ "p05_cosine": 0.9999999730721181,
+ "relative_l2_mean": 0.0002070676065899127,
+ "relative_l2_p95": 0.00023440921164902042,
+ "relative_l2_max": 0.00027016899099782454,
+ "weighted_relative_l2_mean": 0.0002072185978470986,
+ "weighted_relative_frobenius": 0.00020769691233671656,
+ "weighted_mean_cosine": 0.9999999788781782,
+ "max_relative_norm_change": 0.00012017913339375236
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.768897,
+ "warm_p50_ms": 0.228521,
+ "warm_p95_ms": 0.23302279999999997,
+ "warm_p99_ms": 0.25111152999999997,
+ "samples_ms": [
+ 0.232292,
+ 0.229918,
+ 0.22894,
+ 0.227613,
+ 0.228451,
+ 0.229918,
+ 0.229987,
+ 0.229499,
+ 0.258901,
+ 0.232712,
+ 0.229499,
+ 0.227543,
+ 0.229499,
+ 0.228451,
+ 0.229918,
+ 0.229428,
+ 0.228451,
+ 0.228451,
+ 0.228032,
+ 0.228521,
+ 0.228521,
+ 0.228032,
+ 0.229917,
+ 0.229429,
+ 0.22894,
+ 0.238928,
+ 0.229499,
+ 0.228521,
+ 0.228102,
+ 0.228102,
+ 0.227613,
+ 0.228032,
+ 0.227543,
+ 0.228032,
+ 0.228521,
+ 0.228521,
+ 0.230337,
+ 0.230407,
+ 0.228032,
+ 0.228032
+ ],
+ "warm_rows_per_second": 2240494.309056936
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2200767,
+ "unique_rows": 8712,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 273,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 273,
+ "zlib_blocks": 94,
+ "xor_blocks": 179,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8712
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "5f413487c622272d106a2c2cab4f0e33aba0c3ac98cdd387f6108039590b87ab",
+ "table_sha256": "7f973406674ef8687231a08ac83888ca6646e820ad3ecc4feb20a5a423bfd705",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.23608402512036264,
+ "representatives_sha256": "62dc9fb389433ccc5a4329644837c182debd5b33f93e5610c8fed98ee5ca79f3",
+ "structurally_merged_rows": 5420,
+ "backend_artifact_bytes": 2201973,
+ "backend_with_key_registry_bytes": 2420451,
+ "backend_with_datrie_key_index_bytes": 2315685,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900110734293337,
+ "mean_cosine": 0.9986071040782484,
+ "p05_cosine": 0.9931201421840856,
+ "relative_l2_mean": 0.030916340295123496,
+ "relative_l2_p95": 0.11753606067533086,
+ "relative_l2_max": 0.1422308437423335,
+ "weighted_relative_l2_mean": 0.01793043618905993,
+ "weighted_relative_frobenius": 0.009470324713106664,
+ "weighted_mean_cosine": 0.9992361701290867,
+ "max_relative_norm_change": 0.010145610590708034
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900126948932569,
+ "mean_cosine": 0.9986071142742591,
+ "p05_cosine": 0.9931205884476864,
+ "relative_l2_mean": 0.03078874727222894,
+ "relative_l2_p95": 0.11754071778511165,
+ "relative_l2_max": 0.14223837458838182,
+ "weighted_relative_l2_mean": 0.01777303897196514,
+ "weighted_relative_frobenius": 0.009468034050420173,
+ "weighted_mean_cosine": 0.9992361850915106,
+ "max_relative_norm_change": 0.010095191189280194
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.23149,
+ "warm_p50_ms": 0.227962,
+ "warm_p95_ms": 0.25880719999999996,
+ "warm_p99_ms": 0.29290325,
+ "samples_ms": [
+ 0.231384,
+ 0.230407,
+ 0.22901,
+ 0.228521,
+ 0.227962,
+ 0.22901,
+ 0.232293,
+ 0.228102,
+ 0.245981,
+ 0.230406,
+ 0.227962,
+ 0.226635,
+ 0.227613,
+ 0.228102,
+ 0.226635,
+ 0.226985,
+ 0.227055,
+ 0.227124,
+ 0.227613,
+ 0.226216,
+ 0.226636,
+ 0.228032,
+ 0.226635,
+ 0.228102,
+ 0.226636,
+ 0.228102,
+ 0.257784,
+ 0.226635,
+ 0.227194,
+ 0.227613,
+ 0.227473,
+ 0.227054,
+ 0.226635,
+ 0.227612,
+ 0.226146,
+ 0.302273,
+ 0.278248,
+ 0.234737,
+ 0.228451,
+ 0.22894
+ ],
+ "warm_rows_per_second": 2245988.366482133
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2099636,
+ "unique_rows": 8277,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 259,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 259,
+ "zlib_blocks": 110,
+ "xor_blocks": 149,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8277
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "bbf09843f09d773ec55ec2f2a8eb50bbbe9822cf7bd8768a38d0beebb8e76d63",
+ "table_sha256": "a3fc9dde46ac640d4297c0ae901ee7e9a64252dc41af16e982523e999d8412b3",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22631070506758988,
+ "representatives_sha256": "1aa39d9c2a2cba69776ee123ef0a337f933d44e3a82a1a8d36ee472a12a47b61",
+ "structurally_merged_rows": 5855,
+ "backend_artifact_bytes": 2100843,
+ "backend_with_key_registry_bytes": 2319321,
+ "backend_with_datrie_key_index_bytes": 2214555,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800635154178402,
+ "mean_cosine": 0.9981698821459642,
+ "p05_cosine": 0.9912050222029035,
+ "relative_l2_mean": 0.03608720743827846,
+ "relative_l2_p95": 0.13281078096026844,
+ "relative_l2_max": 0.20087697063823262,
+ "weighted_relative_l2_mean": 0.020515703015552056,
+ "weighted_relative_frobenius": 0.009891052383337548,
+ "weighted_mean_cosine": 0.9990175669350744,
+ "max_relative_norm_change": 0.010145610590708034
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800675682083074,
+ "mean_cosine": 0.9981698900290342,
+ "p05_cosine": 0.9912022668645663,
+ "relative_l2_mean": 0.03596599927953503,
+ "relative_l2_p95": 0.13279856767887474,
+ "relative_l2_max": 0.20085791728800229,
+ "weighted_relative_l2_mean": 0.020361497298607506,
+ "weighted_relative_frobenius": 0.009888866721531615,
+ "weighted_mean_cosine": 0.9990175807324861,
+ "max_relative_norm_change": 0.010095191189280194
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.294417,
+ "warm_p50_ms": 0.22900949999999998,
+ "warm_p95_ms": 0.23516634999999997,
+ "warm_p99_ms": 0.25411205,
+ "samples_ms": [
+ 0.234667,
+ 0.231385,
+ 0.231385,
+ 0.22894,
+ 0.228381,
+ 0.231315,
+ 0.244654,
+ 0.229987,
+ 0.230407,
+ 0.228521,
+ 0.229917,
+ 0.228521,
+ 0.229009,
+ 0.228451,
+ 0.22852,
+ 0.228521,
+ 0.228032,
+ 0.22901,
+ 0.22901,
+ 0.229429,
+ 0.228521,
+ 0.228451,
+ 0.22901,
+ 0.260159,
+ 0.232292,
+ 0.229499,
+ 0.229499,
+ 0.229499,
+ 0.22852,
+ 0.230896,
+ 0.22894,
+ 0.228521,
+ 0.228032,
+ 0.22894,
+ 0.22887,
+ 0.228521,
+ 0.228521,
+ 0.228032,
+ 0.233689,
+ 0.23327
+ ],
+ "warm_rows_per_second": 2235715.1122551686
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2081919,
+ "unique_rows": 8202,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 257,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 257,
+ "zlib_blocks": 120,
+ "xor_blocks": 137,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8202
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "64effdba1d8d0a799adf2c143fa824c2282ee666d273bd20d4f6ce45b99d8f39",
+ "table_sha256": "eb149d4b44386cb212e5ec282c6ce92d914fe535222031a1cd0bdb868608aa12",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.23895569099113345,
+ "representatives_sha256": "6bc853a79cffddff151595a7bcea1ed7bf384365d40cb19dbb15de973141547f",
+ "structurally_merged_rows": 5930,
+ "backend_artifact_bytes": 2083126,
+ "backend_with_key_registry_bytes": 2301604,
+ "backend_with_datrie_key_index_bytes": 2196838,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9702653636864929,
+ "mean_cosine": 0.9980383303992661,
+ "p05_cosine": 0.990743200733894,
+ "relative_l2_mean": 0.03727752367335026,
+ "relative_l2_p95": 0.13636840106871834,
+ "relative_l2_max": 0.24369412042518732,
+ "weighted_relative_l2_mean": 0.02125235802853126,
+ "weighted_relative_frobenius": 0.011080112730047647,
+ "weighted_mean_cosine": 0.9989369076135723,
+ "max_relative_norm_change": 0.010145610590708034
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9702650416300592,
+ "mean_cosine": 0.998038341143016,
+ "p05_cosine": 0.990743685168674,
+ "relative_l2_mean": 0.03715741640787692,
+ "relative_l2_p95": 0.13636450062101535,
+ "relative_l2_max": 0.24366716403651428,
+ "weighted_relative_l2_mean": 0.021098848735033624,
+ "weighted_relative_frobenius": 0.011078245050048044,
+ "weighted_mean_cosine": 0.9989369217639886,
+ "max_relative_norm_change": 0.010095191189280194
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.635814,
+ "warm_p50_ms": 0.228102,
+ "warm_p95_ms": 0.24332764999999998,
+ "warm_p99_ms": 0.25433129,
+ "samples_ms": [
+ 0.232292,
+ 0.22901,
+ 0.229918,
+ 0.243188,
+ 0.231384,
+ 0.229918,
+ 0.227962,
+ 0.22894,
+ 0.227054,
+ 0.228521,
+ 0.229009,
+ 0.228032,
+ 0.228521,
+ 0.228102,
+ 0.227124,
+ 0.227613,
+ 0.228521,
+ 0.228102,
+ 0.228033,
+ 0.228032,
+ 0.227543,
+ 0.25967,
+ 0.231803,
+ 0.229429,
+ 0.228102,
+ 0.228102,
+ 0.227963,
+ 0.226146,
+ 0.228521,
+ 0.228032,
+ 0.227962,
+ 0.228102,
+ 0.226147,
+ 0.227124,
+ 0.228032,
+ 0.226566,
+ 0.228451,
+ 0.227124,
+ 0.245981,
+ 0.229009
+ ],
+ "warm_rows_per_second": 2244609.86751541
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2077788,
+ "unique_rows": 8185,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 256,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 256,
+ "zlib_blocks": 125,
+ "xor_blocks": 131,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8185
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "288ae047d4a3ea025f53271f67a80e7c1a9c12fcf36187b3282ce05b4d951009",
+ "table_sha256": "b9fb0439ad88f608319495e30d8c80a65d401bd39dbb59060475c702110aa3db",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22219355800189078,
+ "representatives_sha256": "f037f27f7c62814c89f5cc44386f0e0757a3e83e41b2187d71bbab635d4cd0f1",
+ "structurally_merged_rows": 5947,
+ "backend_artifact_bytes": 2078995,
+ "backend_with_key_registry_bytes": 2297473,
+ "backend_with_datrie_key_index_bytes": 2192707,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9525649397506648,
+ "mean_cosine": 0.99796889285962,
+ "p05_cosine": 0.9905429346239615,
+ "relative_l2_mean": 0.037746117319474796,
+ "relative_l2_p95": 0.1374701495888191,
+ "relative_l2_max": 0.30793962720557705,
+ "weighted_relative_l2_mean": 0.02157451567575557,
+ "weighted_relative_frobenius": 0.01161458655939185,
+ "weighted_mean_cosine": 0.9988891428632327,
+ "max_relative_norm_change": 0.010145610590708034
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9525649504335991,
+ "mean_cosine": 0.9979689029929782,
+ "p05_cosine": 0.9905451002882296,
+ "relative_l2_mean": 0.03762625099845266,
+ "relative_l2_p95": 0.137479103212657,
+ "relative_l2_max": 0.3079354221985442,
+ "weighted_relative_l2_mean": 0.02142115707058293,
+ "weighted_relative_frobenius": 0.01161280946049565,
+ "weighted_mean_cosine": 0.9988891564634181,
+ "max_relative_norm_change": 0.010095191189280194
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.957991,
+ "warm_p50_ms": 0.22845149999999997,
+ "warm_p95_ms": 0.23942624999999998,
+ "warm_p99_ms": 0.25373486,
+ "samples_ms": [
+ 0.250242,
+ 0.232292,
+ 0.229918,
+ 0.22894,
+ 0.22859,
+ 0.22901,
+ 0.22901,
+ 0.227613,
+ 0.228102,
+ 0.226566,
+ 0.22894,
+ 0.228452,
+ 0.228032,
+ 0.228451,
+ 0.228451,
+ 0.22894,
+ 0.228521,
+ 0.238857,
+ 0.22894,
+ 0.22908,
+ 0.227613,
+ 0.227055,
+ 0.227543,
+ 0.22894,
+ 0.227543,
+ 0.228032,
+ 0.228521,
+ 0.227543,
+ 0.228102,
+ 0.227613,
+ 0.228521,
+ 0.228032,
+ 0.228591,
+ 0.227543,
+ 0.255968,
+ 0.23327,
+ 0.228451,
+ 0.227613,
+ 0.226635,
+ 0.227194
+ ],
+ "warm_rows_per_second": 2241175.9169889456
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2077748,
+ "unique_rows": 8185,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 256,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 256,
+ "zlib_blocks": 117,
+ "xor_blocks": 139,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8185
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "d782d297b88e05e492b1903221756f73f25f26fd7a54f257605ea56912bea425",
+ "table_sha256": "d216072fb926e7f2af7600afb31c08323f57ad2822306d0f0a4c2903061c3b87",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2176839008461684,
+ "representatives_sha256": "f52343d6608456aeae58f0697a7c8302e5f9edbd616cce35edc842bb80d63f3c",
+ "structurally_merged_rows": 5947,
+ "backend_artifact_bytes": 2078955,
+ "backend_with_key_registry_bytes": 2297433,
+ "backend_with_datrie_key_index_bytes": 2192667,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9013593202834985,
+ "mean_cosine": 0.9978685166873792,
+ "p05_cosine": 0.9904298560601507,
+ "relative_l2_mean": 0.03818776739965724,
+ "relative_l2_p95": 0.1385518535656877,
+ "relative_l2_max": 0.4453821744456762,
+ "weighted_relative_l2_mean": 0.021859553974485434,
+ "weighted_relative_frobenius": 0.012750560790134983,
+ "weighted_mean_cosine": 0.9988277159065638,
+ "max_relative_norm_change": 0.010145610590708034
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9013408026498358,
+ "mean_cosine": 0.9978685267190944,
+ "p05_cosine": 0.9904296759069757,
+ "relative_l2_mean": 0.03806790055133203,
+ "relative_l2_p95": 0.1385533654179914,
+ "relative_l2_max": 0.445427766006183,
+ "weighted_relative_l2_mean": 0.021706219190303095,
+ "weighted_relative_frobenius": 0.012748906955825902,
+ "weighted_mean_cosine": 0.9988277296235697,
+ "max_relative_norm_change": 0.010095191189280194
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.071623,
+ "warm_p50_ms": 0.228032,
+ "warm_p95_ms": 0.23221234999999996,
+ "warm_p99_ms": 0.24245207,
+ "samples_ms": [
+ 0.231734,
+ 0.230407,
+ 0.228521,
+ 0.228521,
+ 0.227613,
+ 0.228032,
+ 0.229499,
+ 0.228521,
+ 0.22894,
+ 0.227125,
+ 0.228521,
+ 0.227194,
+ 0.241301,
+ 0.228032,
+ 0.227543,
+ 0.228102,
+ 0.228032,
+ 0.227194,
+ 0.227124,
+ 0.227962,
+ 0.228521,
+ 0.228521,
+ 0.227054,
+ 0.228033,
+ 0.228102,
+ 0.227124,
+ 0.227613,
+ 0.228102,
+ 0.227613,
+ 0.227055,
+ 0.243188,
+ 0.228032,
+ 0.228032,
+ 0.227613,
+ 0.228591,
+ 0.228032,
+ 0.22852,
+ 0.226635,
+ 0.226565,
+ 0.227473
+ ],
+ "warm_rows_per_second": 2245298.9054167834
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 132,
+ "storage": {
+ "artifact_bytes": 1865424,
+ "unique_rows": 14132,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "f74165bb32ed3307f423ea4917e648ddb56eef15fa4521f69fd6f16fc3599f66",
+ "table_sha256": "0e2bdbfbc8dca09e94a85dd1f3349a0af8a377df57076f698fe6ceb081197dee",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.036789210978895426,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1866274,
+ "backend_with_key_registry_bytes": 2084752,
+ "backend_with_datrie_key_index_bytes": 1979986,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999471589781962,
+ "mean_cosine": 0.9999794823547403,
+ "p05_cosine": 0.9999697795998994,
+ "relative_l2_mean": 0.0063798520701271065,
+ "relative_l2_p95": 0.007798429717370971,
+ "relative_l2_max": 0.010301694004383433,
+ "weighted_relative_l2_mean": 0.006459010216051861,
+ "weighted_relative_frobenius": 0.0066886524797597335,
+ "weighted_mean_cosine": 0.9999789761503499,
+ "max_relative_norm_change": 0.002403168609907241
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.363315,
+ "warm_p50_ms": 0.121838,
+ "warm_p95_ms": 0.123968,
+ "warm_p99_ms": 0.13491749999999997,
+ "samples_ms": [
+ 0.12341,
+ 0.12348,
+ 0.121244,
+ 0.121245,
+ 0.121663,
+ 0.123968,
+ 0.122502,
+ 0.122502,
+ 0.121035,
+ 0.123549,
+ 0.122642,
+ 0.12348,
+ 0.122502,
+ 0.123968,
+ 0.12299,
+ 0.122502,
+ 0.123061,
+ 0.122571,
+ 0.122571,
+ 0.122013,
+ 0.120616,
+ 0.121035,
+ 0.121035,
+ 0.121524,
+ 0.122013,
+ 0.121105,
+ 0.121105,
+ 0.121105,
+ 0.141918,
+ 0.120266,
+ 0.121105,
+ 0.121524,
+ 0.121524,
+ 0.120686,
+ 0.122152,
+ 0.120267,
+ 0.120616,
+ 0.120616,
+ 0.122571,
+ 0.120756
+ ],
+ "warm_rows_per_second": 4202301.416635204
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1726950,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 228,
+ "zlib_blocks": 0,
+ "xor_blocks": 228,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "f74165bb32ed3307f423ea4917e648ddb56eef15fa4521f69fd6f16fc3599f66",
+ "table_sha256": "e3c61a84ae42232b395b0aec70a7cd0bae29d2b67c4f706876b2204483fc318d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.20328680495731533,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1728203,
+ "backend_with_key_registry_bytes": 1946681,
+ "backend_with_datrie_key_index_bytes": 1841915,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999471589781962,
+ "mean_cosine": 0.9999794823547403,
+ "p05_cosine": 0.9999697795998994,
+ "relative_l2_mean": 0.0063798520701271065,
+ "relative_l2_p95": 0.007798429717370971,
+ "relative_l2_max": 0.010301694004383433,
+ "weighted_relative_l2_mean": 0.006459010216051861,
+ "weighted_relative_frobenius": 0.0066886524797597335,
+ "weighted_mean_cosine": 0.9999789761503499,
+ "max_relative_norm_change": 0.002403168609907241
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.94925,
+ "warm_p50_ms": 0.1317905,
+ "warm_p95_ms": 0.13938575,
+ "warm_p99_ms": 0.14598729000000002,
+ "samples_ms": [
+ 0.139054,
+ 0.132978,
+ 0.134933,
+ 0.132489,
+ 0.132489,
+ 0.132978,
+ 0.132419,
+ 0.131023,
+ 0.131512,
+ 0.146178,
+ 0.133955,
+ 0.132908,
+ 0.129625,
+ 0.131022,
+ 0.131092,
+ 0.130114,
+ 0.131022,
+ 0.133397,
+ 0.132,
+ 0.132,
+ 0.132,
+ 0.130114,
+ 0.130114,
+ 0.132,
+ 0.131581,
+ 0.131512,
+ 0.131512,
+ 0.131023,
+ 0.132,
+ 0.131511,
+ 0.132419,
+ 0.130603,
+ 0.130604,
+ 0.132,
+ 0.131581,
+ 0.130533,
+ 0.131092,
+ 0.131022,
+ 0.132001,
+ 0.145689
+ ],
+ "warm_rows_per_second": 3884953.771326461
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1726950,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 228,
+ "zlib_blocks": 0,
+ "xor_blocks": 228,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "f74165bb32ed3307f423ea4917e648ddb56eef15fa4521f69fd6f16fc3599f66",
+ "table_sha256": "e3c61a84ae42232b395b0aec70a7cd0bae29d2b67c4f706876b2204483fc318d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24753053300082684,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1728202,
+ "backend_with_key_registry_bytes": 1946680,
+ "backend_with_datrie_key_index_bytes": 1841914,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999471589781962,
+ "mean_cosine": 0.9999794823547403,
+ "p05_cosine": 0.9999697795998994,
+ "relative_l2_mean": 0.0063798520701271065,
+ "relative_l2_p95": 0.007798429717370971,
+ "relative_l2_max": 0.010301694004383433,
+ "weighted_relative_l2_mean": 0.006459010216051861,
+ "weighted_relative_frobenius": 0.0066886524797597335,
+ "weighted_mean_cosine": 0.9999789761503499,
+ "max_relative_norm_change": 0.002403168609907241
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.922221,
+ "warm_p50_ms": 0.1315115,
+ "warm_p95_ms": 0.1354263,
+ "warm_p99_ms": 0.15232642,
+ "samples_ms": [
+ 0.136819,
+ 0.134374,
+ 0.132978,
+ 0.132,
+ 0.131511,
+ 0.135353,
+ 0.133886,
+ 0.133886,
+ 0.132908,
+ 0.131511,
+ 0.132,
+ 0.131022,
+ 0.131092,
+ 0.130533,
+ 0.132977,
+ 0.131092,
+ 0.132,
+ 0.130533,
+ 0.132,
+ 0.162241,
+ 0.132419,
+ 0.130115,
+ 0.131092,
+ 0.131092,
+ 0.130045,
+ 0.129556,
+ 0.130533,
+ 0.130533,
+ 0.131511,
+ 0.130603,
+ 0.130115,
+ 0.130115,
+ 0.132,
+ 0.131581,
+ 0.132489,
+ 0.132001,
+ 0.131512,
+ 0.131511,
+ 0.130114,
+ 0.132
+ ],
+ "warm_rows_per_second": 3893195.6520912615
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1726950,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 228,
+ "zlib_blocks": 0,
+ "xor_blocks": 228,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "f74165bb32ed3307f423ea4917e648ddb56eef15fa4521f69fd6f16fc3599f66",
+ "table_sha256": "e3c61a84ae42232b395b0aec70a7cd0bae29d2b67c4f706876b2204483fc318d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.202649642014876,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1728202,
+ "backend_with_key_registry_bytes": 1946680,
+ "backend_with_datrie_key_index_bytes": 1841914,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999471589781962,
+ "mean_cosine": 0.9999794823547403,
+ "p05_cosine": 0.9999697795998994,
+ "relative_l2_mean": 0.0063798520701271065,
+ "relative_l2_p95": 0.007798429717370971,
+ "relative_l2_max": 0.010301694004383433,
+ "weighted_relative_l2_mean": 0.006459010216051861,
+ "weighted_relative_frobenius": 0.0066886524797597335,
+ "weighted_mean_cosine": 0.9999789761503499,
+ "max_relative_norm_change": 0.002403168609907241
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.91426,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.1358653,
+ "warm_p99_ms": 0.14259861,
+ "samples_ms": [
+ 0.135771,
+ 0.137657,
+ 0.133397,
+ 0.131511,
+ 0.131022,
+ 0.133397,
+ 0.130534,
+ 0.131581,
+ 0.131022,
+ 0.132,
+ 0.145758,
+ 0.132908,
+ 0.132001,
+ 0.131093,
+ 0.131093,
+ 0.129207,
+ 0.131023,
+ 0.131022,
+ 0.131511,
+ 0.132,
+ 0.131511,
+ 0.129625,
+ 0.131022,
+ 0.131022,
+ 0.131581,
+ 0.130603,
+ 0.132,
+ 0.131512,
+ 0.131022,
+ 0.131511,
+ 0.131092,
+ 0.131511,
+ 0.131511,
+ 0.132001,
+ 0.130604,
+ 0.131582,
+ 0.131511,
+ 0.131511,
+ 0.131022,
+ 0.131511
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1175823,
+ "unique_rows": 8713,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 141,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 141,
+ "zlib_blocks": 0,
+ "xor_blocks": 141,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8713
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "0e16307f6d7d2f16b5df0bf18bb5e0ad4aa9696151e7d65d3a3cb64efcdc223b",
+ "table_sha256": "123562d52dd4c313e376c968d2224f93c7415d264d6a7974e87c1b10c8d1a495",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14536297111772,
+ "representatives_sha256": "6768f1ff6fc3d7df653ab57688d7db1413898bee807bd7a61a3abfa4c5029f9c",
+ "structurally_merged_rows": 5419,
+ "backend_artifact_bytes": 1177072,
+ "backend_with_key_registry_bytes": 1395550,
+ "backend_with_datrie_key_index_bytes": 1290784,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9899801138449026,
+ "mean_cosine": 0.9985845270900451,
+ "p05_cosine": 0.9930813891868836,
+ "relative_l2_mean": 0.034861390437250256,
+ "relative_l2_p95": 0.11772796904728766,
+ "relative_l2_max": 0.14200820684716933,
+ "weighted_relative_l2_mean": 0.022794419147791192,
+ "weighted_relative_frobenius": 0.011592827136221024,
+ "weighted_mean_cosine": 0.9992138226113876,
+ "max_relative_norm_change": 0.01097468020532888
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900024071397229,
+ "mean_cosine": 0.9985897209075235,
+ "p05_cosine": 0.9930826313595622,
+ "relative_l2_mean": 0.031029785163936886,
+ "relative_l2_p95": 0.11787638870426462,
+ "relative_l2_max": 0.14219781943961757,
+ "weighted_relative_l2_mean": 0.017946976639319694,
+ "weighted_relative_frobenius": 0.009600591901907743,
+ "weighted_mean_cosine": 0.9992252098943923,
+ "max_relative_norm_change": 0.010098747555370645
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.459126,
+ "warm_p50_ms": 0.131023,
+ "warm_p95_ms": 0.13375319999999996,
+ "warm_p99_ms": 0.14542052,
+ "samples_ms": [
+ 0.140521,
+ 0.132489,
+ 0.132,
+ 0.130114,
+ 0.131022,
+ 0.131022,
+ 0.131092,
+ 0.13249,
+ 0.132,
+ 0.131511,
+ 0.132,
+ 0.130604,
+ 0.131023,
+ 0.131023,
+ 0.131023,
+ 0.130534,
+ 0.132,
+ 0.128299,
+ 0.130604,
+ 0.130115,
+ 0.130603,
+ 0.131512,
+ 0.131512,
+ 0.131022,
+ 0.131092,
+ 0.131162,
+ 0.130533,
+ 0.130603,
+ 0.130603,
+ 0.148553,
+ 0.133397,
+ 0.130603,
+ 0.132,
+ 0.132,
+ 0.132489,
+ 0.130603,
+ 0.130044,
+ 0.130533,
+ 0.132,
+ 0.130184
+ ],
+ "warm_rows_per_second": 3907710.8599253567
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1128366,
+ "unique_rows": 8285,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 134,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 134,
+ "zlib_blocks": 4,
+ "xor_blocks": 130,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8285
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "646f01f4709bb7e548edcc943f72979f595c75c76b8723f5e9942df74d389984",
+ "table_sha256": "a221b6fbe42e9142dacbda3d08460e7c331d6c2c40d697dd22fbb58832d0d5ea",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1393656248692423,
+ "representatives_sha256": "8b2a9c93f4b9dafeed1567dbc52b1969bd2d5924cae49daf7b12835ce067e2db",
+ "structurally_merged_rows": 5847,
+ "backend_artifact_bytes": 1129615,
+ "backend_with_key_registry_bytes": 1348093,
+ "backend_with_datrie_key_index_bytes": 1243327,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9801019509718193,
+ "mean_cosine": 0.9981548793445838,
+ "p05_cosine": 0.9912022094948002,
+ "relative_l2_mean": 0.0397624235208118,
+ "relative_l2_p95": 0.1327321045453814,
+ "relative_l2_max": 0.20063492021887994,
+ "weighted_relative_l2_mean": 0.025245128151299166,
+ "weighted_relative_frobenius": 0.011933935921140228,
+ "weighted_mean_cosine": 0.9989989802006313,
+ "max_relative_norm_change": 0.01097468020532888
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800593020711716,
+ "mean_cosine": 0.9981587042292385,
+ "p05_cosine": 0.9911925880905068,
+ "relative_l2_mean": 0.03612972620269922,
+ "relative_l2_p95": 0.13290670551842743,
+ "relative_l2_max": 0.20076566518744676,
+ "weighted_relative_l2_mean": 0.02049709007338642,
+ "weighted_relative_frobenius": 0.010011650874360624,
+ "weighted_mean_cosine": 0.999009687671515,
+ "max_relative_norm_change": 0.010098747555370645
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.518352,
+ "warm_p50_ms": 0.131512,
+ "warm_p95_ms": 0.1358234,
+ "warm_p99_ms": 0.14602146,
+ "samples_ms": [
+ 0.135771,
+ 0.133397,
+ 0.131512,
+ 0.131023,
+ 0.132908,
+ 0.132489,
+ 0.132978,
+ 0.131511,
+ 0.133886,
+ 0.133886,
+ 0.131511,
+ 0.131512,
+ 0.132,
+ 0.131092,
+ 0.131092,
+ 0.132489,
+ 0.131023,
+ 0.131512,
+ 0.131023,
+ 0.132,
+ 0.131511,
+ 0.151905,
+ 0.136819,
+ 0.131511,
+ 0.131512,
+ 0.132,
+ 0.131511,
+ 0.132,
+ 0.131512,
+ 0.131093,
+ 0.130185,
+ 0.131511,
+ 0.131511,
+ 0.132,
+ 0.135283,
+ 0.134305,
+ 0.132,
+ 0.132001,
+ 0.131093,
+ 0.131512
+ ],
+ "warm_rows_per_second": 3893180.850416692
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1119365,
+ "unique_rows": 8202,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 133,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 133,
+ "zlib_blocks": 3,
+ "xor_blocks": 130,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8202
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "3a561747837ca991db9b9b7a7abf98bd3daf970df326ce35ce08132003bf3fb9",
+ "table_sha256": "99a21c2eceaec8b32fc69b4d5de4db8f66e98d2f3a3136cd91e65c3fb877eade",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13420826010406017,
+ "representatives_sha256": "c0e878fe084ee8df1cfa5e4eeefcfe3f0d6371135b2bbe3bd8e033450e686031",
+ "structurally_merged_rows": 5930,
+ "backend_artifact_bytes": 1120614,
+ "backend_with_key_registry_bytes": 1339092,
+ "backend_with_datrie_key_index_bytes": 1234326,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9701486181461639,
+ "mean_cosine": 0.9980143916261239,
+ "p05_cosine": 0.9907430056165512,
+ "relative_l2_mean": 0.041009092049086614,
+ "relative_l2_p95": 0.1365202950335466,
+ "relative_l2_max": 0.24363978297137226,
+ "weighted_relative_l2_mean": 0.0260052352356428,
+ "weighted_relative_frobenius": 0.012924036411778727,
+ "weighted_mean_cosine": 0.9989138264682151,
+ "max_relative_norm_change": 0.01097468020532888
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9701112399545528,
+ "mean_cosine": 0.9980180167163665,
+ "p05_cosine": 0.9906755883005909,
+ "relative_l2_mean": 0.037414039058715175,
+ "relative_l2_p95": 0.13672098797393897,
+ "relative_l2_max": 0.24369098703870892,
+ "weighted_relative_l2_mean": 0.021280622747528337,
+ "weighted_relative_frobenius": 0.01117075773889507,
+ "weighted_mean_cosine": 0.9989244699849454,
+ "max_relative_norm_change": 0.010098747555370645
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.155875,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.13579295,
+ "warm_p99_ms": 0.1521669,
+ "samples_ms": [
+ 0.134934,
+ 0.133397,
+ 0.131511,
+ 0.131092,
+ 0.131092,
+ 0.133956,
+ 0.132908,
+ 0.130603,
+ 0.131511,
+ 0.130115,
+ 0.131512,
+ 0.131581,
+ 0.162381,
+ 0.135353,
+ 0.131581,
+ 0.131092,
+ 0.131511,
+ 0.130114,
+ 0.132,
+ 0.132,
+ 0.132,
+ 0.131093,
+ 0.131511,
+ 0.131022,
+ 0.131022,
+ 0.131022,
+ 0.130114,
+ 0.131022,
+ 0.130603,
+ 0.131022,
+ 0.131511,
+ 0.131092,
+ 0.131092,
+ 0.130533,
+ 0.131511,
+ 0.131093,
+ 0.130604,
+ 0.135772,
+ 0.136191,
+ 0.133467
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1117375,
+ "unique_rows": 8185,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 133,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 132,
+ "zlib_blocks": 3,
+ "xor_blocks": 129,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8185
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "7b92d29e0623f52881105e40964eb31dd3488b45ea812fede290b3e1d3d3419c",
+ "table_sha256": "f7866a75986a6ee885c564b71332a5d937238ba1d0ce4ce7802eed12c2074983",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.15252687805332243,
+ "representatives_sha256": "c1a9fa5dcc227e4066722a5ae828af0d54b42ec976d465b8a1ccf7ad8390f6df",
+ "structurally_merged_rows": 5947,
+ "backend_artifact_bytes": 1118624,
+ "backend_with_key_registry_bytes": 1337102,
+ "backend_with_datrie_key_index_bytes": 1232336,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9526420225325654,
+ "mean_cosine": 0.9979460880636477,
+ "p05_cosine": 0.9905713481499385,
+ "relative_l2_mean": 0.04146497261643976,
+ "relative_l2_p95": 0.13751176348738786,
+ "relative_l2_max": 0.30767438969030325,
+ "weighted_relative_l2_mean": 0.02631967532395068,
+ "weighted_relative_frobenius": 0.013384670786946199,
+ "weighted_mean_cosine": 0.9988666273981026,
+ "max_relative_norm_change": 0.01097468020532888
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9526048497143683,
+ "mean_cosine": 0.9979496431001218,
+ "p05_cosine": 0.9905341890130446,
+ "relative_l2_mean": 0.03787774947286743,
+ "relative_l2_p95": 0.13756244760914405,
+ "relative_l2_max": 0.30787719238090006,
+ "weighted_relative_l2_mean": 0.021600008171405432,
+ "weighted_relative_frobenius": 0.011701695336837683,
+ "weighted_mean_cosine": 0.9988772090881424,
+ "max_relative_norm_change": 0.010098747555370645
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.398644,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.13712965,
+ "warm_p99_ms": 0.14862224,
+ "samples_ms": [
+ 0.136749,
+ 0.135352,
+ 0.13626,
+ 0.151346,
+ 0.134794,
+ 0.132,
+ 0.132,
+ 0.132978,
+ 0.131093,
+ 0.132,
+ 0.132489,
+ 0.131511,
+ 0.132978,
+ 0.132,
+ 0.131511,
+ 0.131023,
+ 0.132489,
+ 0.131581,
+ 0.131511,
+ 0.131511,
+ 0.131023,
+ 0.131582,
+ 0.131512,
+ 0.130604,
+ 0.130604,
+ 0.133467,
+ 0.132978,
+ 0.132908,
+ 0.133397,
+ 0.132489,
+ 0.132001,
+ 0.133397,
+ 0.144362,
+ 0.133886,
+ 0.131511,
+ 0.132,
+ 0.133397,
+ 0.132,
+ 0.132,
+ 0.131512
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1117462,
+ "unique_rows": 8184,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 132,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 132,
+ "zlib_blocks": 5,
+ "xor_blocks": 127,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8184
+ },
+ "source_float32_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178",
+ "decoded_float32_sha256": "82400842112d5cc5e60fe9ebbe01788530c547bfffe6a2cc67fb260f7c00c5f2",
+ "table_sha256": "53dc99e9f14b3b221ce077544c3ce9041959917141f59efd34d46b6f36f9b6d9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.13213550811633468,
+ "representatives_sha256": "bfb33af6533f0fd7d63a625008e67a6a91770e6f159b62efed6590253854b2b2",
+ "structurally_merged_rows": 5948,
+ "backend_artifact_bytes": 1118711,
+ "backend_with_key_registry_bytes": 1337189,
+ "backend_with_datrie_key_index_bytes": 1232423,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9011804812433917,
+ "mean_cosine": 0.9978587097768464,
+ "p05_cosine": 0.9904579582661998,
+ "relative_l2_mean": 0.04186114631853476,
+ "relative_l2_p95": 0.1381574059207003,
+ "relative_l2_max": 0.4456799429253098,
+ "weighted_relative_l2_mean": 0.02658140480970069,
+ "weighted_relative_frobenius": 0.01430581292194207,
+ "weighted_mean_cosine": 0.9988116744021801,
+ "max_relative_norm_change": 0.01097468020532888
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9013204637301045,
+ "mean_cosine": 0.9978622987668155,
+ "p05_cosine": 0.990461845214681,
+ "relative_l2_mean": 0.03827414125471684,
+ "relative_l2_p95": 0.13832441220687874,
+ "relative_l2_max": 0.44548952244239215,
+ "weighted_relative_l2_mean": 0.02186235257397199,
+ "weighted_relative_frobenius": 0.012744182905516623,
+ "weighted_mean_cosine": 0.9988223194633946,
+ "max_relative_norm_change": 0.010098747555370645
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.264339,
+ "warm_p50_ms": 0.131511,
+ "warm_p95_ms": 0.13409905,
+ "warm_p99_ms": 0.14215104,
+ "samples_ms": [
+ 0.138147,
+ 0.133886,
+ 0.133467,
+ 0.130604,
+ 0.132419,
+ 0.132489,
+ 0.13242,
+ 0.132,
+ 0.130114,
+ 0.130534,
+ 0.131511,
+ 0.131022,
+ 0.131022,
+ 0.131092,
+ 0.131581,
+ 0.131512,
+ 0.13249,
+ 0.132978,
+ 0.130603,
+ 0.131511,
+ 0.132,
+ 0.132489,
+ 0.130534,
+ 0.132,
+ 0.130603,
+ 0.131092,
+ 0.129625,
+ 0.130044,
+ 0.132489,
+ 0.144711,
+ 0.131023,
+ 0.131023,
+ 0.132419,
+ 0.131511,
+ 0.132,
+ 0.129066,
+ 0.130044,
+ 0.131511,
+ 0.131511,
+ 0.129625
+ ],
+ "warm_rows_per_second": 3893210.453878383
+ }
+ }
+ ],
+ "baseline_quality": {},
+ "training_row_sha256": "c57dd713d91fb69fec6853aa0b34ca8b479f86acffb9aca5ab48614b45217178"
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 11,
+ "rows": 14132,
+ "dim": 128,
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "source_float32_bytes": 7235584,
+ "key_registry_bytes": 218478,
+ "unchanged_backbone_artifact_bytes": 0,
+ "alternative_datrie_key_index_bytes": 113712,
+ "query_rows": 512,
+ "query_ids_sha256": "48cdbdeffaad5c50239fec7524368c85bd12fc21eefeb42effaa6c7c5ff0c323",
+ "frequency_trace_sha256": "fc8afe9c1f9a53dfd826267bdad324151b311d81a4be4e4834bd129ccf363886",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 512,
+ "storage": {
+ "artifact_bytes": 7235584,
+ "unique_rows": 14132,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "table_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.03935490385629237,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 7236389,
+ "backend_with_key_registry_bytes": 7454867,
+ "backend_with_datrie_key_index_bytes": 7350101,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.379937,
+ "warm_p50_ms": 0.062473,
+ "warm_p95_ms": 0.0648165,
+ "warm_p99_ms": 0.08188186999999998,
+ "samples_ms": [
+ 0.064813,
+ 0.064393,
+ 0.064813,
+ 0.064812,
+ 0.064254,
+ 0.09275,
+ 0.064883,
+ 0.063416,
+ 0.064393,
+ 0.064813,
+ 0.064813,
+ 0.064254,
+ 0.064324,
+ 0.062438,
+ 0.063835,
+ 0.062857,
+ 0.062438,
+ 0.06195,
+ 0.062019,
+ 0.062438,
+ 0.061949,
+ 0.062368,
+ 0.062019,
+ 0.062508,
+ 0.06153,
+ 0.061879,
+ 0.061949,
+ 0.061949,
+ 0.062019,
+ 0.061111,
+ 0.062858,
+ 0.061949,
+ 0.062368,
+ 0.062997,
+ 0.062019,
+ 0.062089,
+ 0.062438,
+ 0.062438,
+ 0.062858,
+ 0.063346
+ ],
+ "warm_rows_per_second": 8195540.473484546
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 6812385,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 884,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 884,
+ "zlib_blocks": 49,
+ "xor_blocks": 835,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "table_sha256": "3c308989d9e4435c9b64bc1f8042a24b58e6af06c675322b29228495836ae5d3",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6078370399773121,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6813594,
+ "backend_with_key_registry_bytes": 7032072,
+ "backend_with_datrie_key_index_bytes": 6927306,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 13.036797,
+ "warm_p50_ms": 0.071936,
+ "warm_p95_ms": 0.07655045,
+ "warm_p99_ms": 0.07840512999999999,
+ "samples_ms": [
+ 0.079549,
+ 0.076616,
+ 0.075638,
+ 0.075639,
+ 0.074731,
+ 0.073822,
+ 0.07473,
+ 0.074311,
+ 0.0748,
+ 0.076547,
+ 0.073822,
+ 0.074311,
+ 0.07473,
+ 0.071517,
+ 0.071448,
+ 0.071448,
+ 0.071936,
+ 0.071029,
+ 0.071518,
+ 0.071936,
+ 0.071448,
+ 0.071518,
+ 0.07012,
+ 0.071378,
+ 0.073334,
+ 0.072915,
+ 0.071936,
+ 0.071937,
+ 0.071029,
+ 0.070958,
+ 0.071448,
+ 0.071937,
+ 0.071378,
+ 0.07047,
+ 0.071937,
+ 0.069981,
+ 0.071378,
+ 0.071448,
+ 0.071377,
+ 0.071867
+ ],
+ "warm_rows_per_second": 7117437.722419929
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 6812385,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 884,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 884,
+ "zlib_blocks": 49,
+ "xor_blocks": 835,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "table_sha256": "3c308989d9e4435c9b64bc1f8042a24b58e6af06c675322b29228495836ae5d3",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6630070290993899,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6813593,
+ "backend_with_key_registry_bytes": 7032071,
+ "backend_with_datrie_key_index_bytes": 6927305,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.959691,
+ "warm_p50_ms": 0.071378,
+ "warm_p95_ms": 0.076197,
+ "warm_p99_ms": 0.08105382,
+ "samples_ms": [
+ 0.076197,
+ 0.076127,
+ 0.074311,
+ 0.073823,
+ 0.07522,
+ 0.073753,
+ 0.073822,
+ 0.073822,
+ 0.076197,
+ 0.073822,
+ 0.074311,
+ 0.074241,
+ 0.073823,
+ 0.07047,
+ 0.0704,
+ 0.071378,
+ 0.070958,
+ 0.071029,
+ 0.069562,
+ 0.070959,
+ 0.069074,
+ 0.070959,
+ 0.074241,
+ 0.071378,
+ 0.071378,
+ 0.071448,
+ 0.070959,
+ 0.07047,
+ 0.071448,
+ 0.070959,
+ 0.07054,
+ 0.069492,
+ 0.084159,
+ 0.072425,
+ 0.073333,
+ 0.07054,
+ 0.07054,
+ 0.07047,
+ 0.070959,
+ 0.069073
+ ],
+ "warm_rows_per_second": 7173078.539606042
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 6812385,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 884,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 884,
+ "zlib_blocks": 49,
+ "xor_blocks": 835,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "table_sha256": "3c308989d9e4435c9b64bc1f8042a24b58e6af06c675322b29228495836ae5d3",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6472038871143013,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 6813593,
+ "backend_with_key_registry_bytes": 7032071,
+ "backend_with_datrie_key_index_bytes": 6927305,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.96018,
+ "warm_p50_ms": 0.071448,
+ "warm_p95_ms": 0.07561979999999997,
+ "warm_p99_ms": 0.12001117999999995,
+ "samples_ms": [
+ 0.081924,
+ 0.075288,
+ 0.074241,
+ 0.07473,
+ 0.071029,
+ 0.072007,
+ 0.072425,
+ 0.072425,
+ 0.072006,
+ 0.144362,
+ 0.073822,
+ 0.072426,
+ 0.071448,
+ 0.071028,
+ 0.070959,
+ 0.072426,
+ 0.071028,
+ 0.071378,
+ 0.071378,
+ 0.072425,
+ 0.070959,
+ 0.073334,
+ 0.073334,
+ 0.072915,
+ 0.070539,
+ 0.071098,
+ 0.07061,
+ 0.07047,
+ 0.071448,
+ 0.07047,
+ 0.071448,
+ 0.07047,
+ 0.071029,
+ 0.071937,
+ 0.071447,
+ 0.07054,
+ 0.069981,
+ 0.07047,
+ 0.07047,
+ 0.071448
+ ],
+ "warm_rows_per_second": 7166050.834173105
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4312855,
+ "unique_rows": 8755,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 548,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 548,
+ "zlib_blocks": 335,
+ "xor_blocks": 213,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8755
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "e7fe4379c53dd2f901f93b06027b9675225cf138d76c14febb466dffa8a50e74",
+ "table_sha256": "dd763eb9308f1f0a4ced7bc88972e9e793c3ada18bae66b28b048d89b212ab9b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4010178688913584,
+ "representatives_sha256": "018390edeae40939895cd7c39faa236b68fafee217da3bbfcd60df4c223ab3f8",
+ "structurally_merged_rows": 5377,
+ "backend_artifact_bytes": 4314062,
+ "backend_with_key_registry_bytes": 4532540,
+ "backend_with_datrie_key_index_bytes": 4427774,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900010721780895,
+ "mean_cosine": 0.9986084953232017,
+ "p05_cosine": 0.993215256177573,
+ "relative_l2_mean": 0.03070619265155565,
+ "relative_l2_p95": 0.11651034927530395,
+ "relative_l2_max": 0.1419349417581233,
+ "weighted_relative_l2_mean": 0.017809658251168703,
+ "weighted_relative_frobenius": 0.009554490322035257,
+ "weighted_mean_cosine": 0.9992321028745716,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900010721780895,
+ "mean_cosine": 0.9986084953232017,
+ "p05_cosine": 0.993215256177573,
+ "relative_l2_mean": 0.03070619265155565,
+ "relative_l2_p95": 0.11651034927530395,
+ "relative_l2_max": 0.1419349417581233,
+ "weighted_relative_l2_mean": 0.017809658251168703,
+ "weighted_relative_frobenius": 0.009554490322035257,
+ "weighted_mean_cosine": 0.9992321028745716,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.352991,
+ "warm_p50_ms": 0.07047,
+ "warm_p95_ms": 0.0748419,
+ "warm_p99_ms": 0.07708675,
+ "samples_ms": [
+ 0.075638,
+ 0.072844,
+ 0.0748,
+ 0.073753,
+ 0.071867,
+ 0.073822,
+ 0.072914,
+ 0.073822,
+ 0.073333,
+ 0.072355,
+ 0.072286,
+ 0.071867,
+ 0.071866,
+ 0.078013,
+ 0.071937,
+ 0.069143,
+ 0.069632,
+ 0.069003,
+ 0.069631,
+ 0.068584,
+ 0.069492,
+ 0.069632,
+ 0.07054,
+ 0.069981,
+ 0.069981,
+ 0.070889,
+ 0.071448,
+ 0.0748,
+ 0.07047,
+ 0.07047,
+ 0.068654,
+ 0.069492,
+ 0.069492,
+ 0.069562,
+ 0.070051,
+ 0.069003,
+ 0.069562,
+ 0.068654,
+ 0.07047,
+ 0.069423
+ ],
+ "warm_rows_per_second": 7265503.050943663
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4091660,
+ "unique_rows": 8289,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 519,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 518,
+ "zlib_blocks": 347,
+ "xor_blocks": 171,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8289
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "4f2e683d7805e5d20af336ad1ab4e4f9a34340d282bd9f57cc6ef4765664588c",
+ "table_sha256": "d6b55db5687508cccadab3e8564807b43d8397f4a2469dec2e421d674e16dbf9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3817510849330574,
+ "representatives_sha256": "a901c4c315d184a9ea7e4d0d9d8e32fcffca0eb7029e87fd2900ce22354f6c12",
+ "structurally_merged_rows": 5843,
+ "backend_artifact_bytes": 4092867,
+ "backend_with_key_registry_bytes": 4311345,
+ "backend_with_datrie_key_index_bytes": 4206579,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9801304891798318,
+ "mean_cosine": 0.9981494860596879,
+ "p05_cosine": 0.9911426784933394,
+ "relative_l2_mean": 0.036195062260779946,
+ "relative_l2_p95": 0.13323938209852676,
+ "relative_l2_max": 0.20007905691257175,
+ "weighted_relative_l2_mean": 0.020541834143205002,
+ "weighted_relative_frobenius": 0.009887839452694107,
+ "weighted_mean_cosine": 0.9990035141227676,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9801304891798318,
+ "mean_cosine": 0.9981494860596879,
+ "p05_cosine": 0.9911426784933394,
+ "relative_l2_mean": 0.036195062260779946,
+ "relative_l2_p95": 0.13323938209852676,
+ "relative_l2_max": 0.20007905691257175,
+ "weighted_relative_l2_mean": 0.020541834143205002,
+ "weighted_relative_frobenius": 0.009887839452694107,
+ "weighted_mean_cosine": 0.9990035141227676,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.468858,
+ "warm_p50_ms": 0.070749,
+ "warm_p95_ms": 0.11067709999999997,
+ "warm_p99_ms": 0.13698629999999998,
+ "samples_ms": [
+ 0.078083,
+ 0.07473,
+ 0.072914,
+ 0.102247,
+ 0.148971,
+ 0.118241,
+ 0.108813,
+ 0.109301,
+ 0.110279,
+ 0.105949,
+ 0.103644,
+ 0.101759,
+ 0.1012,
+ 0.102178,
+ 0.07948,
+ 0.069912,
+ 0.073334,
+ 0.073334,
+ 0.071867,
+ 0.069981,
+ 0.07054,
+ 0.068654,
+ 0.067607,
+ 0.068166,
+ 0.069003,
+ 0.069492,
+ 0.068654,
+ 0.068096,
+ 0.068165,
+ 0.068584,
+ 0.067746,
+ 0.068165,
+ 0.068584,
+ 0.067606,
+ 0.069492,
+ 0.069143,
+ 0.068165,
+ 0.070958,
+ 0.071378,
+ 0.069911
+ ],
+ "warm_rows_per_second": 7236851.40426013
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4063691,
+ "unique_rows": 8231,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 515,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 515,
+ "zlib_blocks": 359,
+ "xor_blocks": 156,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8231
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "f299252d533d07c324070ebea99a861f94d69e0aa6c3b26fae7871addc807785",
+ "table_sha256": "268961ca0d0669b1a658f202a2e302a8c24e26e2530452a187ffef40403acaff",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.38775255205109715,
+ "representatives_sha256": "709636525f6f6787155801add12c053730f005d6da5439518e9bee1dbdb81e10",
+ "structurally_merged_rows": 5901,
+ "backend_artifact_bytes": 4064898,
+ "backend_with_key_registry_bytes": 4283376,
+ "backend_with_datrie_key_index_bytes": 4178610,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9700633688165851,
+ "mean_cosine": 0.998046379372092,
+ "p05_cosine": 0.9907880633674943,
+ "relative_l2_mean": 0.03711988639603293,
+ "relative_l2_p95": 0.13578873025592136,
+ "relative_l2_max": 0.24521876248037364,
+ "weighted_relative_l2_mean": 0.021030742895474887,
+ "weighted_relative_frobenius": 0.010116847994454945,
+ "weighted_mean_cosine": 0.9989485428849026,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9700633688165851,
+ "mean_cosine": 0.998046379372092,
+ "p05_cosine": 0.9907880633674943,
+ "relative_l2_mean": 0.03711988639603293,
+ "relative_l2_p95": 0.13578873025592136,
+ "relative_l2_max": 0.24521876248037364,
+ "weighted_relative_l2_mean": 0.021030742895474887,
+ "weighted_relative_frobenius": 0.010116847994454945,
+ "weighted_mean_cosine": 0.9989485428849026,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.594095,
+ "warm_p50_ms": 0.07054,
+ "warm_p95_ms": 0.07780269999999999,
+ "warm_p99_ms": 0.08075529000000001,
+ "samples_ms": [
+ 0.080946,
+ 0.076686,
+ 0.074241,
+ 0.074241,
+ 0.073752,
+ 0.072914,
+ 0.071517,
+ 0.073403,
+ 0.071937,
+ 0.071518,
+ 0.071028,
+ 0.069981,
+ 0.068584,
+ 0.070539,
+ 0.071448,
+ 0.071029,
+ 0.069562,
+ 0.069562,
+ 0.069143,
+ 0.069562,
+ 0.07047,
+ 0.069981,
+ 0.069074,
+ 0.07061,
+ 0.070539,
+ 0.069981,
+ 0.069562,
+ 0.07054,
+ 0.069562,
+ 0.069073,
+ 0.069562,
+ 0.07054,
+ 0.073403,
+ 0.077663,
+ 0.072844,
+ 0.07047,
+ 0.069562,
+ 0.069562,
+ 0.080457,
+ 0.071028
+ ],
+ "warm_rows_per_second": 7258293.166997449
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4049073,
+ "unique_rows": 8200,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 513,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 513,
+ "zlib_blocks": 356,
+ "xor_blocks": 157,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8200
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "88b89ca0c6e9bfdc07772c720877973a90980360086049102eb00191b22979e2",
+ "table_sha256": "4481d38ddba5240aed2a3b28c7e491a4e8530a1776a632472f4710f284fea1a9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.37774100410752,
+ "representatives_sha256": "15e683ceadbc4d78f01e5e7e56e96b69fb47cf1ef552f21877aad4aa2835dba4",
+ "structurally_merged_rows": 5932,
+ "backend_artifact_bytes": 4050280,
+ "backend_with_key_registry_bytes": 4268758,
+ "backend_with_datrie_key_index_bytes": 4163992,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9514328140470131,
+ "mean_cosine": 0.9979344438076095,
+ "p05_cosine": 0.990552467538009,
+ "relative_l2_mean": 0.03790951931057653,
+ "relative_l2_p95": 0.1375593224122512,
+ "relative_l2_max": 0.31084382639637137,
+ "weighted_relative_l2_mean": 0.021484572239686486,
+ "weighted_relative_frobenius": 0.01135935648014331,
+ "weighted_mean_cosine": 0.9988836707971713,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9514328140470131,
+ "mean_cosine": 0.9979344438076095,
+ "p05_cosine": 0.990552467538009,
+ "relative_l2_mean": 0.03790951931057653,
+ "relative_l2_p95": 0.1375593224122512,
+ "relative_l2_max": 0.31084382639637137,
+ "weighted_relative_l2_mean": 0.021484572239686486,
+ "weighted_relative_frobenius": 0.01135935648014331,
+ "weighted_mean_cosine": 0.9988836707971713,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.69718,
+ "warm_p50_ms": 0.069562,
+ "warm_p95_ms": 0.07391629999999999,
+ "warm_p99_ms": 0.09185469999999998,
+ "samples_ms": [
+ 0.075708,
+ 0.071936,
+ 0.073403,
+ 0.073333,
+ 0.069632,
+ 0.069073,
+ 0.069003,
+ 0.07047,
+ 0.071517,
+ 0.102178,
+ 0.071448,
+ 0.070889,
+ 0.069422,
+ 0.069143,
+ 0.069562,
+ 0.068514,
+ 0.07054,
+ 0.071029,
+ 0.069562,
+ 0.069073,
+ 0.069562,
+ 0.073822,
+ 0.071936,
+ 0.071867,
+ 0.071448,
+ 0.069143,
+ 0.069074,
+ 0.069562,
+ 0.068235,
+ 0.068584,
+ 0.069562,
+ 0.069073,
+ 0.068095,
+ 0.069073,
+ 0.068584,
+ 0.069073,
+ 0.069562,
+ 0.068585,
+ 0.068584,
+ 0.068584
+ ],
+ "warm_rows_per_second": 7360340.415744228
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 512,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 7235584,
+ "artifact_bytes": 4044516,
+ "unique_rows": 8191,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 512,
+ "block_rows": 16,
+ "blocks": 512,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 512,
+ "zlib_blocks": 362,
+ "xor_blocks": 150,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8191
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "21a0609cab73c8e512b40350268db37db4abb653c280b6a31b32e699609d2d5b",
+ "table_sha256": "8a070942214cfdc28e214b09ba33aae31faa4f7bf54ad72c8132c6d848bff681",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.37253565806895494,
+ "representatives_sha256": "c9fedb493c3925ad399707e17a6a860b796aba5a8e625de6672f3ce9e4cf020c",
+ "structurally_merged_rows": 5941,
+ "backend_artifact_bytes": 4045723,
+ "backend_with_key_registry_bytes": 4264201,
+ "backend_with_datrie_key_index_bytes": 4159435,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9052723324331029,
+ "mean_cosine": 0.997872344589949,
+ "p05_cosine": 0.9904502799530508,
+ "relative_l2_mean": 0.038233152597705405,
+ "relative_l2_p95": 0.13817922336006339,
+ "relative_l2_max": 0.43421287618849197,
+ "weighted_relative_l2_mean": 0.02169441101934335,
+ "weighted_relative_frobenius": 0.01190780425189542,
+ "weighted_mean_cosine": 0.9988430147109009,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9052723324331029,
+ "mean_cosine": 0.997872344589949,
+ "p05_cosine": 0.9904502799530508,
+ "relative_l2_mean": 0.038233152597705405,
+ "relative_l2_p95": 0.13817922336006339,
+ "relative_l2_max": 0.43421287618849197,
+ "weighted_relative_l2_mean": 0.02169441101934335,
+ "weighted_relative_frobenius": 0.01190780425189542,
+ "weighted_mean_cosine": 0.9988430147109009,
+ "max_relative_norm_change": 0.0100968194450101
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.546253,
+ "warm_p50_ms": 0.07141249999999999,
+ "warm_p95_ms": 0.09389114999999999,
+ "warm_p99_ms": 0.09756388,
+ "samples_ms": [
+ 0.078501,
+ 0.075289,
+ 0.071378,
+ 0.075219,
+ 0.071936,
+ 0.07054,
+ 0.072844,
+ 0.073333,
+ 0.075289,
+ 0.075219,
+ 0.071378,
+ 0.071936,
+ 0.072425,
+ 0.072844,
+ 0.068235,
+ 0.069073,
+ 0.069493,
+ 0.069143,
+ 0.097918,
+ 0.09701,
+ 0.093727,
+ 0.07473,
+ 0.072425,
+ 0.072356,
+ 0.069981,
+ 0.068654,
+ 0.069143,
+ 0.069492,
+ 0.068654,
+ 0.069981,
+ 0.068514,
+ 0.071448,
+ 0.071448,
+ 0.071447,
+ 0.070959,
+ 0.069004,
+ 0.069492,
+ 0.069143,
+ 0.069073,
+ 0.069073
+ ],
+ "warm_rows_per_second": 7169613.162961667
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 3617792,
+ "unique_rows": 14132,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "58a8afde490d0e308f8add0008e5239d0de80c72e72d6d5ff0e13f47acc810a9",
+ "table_sha256": "891a04cf80ccfbc91b7e235af0b0a5c951005e27e177a089c05291a8bae10358",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.045193221885710955,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3618598,
+ "backend_with_key_registry_bytes": 3837076,
+ "backend_with_datrie_key_index_bytes": 3732310,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999646369165,
+ "mean_cosine": 0.9999999790364589,
+ "p05_cosine": 0.9999999731862095,
+ "relative_l2_mean": 0.0002065451703408264,
+ "relative_l2_p95": 0.00023398282848556456,
+ "relative_l2_max": 0.0002720648220136451,
+ "weighted_relative_l2_mean": 0.00020648960705042698,
+ "weighted_relative_frobenius": 0.00020887876201816394,
+ "weighted_mean_cosine": 0.9999999790633699,
+ "max_relative_norm_change": 0.0001369949692249569
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.498598,
+ "warm_p50_ms": 0.2204195,
+ "warm_p95_ms": 0.3327629999999999,
+ "warm_p99_ms": 0.35659279,
+ "samples_ms": [
+ 0.221886,
+ 0.219512,
+ 0.220978,
+ 0.22,
+ 0.220978,
+ 0.235155,
+ 0.22007,
+ 0.220489,
+ 0.219092,
+ 0.219581,
+ 0.219581,
+ 0.219581,
+ 0.21993,
+ 0.219442,
+ 0.219512,
+ 0.220001,
+ 0.220001,
+ 0.219093,
+ 0.218604,
+ 0.220071,
+ 0.219512,
+ 0.235645,
+ 0.331677,
+ 0.353397,
+ 0.23327,
+ 0.22035,
+ 0.22042,
+ 0.271124,
+ 0.318476,
+ 0.2332,
+ 0.22,
+ 0.220419,
+ 0.219512,
+ 0.220489,
+ 0.358636,
+ 0.274896,
+ 0.282927,
+ 0.27727,
+ 0.265467,
+ 0.317499
+ ],
+ "warm_rows_per_second": 2322843.4870780488
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 3408317,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 442,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 442,
+ "zlib_blocks": 0,
+ "xor_blocks": 442,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "58a8afde490d0e308f8add0008e5239d0de80c72e72d6d5ff0e13f47acc810a9",
+ "table_sha256": "95c82541f2f4d0148a5a7809366cc3421dd23b8b56deae7a8190562fb158b720",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.36331925401464105,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3409526,
+ "backend_with_key_registry_bytes": 3628004,
+ "backend_with_datrie_key_index_bytes": 3523238,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999646369165,
+ "mean_cosine": 0.9999999790364589,
+ "p05_cosine": 0.9999999731862095,
+ "relative_l2_mean": 0.0002065451703408264,
+ "relative_l2_p95": 0.00023398282848556456,
+ "relative_l2_max": 0.0002720648220136451,
+ "weighted_relative_l2_mean": 0.00020648960705042698,
+ "weighted_relative_frobenius": 0.00020887876201816394,
+ "weighted_mean_cosine": 0.9999999790633699,
+ "max_relative_norm_change": 0.0001369949692249569
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.808566,
+ "warm_p50_ms": 0.22900949999999998,
+ "warm_p95_ms": 0.23397869999999998,
+ "warm_p99_ms": 0.25061204,
+ "samples_ms": [
+ 0.233619,
+ 0.230406,
+ 0.230477,
+ 0.22894,
+ 0.22852,
+ 0.22894,
+ 0.228102,
+ 0.227473,
+ 0.22894,
+ 0.22894,
+ 0.228521,
+ 0.22894,
+ 0.22894,
+ 0.228451,
+ 0.22901,
+ 0.256877,
+ 0.231315,
+ 0.228032,
+ 0.229429,
+ 0.231384,
+ 0.2332,
+ 0.230826,
+ 0.229009,
+ 0.22908,
+ 0.230895,
+ 0.227613,
+ 0.229499,
+ 0.229428,
+ 0.229918,
+ 0.228032,
+ 0.22901,
+ 0.228032,
+ 0.240813,
+ 0.229987,
+ 0.229918,
+ 0.228032,
+ 0.228102,
+ 0.230407,
+ 0.228521,
+ 0.22894
+ ],
+ "warm_rows_per_second": 2235715.1122551686
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 3408317,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 442,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 442,
+ "zlib_blocks": 0,
+ "xor_blocks": 442,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "58a8afde490d0e308f8add0008e5239d0de80c72e72d6d5ff0e13f47acc810a9",
+ "table_sha256": "95c82541f2f4d0148a5a7809366cc3421dd23b8b56deae7a8190562fb158b720",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.39766583405435085,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3409525,
+ "backend_with_key_registry_bytes": 3628003,
+ "backend_with_datrie_key_index_bytes": 3523237,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999646369165,
+ "mean_cosine": 0.9999999790364589,
+ "p05_cosine": 0.9999999731862095,
+ "relative_l2_mean": 0.0002065451703408264,
+ "relative_l2_p95": 0.00023398282848556456,
+ "relative_l2_max": 0.0002720648220136451,
+ "weighted_relative_l2_mean": 0.00020648960705042698,
+ "weighted_relative_frobenius": 0.00020887876201816394,
+ "weighted_mean_cosine": 0.9999999790633699,
+ "max_relative_norm_change": 0.0001369949692249569
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.791944,
+ "warm_p50_ms": 0.229918,
+ "warm_p95_ms": 0.23502399999999998,
+ "warm_p99_ms": 0.24341629,
+ "samples_ms": [
+ 0.2332,
+ 0.230406,
+ 0.230826,
+ 0.230407,
+ 0.229918,
+ 0.229918,
+ 0.230407,
+ 0.230407,
+ 0.234598,
+ 0.230896,
+ 0.243607,
+ 0.230337,
+ 0.229079,
+ 0.229988,
+ 0.229498,
+ 0.229918,
+ 0.229499,
+ 0.229429,
+ 0.229918,
+ 0.229499,
+ 0.229987,
+ 0.229499,
+ 0.22901,
+ 0.230895,
+ 0.230407,
+ 0.229918,
+ 0.229917,
+ 0.243118,
+ 0.230407,
+ 0.229987,
+ 0.229987,
+ 0.229499,
+ 0.229498,
+ 0.229429,
+ 0.22894,
+ 0.22852,
+ 0.22901,
+ 0.229429,
+ 0.229429,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2226880.887968754
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 3408317,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 442,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 442,
+ "zlib_blocks": 0,
+ "xor_blocks": 442,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "58a8afde490d0e308f8add0008e5239d0de80c72e72d6d5ff0e13f47acc810a9",
+ "table_sha256": "95c82541f2f4d0148a5a7809366cc3421dd23b8b56deae7a8190562fb158b720",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3406867440789938,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3409525,
+ "backend_with_key_registry_bytes": 3628003,
+ "backend_with_datrie_key_index_bytes": 3523237,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999646369165,
+ "mean_cosine": 0.9999999790364589,
+ "p05_cosine": 0.9999999731862095,
+ "relative_l2_mean": 0.0002065451703408264,
+ "relative_l2_p95": 0.00023398282848556456,
+ "relative_l2_max": 0.0002720648220136451,
+ "weighted_relative_l2_mean": 0.00020648960705042698,
+ "weighted_relative_frobenius": 0.00020887876201816394,
+ "weighted_mean_cosine": 0.9999999790633699,
+ "max_relative_norm_change": 0.0001369949692249569
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.762191,
+ "warm_p50_ms": 0.22900949999999998,
+ "warm_p95_ms": 0.23327354999999997,
+ "warm_p99_ms": 0.24449718,
+ "samples_ms": [
+ 0.232851,
+ 0.230895,
+ 0.227613,
+ 0.22901,
+ 0.22901,
+ 0.22901,
+ 0.22894,
+ 0.24654,
+ 0.231804,
+ 0.22908,
+ 0.229498,
+ 0.228102,
+ 0.229918,
+ 0.22901,
+ 0.227613,
+ 0.22901,
+ 0.229009,
+ 0.22908,
+ 0.227612,
+ 0.228102,
+ 0.227543,
+ 0.228032,
+ 0.227962,
+ 0.228032,
+ 0.241302,
+ 0.229918,
+ 0.228521,
+ 0.227613,
+ 0.22859,
+ 0.22908,
+ 0.22901,
+ 0.22901,
+ 0.228521,
+ 0.227613,
+ 0.228521,
+ 0.229499,
+ 0.228101,
+ 0.228102,
+ 0.226635,
+ 0.229499
+ ],
+ "warm_rows_per_second": 2235715.1122551686
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2210511,
+ "unique_rows": 8749,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 274,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 274,
+ "zlib_blocks": 87,
+ "xor_blocks": 187,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8749
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "03a606c976a608624362337ad63b9de4e37052f8605cd31768a0f4f154a9ce53",
+ "table_sha256": "15c98fde434b647d340bc6a2e86d7a9a5117e86df9d4fa59230bbe4ddd7f4983",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2351542259566486,
+ "representatives_sha256": "956043b03b29a55c6ba1c8769f797812d95a669298725966e612a729bc0f1305",
+ "structurally_merged_rows": 5383,
+ "backend_artifact_bytes": 2211717,
+ "backend_with_key_registry_bytes": 2430195,
+ "backend_with_datrie_key_index_bytes": 2325429,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.990007056186391,
+ "mean_cosine": 0.9986070309700306,
+ "p05_cosine": 0.9932159084086382,
+ "relative_l2_mean": 0.03086968019594394,
+ "relative_l2_p95": 0.11652666725147992,
+ "relative_l2_max": 0.1419611647297449,
+ "weighted_relative_l2_mean": 0.017984365153411806,
+ "weighted_relative_frobenius": 0.009560715856289807,
+ "weighted_mean_cosine": 0.9992313660358744,
+ "max_relative_norm_change": 0.010105117790425226
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900091705454804,
+ "mean_cosine": 0.9986070378401536,
+ "p05_cosine": 0.9932155274809832,
+ "relative_l2_mean": 0.03074185093056735,
+ "relative_l2_p95": 0.11654768876388556,
+ "relative_l2_max": 0.14195528458548512,
+ "weighted_relative_l2_mean": 0.017827325370210853,
+ "weighted_relative_frobenius": 0.009558478461585218,
+ "weighted_mean_cosine": 0.999231380401188,
+ "max_relative_norm_change": 0.010089174934655571
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.205928,
+ "warm_p50_ms": 0.228521,
+ "warm_p95_ms": 0.23315485,
+ "warm_p99_ms": 0.24787425999999999,
+ "samples_ms": [
+ 0.232781,
+ 0.229917,
+ 0.22901,
+ 0.229987,
+ 0.22901,
+ 0.22894,
+ 0.228102,
+ 0.229499,
+ 0.22901,
+ 0.253594,
+ 0.232851,
+ 0.229498,
+ 0.228521,
+ 0.228521,
+ 0.228521,
+ 0.228032,
+ 0.228032,
+ 0.228032,
+ 0.227473,
+ 0.227962,
+ 0.228032,
+ 0.228521,
+ 0.228032,
+ 0.226636,
+ 0.228521,
+ 0.228591,
+ 0.238928,
+ 0.228521,
+ 0.22894,
+ 0.22894,
+ 0.229009,
+ 0.227543,
+ 0.229429,
+ 0.228451,
+ 0.227543,
+ 0.227124,
+ 0.228102,
+ 0.228521,
+ 0.22894,
+ 0.228032
+ ],
+ "warm_rows_per_second": 2240494.309056936
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2101413,
+ "unique_rows": 8282,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 259,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 259,
+ "zlib_blocks": 116,
+ "xor_blocks": 143,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8282
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "3a96c4256fcc02f1a5fcda34489b58d74a8e15808fe6d18b86efcddcfc8c0573",
+ "table_sha256": "96d36c87ddd5c34b913ae9b5f1b4e70fa7e2dbac86518a884510291d732ff463",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2250868040136993,
+ "representatives_sha256": "76c124d99e9a0715c870dcde178039f14872e2a220d3aa66ae8ef69cb4241b1e",
+ "structurally_merged_rows": 5850,
+ "backend_artifact_bytes": 2102620,
+ "backend_with_key_registry_bytes": 2321098,
+ "backend_with_datrie_key_index_bytes": 2216332,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.980132361987804,
+ "mean_cosine": 0.9981471850465657,
+ "p05_cosine": 0.991141640232687,
+ "relative_l2_mean": 0.03636265467885292,
+ "relative_l2_p95": 0.13322974873587912,
+ "relative_l2_max": 0.20006871358342282,
+ "weighted_relative_l2_mean": 0.02071855969197664,
+ "weighted_relative_frobenius": 0.00989450479246472,
+ "weighted_mean_cosine": 0.9990023627562143,
+ "max_relative_norm_change": 0.010138660244752193
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9801310084676192,
+ "mean_cosine": 0.9981471926948611,
+ "p05_cosine": 0.9911425797513979,
+ "relative_l2_mean": 0.03624166457588246,
+ "relative_l2_p95": 0.13323903332007203,
+ "relative_l2_max": 0.2000592916691895,
+ "weighted_relative_l2_mean": 0.020564924203437772,
+ "weighted_relative_frobenius": 0.009892340007271048,
+ "weighted_mean_cosine": 0.9990023773552794,
+ "max_relative_norm_change": 0.010098746211067064
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.711102,
+ "warm_p50_ms": 0.2287655,
+ "warm_p95_ms": 0.23453419999999997,
+ "warm_p99_ms": 0.25344405,
+ "samples_ms": [
+ 0.231873,
+ 0.230407,
+ 0.228521,
+ 0.228521,
+ 0.22901,
+ 0.229499,
+ 0.22901,
+ 0.261207,
+ 0.232782,
+ 0.229499,
+ 0.22894,
+ 0.228591,
+ 0.228032,
+ 0.228102,
+ 0.228452,
+ 0.22894,
+ 0.22894,
+ 0.22908,
+ 0.228032,
+ 0.227543,
+ 0.228032,
+ 0.229499,
+ 0.228451,
+ 0.22894,
+ 0.241302,
+ 0.232293,
+ 0.229988,
+ 0.22859,
+ 0.227613,
+ 0.227543,
+ 0.227543,
+ 0.226565,
+ 0.228521,
+ 0.231315,
+ 0.234178,
+ 0.230826,
+ 0.227543,
+ 0.227473,
+ 0.228521,
+ 0.228521
+ ],
+ "warm_rows_per_second": 2238099.7134620384
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2087813,
+ "unique_rows": 8224,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 257,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 257,
+ "zlib_blocks": 120,
+ "xor_blocks": 137,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8224
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "d07ef498733e1452d6be269c51990edcb0f3e4f12c6491ff315e2f8b71ea9467",
+ "table_sha256": "dd18f5186e790d904a239a15682d765e8ed1ac0798b9240057bffac5170bf58e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2296391341369599,
+ "representatives_sha256": "e728b6c67e5b22f9069b02a96ecb6b4e023a9c5b4c19df4e2427c9a29c3ce14f",
+ "structurally_merged_rows": 5908,
+ "backend_artifact_bytes": 2089020,
+ "backend_with_key_registry_bytes": 2307498,
+ "backend_with_datrie_key_index_bytes": 2202732,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9700701881194705,
+ "mean_cosine": 0.9980440754357444,
+ "p05_cosine": 0.9907911042391875,
+ "relative_l2_mean": 0.03728664315912571,
+ "relative_l2_p95": 0.1357957115925837,
+ "relative_l2_max": 0.24519072193375074,
+ "weighted_relative_l2_mean": 0.021207044963463435,
+ "weighted_relative_frobenius": 0.010123420958298426,
+ "weighted_mean_cosine": 0.9989473888952469,
+ "max_relative_norm_change": 0.010138660244752193
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9700687232478249,
+ "mean_cosine": 0.9980440826393321,
+ "p05_cosine": 0.9907918330818589,
+ "relative_l2_mean": 0.037166506643704786,
+ "relative_l2_p95": 0.1357704486930214,
+ "relative_l2_max": 0.2451995418904606,
+ "weighted_relative_l2_mean": 0.021053848111204372,
+ "weighted_relative_frobenius": 0.010121302767534777,
+ "weighted_mean_cosine": 0.9989474033210557,
+ "max_relative_norm_change": 0.010098746211067064
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.577147,
+ "warm_p50_ms": 0.228521,
+ "warm_p95_ms": 0.23540029999999998,
+ "warm_p99_ms": 0.24392851,
+ "samples_ms": [
+ 0.233619,
+ 0.229918,
+ 0.235086,
+ 0.22894,
+ 0.228521,
+ 0.229009,
+ 0.234108,
+ 0.241372,
+ 0.232711,
+ 0.230895,
+ 0.228521,
+ 0.227962,
+ 0.229848,
+ 0.227543,
+ 0.232851,
+ 0.22901,
+ 0.228032,
+ 0.227124,
+ 0.22859,
+ 0.234178,
+ 0.228521,
+ 0.228521,
+ 0.227543,
+ 0.226565,
+ 0.245563,
+ 0.232223,
+ 0.22859,
+ 0.228102,
+ 0.227613,
+ 0.228591,
+ 0.228521,
+ 0.228102,
+ 0.227613,
+ 0.227474,
+ 0.228451,
+ 0.228521,
+ 0.2332,
+ 0.227124,
+ 0.227613,
+ 0.227124
+ ],
+ "warm_rows_per_second": 2240494.309056936
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2080306,
+ "unique_rows": 8193,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 257,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 256,
+ "zlib_blocks": 125,
+ "xor_blocks": 131,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8193
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "b87db8420e7476028063309e419abcc76c4e7f8016bfc1cbaad613c73bd5c846",
+ "table_sha256": "ec7a5524d2373dd2bf37f093e78c7991f40cc51c8bf21ab740bc8200758eabbf",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22681321296840906,
+ "representatives_sha256": "94dd6e62f140d247194a90ba465fb0fe89b0bf82a4963c9a54f377809fcf6257",
+ "structurally_merged_rows": 5939,
+ "backend_artifact_bytes": 2081513,
+ "backend_with_key_registry_bytes": 2299991,
+ "backend_with_datrie_key_index_bytes": 2195225,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9514374218821062,
+ "mean_cosine": 0.9979321385598214,
+ "p05_cosine": 0.9905512436111426,
+ "relative_l2_mean": 0.03807581807203882,
+ "relative_l2_p95": 0.1375590599343331,
+ "relative_l2_max": 0.31082112559827385,
+ "weighted_relative_l2_mean": 0.02166062335741108,
+ "weighted_relative_frobenius": 0.011365209163908348,
+ "weighted_mean_cosine": 0.9988825164132097,
+ "max_relative_norm_change": 0.010138660244752193
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9514326597019475,
+ "mean_cosine": 0.9979321468354201,
+ "p05_cosine": 0.990549616265017,
+ "relative_l2_mean": 0.03795614209728174,
+ "relative_l2_p95": 0.13755454468164807,
+ "relative_l2_max": 0.310841627963749,
+ "weighted_relative_l2_mean": 0.021507681587830875,
+ "weighted_relative_frobenius": 0.011363347775666536,
+ "weighted_mean_cosine": 0.9988825303587557,
+ "max_relative_norm_change": 0.010098746211067064
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.828436,
+ "warm_p50_ms": 0.22894,
+ "warm_p95_ms": 0.24172444999999998,
+ "warm_p99_ms": 0.25124866,
+ "samples_ms": [
+ 0.241721,
+ 0.231803,
+ 0.230825,
+ 0.229429,
+ 0.228102,
+ 0.228591,
+ 0.24179,
+ 0.228521,
+ 0.229918,
+ 0.22887,
+ 0.22894,
+ 0.228451,
+ 0.22894,
+ 0.22894,
+ 0.22901,
+ 0.229009,
+ 0.22901,
+ 0.234179,
+ 0.228521,
+ 0.228381,
+ 0.230896,
+ 0.230896,
+ 0.233759,
+ 0.257296,
+ 0.232293,
+ 0.22894,
+ 0.22852,
+ 0.228521,
+ 0.227054,
+ 0.227544,
+ 0.23327,
+ 0.228521,
+ 0.228101,
+ 0.227543,
+ 0.233689,
+ 0.22894,
+ 0.227054,
+ 0.229498,
+ 0.22894,
+ 0.227544
+ ],
+ "warm_rows_per_second": 2236393.8149733553
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3617792,
+ "artifact_bytes": 2078125,
+ "unique_rows": 8184,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 256,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 256,
+ "zlib_blocks": 125,
+ "xor_blocks": 131,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8184
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "1e349ad3c6fd5a154e79b366f7504af81c349c3b150362d02a3039fdc4cd9ace",
+ "table_sha256": "c95ad8defc43336ab7732dc3dde94f4bc3816dc0f320adfa4a7f0ce39680e21e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22549356101080775,
+ "representatives_sha256": "72ae588e623f4aa824b05b1c4d5b6d93d1de2059eb335a7c684f32a53a74f5ce",
+ "structurally_merged_rows": 5948,
+ "backend_artifact_bytes": 2079332,
+ "backend_with_key_registry_bytes": 2297810,
+ "backend_with_datrie_key_index_bytes": 2193044,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9052630677527462,
+ "mean_cosine": 0.9978700376740262,
+ "p05_cosine": 0.9904479782806821,
+ "relative_l2_mean": 0.038399326678778566,
+ "relative_l2_p95": 0.13818826505076168,
+ "relative_l2_max": 0.43421943108487454,
+ "weighted_relative_l2_mean": 0.021870386623126264,
+ "weighted_relative_frobenius": 0.011913401642523774,
+ "weighted_mean_cosine": 0.9988418596550673,
+ "max_relative_norm_change": 0.010138660244752193
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9052623408990954,
+ "mean_cosine": 0.9978700427661503,
+ "p05_cosine": 0.9904473239689596,
+ "relative_l2_mean": 0.038279788666925,
+ "relative_l2_p95": 0.13819983197682933,
+ "relative_l2_max": 0.4342322325789638,
+ "weighted_relative_l2_mean": 0.0217175255331657,
+ "weighted_relative_frobenius": 0.011911638782710864,
+ "weighted_mean_cosine": 0.9988418721406954,
+ "max_relative_norm_change": 0.010098746211067064
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.953731,
+ "warm_p50_ms": 0.229883,
+ "warm_p95_ms": 0.3456933,
+ "warm_p99_ms": 0.37220983999999996,
+ "samples_ms": [
+ 0.351931,
+ 0.345365,
+ 0.345365,
+ 0.345365,
+ 0.34341,
+ 0.341035,
+ 0.385175,
+ 0.342083,
+ 0.229429,
+ 0.228451,
+ 0.228032,
+ 0.22887,
+ 0.234178,
+ 0.231873,
+ 0.231803,
+ 0.230337,
+ 0.227613,
+ 0.228521,
+ 0.22901,
+ 0.22901,
+ 0.228032,
+ 0.228521,
+ 0.274407,
+ 0.238438,
+ 0.232711,
+ 0.230826,
+ 0.229009,
+ 0.228521,
+ 0.229009,
+ 0.227543,
+ 0.228451,
+ 0.229429,
+ 0.287187,
+ 0.287188,
+ 0.227613,
+ 0.227543,
+ 0.227054,
+ 0.227963,
+ 0.264489,
+ 0.233619
+ ],
+ "warm_rows_per_second": 2227219.933618406
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 132,
+ "storage": {
+ "artifact_bytes": 1865424,
+ "unique_rows": 14132,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "1c37947f29f3da25c424fdd730690d3a4a1f55296f4034cf343a6892d5710e81",
+ "table_sha256": "c32e129ca5dd95b920e1532b4fa43658b2d93d5a6a71a0c7ba8b53ed4163e1c5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.03687532618641853,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1866274,
+ "backend_with_key_registry_bytes": 2084752,
+ "backend_with_datrie_key_index_bytes": 1979986,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999416546250693,
+ "mean_cosine": 0.9999789981841369,
+ "p05_cosine": 0.9999686233018587,
+ "relative_l2_mean": 0.006454883173090688,
+ "relative_l2_p95": 0.007945469397262879,
+ "relative_l2_max": 0.010833502058372456,
+ "weighted_relative_l2_mean": 0.006481248993400622,
+ "weighted_relative_frobenius": 0.00659738924794254,
+ "weighted_mean_cosine": 0.9999788190130319,
+ "max_relative_norm_change": 0.0025646664461764367
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.370928,
+ "warm_p50_ms": 0.12211749999999999,
+ "warm_p95_ms": 0.13324419999999998,
+ "warm_p99_ms": 0.15490753,
+ "samples_ms": [
+ 0.12725,
+ 0.124387,
+ 0.122991,
+ 0.122013,
+ 0.122571,
+ 0.122013,
+ 0.12348,
+ 0.139613,
+ 0.124457,
+ 0.122502,
+ 0.132,
+ 0.122013,
+ 0.121105,
+ 0.121594,
+ 0.121664,
+ 0.122152,
+ 0.122083,
+ 0.123969,
+ 0.122572,
+ 0.121523,
+ 0.122502,
+ 0.124388,
+ 0.121594,
+ 0.123968,
+ 0.124387,
+ 0.122082,
+ 0.121664,
+ 0.132909,
+ 0.122083,
+ 0.123969,
+ 0.122083,
+ 0.121524,
+ 0.122083,
+ 0.122083,
+ 0.122082,
+ 0.122572,
+ 0.122013,
+ 0.121943,
+ 0.122013,
+ 0.164686
+ ],
+ "warm_rows_per_second": 4192683.2763526933
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1724422,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 228,
+ "zlib_blocks": 0,
+ "xor_blocks": 228,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "1c37947f29f3da25c424fdd730690d3a4a1f55296f4034cf343a6892d5710e81",
+ "table_sha256": "1ac10c201ab01b98342763b408901006a5b71b11a5df1916a5290c3fb6629807",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.20558535307645798,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1725675,
+ "backend_with_key_registry_bytes": 1944153,
+ "backend_with_datrie_key_index_bytes": 1839387,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999416546250693,
+ "mean_cosine": 0.9999789981841369,
+ "p05_cosine": 0.9999686233018587,
+ "relative_l2_mean": 0.006454883173090688,
+ "relative_l2_p95": 0.007945469397262879,
+ "relative_l2_max": 0.010833502058372456,
+ "weighted_relative_l2_mean": 0.006481248993400622,
+ "weighted_relative_frobenius": 0.00659738924794254,
+ "weighted_mean_cosine": 0.9999788190130319,
+ "max_relative_norm_change": 0.0025646664461764367
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.009314,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.1425711,
+ "warm_p99_ms": 0.15757546,
+ "samples_ms": [
+ 0.165175,
+ 0.132908,
+ 0.133467,
+ 0.132,
+ 0.131092,
+ 0.132,
+ 0.130603,
+ 0.13249,
+ 0.132489,
+ 0.132,
+ 0.132489,
+ 0.131023,
+ 0.131023,
+ 0.132,
+ 0.134864,
+ 0.131511,
+ 0.130604,
+ 0.132,
+ 0.132,
+ 0.131582,
+ 0.129277,
+ 0.129695,
+ 0.130184,
+ 0.131511,
+ 0.131511,
+ 0.142407,
+ 0.137238,
+ 0.131512,
+ 0.131512,
+ 0.145689,
+ 0.132489,
+ 0.130115,
+ 0.130673,
+ 0.136749,
+ 0.134304,
+ 0.132489,
+ 0.131093,
+ 0.132001,
+ 0.131092,
+ 0.130533
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1724422,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 228,
+ "zlib_blocks": 0,
+ "xor_blocks": 228,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "1c37947f29f3da25c424fdd730690d3a4a1f55296f4034cf343a6892d5710e81",
+ "table_sha256": "1ac10c201ab01b98342763b408901006a5b71b11a5df1916a5290c3fb6629807",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2546287889126688,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1725674,
+ "backend_with_key_registry_bytes": 1944152,
+ "backend_with_datrie_key_index_bytes": 1839386,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999416546250693,
+ "mean_cosine": 0.9999789981841369,
+ "p05_cosine": 0.9999686233018587,
+ "relative_l2_mean": 0.006454883173090688,
+ "relative_l2_p95": 0.007945469397262879,
+ "relative_l2_max": 0.010833502058372456,
+ "weighted_relative_l2_mean": 0.006481248993400622,
+ "weighted_relative_frobenius": 0.00659738924794254,
+ "weighted_mean_cosine": 0.9999788190130319,
+ "max_relative_norm_change": 0.0025646664461764367
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.976279,
+ "warm_p50_ms": 0.131023,
+ "warm_p95_ms": 0.13479744999999999,
+ "warm_p99_ms": 0.16570826,
+ "samples_ms": [
+ 0.134794,
+ 0.132489,
+ 0.132489,
+ 0.131511,
+ 0.131022,
+ 0.132908,
+ 0.130604,
+ 0.131093,
+ 0.131511,
+ 0.131511,
+ 0.131511,
+ 0.130114,
+ 0.129206,
+ 0.131022,
+ 0.130114,
+ 0.131022,
+ 0.131511,
+ 0.131092,
+ 0.131092,
+ 0.128298,
+ 0.131092,
+ 0.130603,
+ 0.130045,
+ 0.131023,
+ 0.131023,
+ 0.130603,
+ 0.129137,
+ 0.185429,
+ 0.133886,
+ 0.131581,
+ 0.130603,
+ 0.134863,
+ 0.130604,
+ 0.130603,
+ 0.130603,
+ 0.132489,
+ 0.130603,
+ 0.130603,
+ 0.129625,
+ 0.134794
+ ],
+ "warm_rows_per_second": 3907710.8599253567
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1724422,
+ "unique_rows": 14132,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 228,
+ "zlib_blocks": 0,
+ "xor_blocks": 228,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 14132
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "1c37947f29f3da25c424fdd730690d3a4a1f55296f4034cf343a6892d5710e81",
+ "table_sha256": "1ac10c201ab01b98342763b408901006a5b71b11a5df1916a5290c3fb6629807",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.20552047016099095,
+ "representatives_sha256": "99f43c31125f55c54c27ed391ba505ce31e22821afc532165dd58cdc508f69fc",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1725674,
+ "backend_with_key_registry_bytes": 1944152,
+ "backend_with_datrie_key_index_bytes": 1839386,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999416546250693,
+ "mean_cosine": 0.9999789981841369,
+ "p05_cosine": 0.9999686233018587,
+ "relative_l2_mean": 0.006454883173090688,
+ "relative_l2_p95": 0.007945469397262879,
+ "relative_l2_max": 0.010833502058372456,
+ "weighted_relative_l2_mean": 0.006481248993400622,
+ "weighted_relative_frobenius": 0.00659738924794254,
+ "weighted_mean_cosine": 0.9999788190130319,
+ "max_relative_norm_change": 0.0025646664461764367
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 9.953021,
+ "warm_p50_ms": 0.1301495,
+ "warm_p95_ms": 0.1325588,
+ "warm_p99_ms": 0.14138372999999999,
+ "samples_ms": [
+ 0.133885,
+ 0.132,
+ 0.131582,
+ 0.132489,
+ 0.132,
+ 0.129625,
+ 0.131511,
+ 0.131511,
+ 0.132,
+ 0.131512,
+ 0.132,
+ 0.131022,
+ 0.131511,
+ 0.130114,
+ 0.130533,
+ 0.130533,
+ 0.130533,
+ 0.132,
+ 0.129625,
+ 0.129625,
+ 0.129625,
+ 0.129207,
+ 0.129207,
+ 0.129626,
+ 0.146178,
+ 0.132489,
+ 0.129695,
+ 0.129206,
+ 0.130185,
+ 0.128718,
+ 0.129625,
+ 0.131022,
+ 0.130114,
+ 0.129137,
+ 0.129066,
+ 0.129555,
+ 0.128648,
+ 0.129626,
+ 0.130045,
+ 0.130045
+ ],
+ "warm_rows_per_second": 3933937.510324665
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1182283,
+ "unique_rows": 8788,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 142,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 142,
+ "zlib_blocks": 0,
+ "xor_blocks": 142,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8788
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "a824e7dc6aae4c396c20672cde81a8e8b01bac06024c1aed441d90ac9326e4c1",
+ "table_sha256": "74a174e8442e68b19a8fd008323f91f7c9708a1adf0e3f06602bbe691c214dbb",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1529743520077318,
+ "representatives_sha256": "d41b79fe8448d6f9dde66b6fa9a9fa774533b3195e2a343075490556bf0899af",
+ "structurally_merged_rows": 5344,
+ "backend_artifact_bytes": 1183532,
+ "backend_with_key_registry_bytes": 1402010,
+ "backend_with_datrie_key_index_bytes": 1297244,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9899108723181153,
+ "mean_cosine": 0.9985974036524515,
+ "p05_cosine": 0.9932142209859959,
+ "relative_l2_mean": 0.03463968394059213,
+ "relative_l2_p95": 0.11650950709251426,
+ "relative_l2_max": 0.142252491783661,
+ "weighted_relative_l2_mean": 0.022743038805837456,
+ "weighted_relative_frobenius": 0.011588159396950687,
+ "weighted_mean_cosine": 0.9992156948236537,
+ "max_relative_norm_change": 0.011184456409277174
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9900065524598948,
+ "mean_cosine": 0.9986026066492123,
+ "p05_cosine": 0.9931883698886925,
+ "relative_l2_mean": 0.030733278485272696,
+ "relative_l2_p95": 0.11685027491459732,
+ "relative_l2_max": 0.14242139704155687,
+ "weighted_relative_l2_mean": 0.01787775152113955,
+ "weighted_relative_frobenius": 0.00966408415736297,
+ "weighted_mean_cosine": 0.9992268155290017,
+ "max_relative_norm_change": 0.010090977690148559
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.00801,
+ "warm_p50_ms": 0.13130199999999997,
+ "warm_p95_ms": 0.14217989999999997,
+ "warm_p99_ms": 0.15465412,
+ "samples_ms": [
+ 0.137658,
+ 0.134794,
+ 0.133816,
+ 0.132,
+ 0.131511,
+ 0.131093,
+ 0.131023,
+ 0.130533,
+ 0.147156,
+ 0.132978,
+ 0.131022,
+ 0.132977,
+ 0.132978,
+ 0.130533,
+ 0.131093,
+ 0.130044,
+ 0.130045,
+ 0.133397,
+ 0.132,
+ 0.129626,
+ 0.130114,
+ 0.141428,
+ 0.129626,
+ 0.130604,
+ 0.130115,
+ 0.131581,
+ 0.130045,
+ 0.141918,
+ 0.132,
+ 0.132,
+ 0.131092,
+ 0.130533,
+ 0.129556,
+ 0.130114,
+ 0.130114,
+ 0.130184,
+ 0.132489,
+ 0.159448,
+ 0.135353,
+ 0.132489
+ ],
+ "warm_rows_per_second": 3899407.472848853
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1130550,
+ "unique_rows": 8319,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 135,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 135,
+ "zlib_blocks": 0,
+ "xor_blocks": 135,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8319
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "e04a9bbbe6ed08fbca27d3b0986983f6b93ab737f346e518120b5b64524fb1fa",
+ "table_sha256": "8e396b9ef65e52b1f577096e33001d09d9b1c029d3f9411fa7b35f45ebd20b46",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.16777170891873538,
+ "representatives_sha256": "05d1b683b3b51ddfaacc059479c1618a42851ef44dc661ef671327ceb3b453cd",
+ "structurally_merged_rows": 5813,
+ "backend_artifact_bytes": 1131799,
+ "backend_with_key_registry_bytes": 1350277,
+ "backend_with_datrie_key_index_bytes": 1245511,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9798827077614866,
+ "mean_cosine": 0.9981339639644252,
+ "p05_cosine": 0.991126565017416,
+ "relative_l2_mean": 0.03996771697003143,
+ "relative_l2_p95": 0.13336952753736225,
+ "relative_l2_max": 0.20134611295850205,
+ "weighted_relative_l2_mean": 0.025395107328377076,
+ "weighted_relative_frobenius": 0.011866439446284533,
+ "weighted_mean_cosine": 0.9989849126118469,
+ "max_relative_norm_change": 0.011545051309837495
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9800295519620152,
+ "mean_cosine": 0.9981378180939796,
+ "p05_cosine": 0.9911077709660768,
+ "relative_l2_mean": 0.03627773874662921,
+ "relative_l2_p95": 0.133575330698548,
+ "relative_l2_max": 0.20066482239216804,
+ "weighted_relative_l2_mean": 0.020637454063190525,
+ "weighted_relative_frobenius": 0.009996923230353167,
+ "weighted_mean_cosine": 0.9989953671888233,
+ "max_relative_norm_change": 0.010090977690148559
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.401926,
+ "warm_p50_ms": 0.132,
+ "warm_p95_ms": 0.13689315,
+ "warm_p99_ms": 0.14451251999999998,
+ "samples_ms": [
+ 0.13675,
+ 0.13675,
+ 0.139613,
+ 0.134863,
+ 0.132978,
+ 0.133466,
+ 0.131092,
+ 0.131092,
+ 0.131092,
+ 0.130533,
+ 0.132489,
+ 0.132839,
+ 0.147645,
+ 0.132908,
+ 0.131022,
+ 0.131022,
+ 0.131511,
+ 0.131581,
+ 0.131511,
+ 0.131093,
+ 0.131093,
+ 0.133397,
+ 0.135771,
+ 0.133956,
+ 0.132001,
+ 0.131511,
+ 0.132489,
+ 0.133397,
+ 0.131023,
+ 0.131512,
+ 0.129556,
+ 0.133397,
+ 0.132,
+ 0.132908,
+ 0.13633,
+ 0.131511,
+ 0.131092,
+ 0.132,
+ 0.131512,
+ 0.131163
+ ],
+ "warm_rows_per_second": 3878787.8787878784
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1124396,
+ "unique_rows": 8266,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 134,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 134,
+ "zlib_blocks": 3,
+ "xor_blocks": 131,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8266
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "9ec201def617a752929d2a8bc3e6b099aaf9171a6414ac8398124787fabab7ea",
+ "table_sha256": "375a9969194d26155e350b4501a22a1ff185c0d31b3fc2b4a09f952c66eafc73",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14849828905425966,
+ "representatives_sha256": "de61c1c712da640b57e4078fb10b81abcb058b67e42210c6fed8cb23b53e1c9a",
+ "structurally_merged_rows": 5866,
+ "backend_artifact_bytes": 1125645,
+ "backend_with_key_registry_bytes": 1344123,
+ "backend_with_datrie_key_index_bytes": 1239357,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9701382468354772,
+ "mean_cosine": 0.9980390119271941,
+ "p05_cosine": 0.9908073065862646,
+ "relative_l2_mean": 0.040792527914039235,
+ "relative_l2_p95": 0.13563235078609337,
+ "relative_l2_max": 0.24497082414096052,
+ "weighted_relative_l2_mean": 0.025803739344528464,
+ "weighted_relative_frobenius": 0.011907506555993772,
+ "weighted_mean_cosine": 0.9989378709925603,
+ "max_relative_norm_change": 0.011545051309837495
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.970420856601294,
+ "mean_cosine": 0.9980426977362156,
+ "p05_cosine": 0.990776119815563,
+ "relative_l2_mean": 0.037126568315976424,
+ "relative_l2_p95": 0.13598857146171736,
+ "relative_l2_max": 0.2437800342281442,
+ "weighted_relative_l2_mean": 0.021057985508379482,
+ "weighted_relative_frobenius": 0.010045723189348644,
+ "weighted_mean_cosine": 0.9989482421793255,
+ "max_relative_norm_change": 0.010090977690148559
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.427,
+ "warm_p50_ms": 0.132489,
+ "warm_p95_ms": 0.1387293,
+ "warm_p99_ms": 0.15526165,
+ "samples_ms": [
+ 0.134305,
+ 0.132978,
+ 0.132489,
+ 0.132489,
+ 0.133886,
+ 0.140521,
+ 0.133467,
+ 0.131092,
+ 0.131511,
+ 0.132489,
+ 0.164686,
+ 0.13626,
+ 0.132908,
+ 0.131512,
+ 0.131512,
+ 0.132489,
+ 0.132489,
+ 0.13242,
+ 0.132,
+ 0.132,
+ 0.132,
+ 0.131511,
+ 0.131023,
+ 0.132908,
+ 0.129626,
+ 0.131511,
+ 0.136679,
+ 0.137238,
+ 0.136749,
+ 0.132489,
+ 0.138635,
+ 0.132908,
+ 0.13633,
+ 0.135283,
+ 0.132489,
+ 0.132978,
+ 0.131022,
+ 0.131511,
+ 0.129626,
+ 0.131511
+ ],
+ "warm_rows_per_second": 3864471.7674674885
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1120791,
+ "unique_rows": 8235,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 133,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 133,
+ "zlib_blocks": 1,
+ "xor_blocks": 132,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8235
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "7e9b1ab14ac2b25cb9b467d50e5696269bfef612d85ba7b6cdd555653548f7d6",
+ "table_sha256": "6921034e96171757cebfcf8d3354d7f70c8aa5356ed9a946eef5064fa7dfd43b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.17001997190527618,
+ "representatives_sha256": "07f3594db208298cb4d28797bce605bf0ea698d232d93c0642adcd10c670e55c",
+ "structurally_merged_rows": 5897,
+ "backend_artifact_bytes": 1122040,
+ "backend_with_key_registry_bytes": 1340518,
+ "backend_with_datrie_key_index_bytes": 1235752,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.9513417357401348,
+ "mean_cosine": 0.9979278119096603,
+ "p05_cosine": 0.9905879251663541,
+ "relative_l2_mean": 0.0415656151939831,
+ "relative_l2_p95": 0.13721894132995344,
+ "relative_l2_max": 0.3112503293611566,
+ "weighted_relative_l2_mean": 0.026248778803007623,
+ "weighted_relative_frobenius": 0.013052340782278507,
+ "weighted_mean_cosine": 0.9988733671033254,
+ "max_relative_norm_change": 0.011545051309837495
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9513428546426225,
+ "mean_cosine": 0.9979314301481493,
+ "p05_cosine": 0.990524673728532,
+ "relative_l2_mean": 0.037913106122702,
+ "relative_l2_p95": 0.13763929961460214,
+ "relative_l2_max": 0.3113047564903937,
+ "weighted_relative_l2_mean": 0.02151018282355247,
+ "weighted_relative_frobenius": 0.011379386283756232,
+ "weighted_mean_cosine": 0.998883705182836,
+ "max_relative_norm_change": 0.010090977690148559
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.614873,
+ "warm_p50_ms": 0.131791,
+ "warm_p95_ms": 0.14005999999999996,
+ "warm_p99_ms": 0.16009847,
+ "samples_ms": [
+ 0.139613,
+ 0.136749,
+ 0.138635,
+ 0.132978,
+ 0.131582,
+ 0.132,
+ 0.133397,
+ 0.133816,
+ 0.148553,
+ 0.132908,
+ 0.131092,
+ 0.131023,
+ 0.130534,
+ 0.132978,
+ 0.131511,
+ 0.131022,
+ 0.132489,
+ 0.132489,
+ 0.135772,
+ 0.131512,
+ 0.131023,
+ 0.131511,
+ 0.132489,
+ 0.132908,
+ 0.130115,
+ 0.130533,
+ 0.132,
+ 0.131022,
+ 0.135771,
+ 0.132489,
+ 0.130534,
+ 0.131512,
+ 0.129207,
+ 0.129626,
+ 0.129626,
+ 0.130534,
+ 0.130604,
+ 0.130045,
+ 0.16748,
+ 0.135283
+ ],
+ "warm_rows_per_second": 3884939.032255617
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 14132,
+ 128
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 132,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1865424,
+ "artifact_bytes": 1119674,
+ "unique_rows": 8227,
+ "row_map_bytes": 113056,
+ "n_rows": 14132,
+ "row_bytes": 132,
+ "block_rows": 62,
+ "blocks": 133,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 133,
+ "zlib_blocks": 2,
+ "xor_blocks": 131,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 8227
+ },
+ "source_float32_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c",
+ "decoded_float32_sha256": "10f98491cca5738c4fb306c1e251797790526dfe45938ba06e2ee2efe3cc4cfd",
+ "table_sha256": "bd9783a3d3c6e18c205114125d51e249e7ca7ce9ada9225c6336f26f6999f596",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1467373101040721,
+ "representatives_sha256": "60748e55b9acc28c24131d6a450e53eb3159eb3c5a10b3cce254f21bcc2ced32",
+ "structurally_merged_rows": 5905,
+ "backend_artifact_bytes": 1120923,
+ "backend_with_key_registry_bytes": 1339401,
+ "backend_with_datrie_key_index_bytes": 1234635,
+ "full_model_serving_artifact_bytes": null,
+ "fidelity_vs_float32": {
+ "rows": 14132,
+ "minimum_cosine": 0.905058789974138,
+ "mean_cosine": 0.9978704723170204,
+ "p05_cosine": 0.9904815039656957,
+ "relative_l2_mean": 0.04185979944478432,
+ "relative_l2_p95": 0.13801117982453734,
+ "relative_l2_max": 0.43474475731290524,
+ "weighted_relative_l2_mean": 0.02644353692266584,
+ "weighted_relative_frobenius": 0.013530645935659193,
+ "weighted_mean_cosine": 0.9988350872672367,
+ "max_relative_norm_change": 0.011545051309837495
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 14132,
+ "minimum_cosine": 0.9052706198580938,
+ "mean_cosine": 0.9978740643616799,
+ "p05_cosine": 0.9904604786044254,
+ "relative_l2_mean": 0.03821097608579753,
+ "relative_l2_p95": 0.13834025060732943,
+ "relative_l2_max": 0.43431985848158156,
+ "weighted_relative_l2_mean": 0.021707155072084528,
+ "weighted_relative_frobenius": 0.011924654731844517,
+ "weighted_mean_cosine": 0.9988454252007132,
+ "max_relative_norm_change": 0.010090977690148559
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.640924,
+ "warm_p50_ms": 0.131092,
+ "warm_p95_ms": 0.1376161,
+ "warm_p99_ms": 0.15888297999999998,
+ "samples_ms": [
+ 0.136261,
+ 0.135283,
+ 0.137168,
+ 0.132001,
+ 0.167899,
+ 0.13668,
+ 0.137239,
+ 0.132,
+ 0.132978,
+ 0.13242,
+ 0.133467,
+ 0.132,
+ 0.130603,
+ 0.130114,
+ 0.131092,
+ 0.132,
+ 0.131511,
+ 0.131581,
+ 0.131092,
+ 0.129206,
+ 0.130115,
+ 0.130534,
+ 0.131022,
+ 0.131022,
+ 0.131581,
+ 0.130114,
+ 0.131022,
+ 0.131092,
+ 0.131022,
+ 0.130533,
+ 0.130603,
+ 0.130603,
+ 0.129695,
+ 0.130114,
+ 0.144781,
+ 0.132489,
+ 0.131022,
+ 0.130114,
+ 0.130603,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3905654.0444878405
+ }
+ }
+ ],
+ "baseline_quality": {},
+ "training_row_sha256": "d0f54618325a4f2c9ffa3ab5779125a3b6d0c7192249f54dad23a6ea5e81304c"
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 3,
+ "rows": 11760,
+ "dim": 64,
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "source_float32_bytes": 3010560,
+ "key_registry_bytes": 161725,
+ "unchanged_backbone_artifact_bytes": 294686,
+ "alternative_datrie_key_index_bytes": 95280,
+ "query_rows": 512,
+ "query_ids_sha256": "8ba8e1ab159a822d9159f679166b91955ebc405920cda85ccf24f213c3eff2ad",
+ "frequency_trace_sha256": "73ddaf0153da6d907ac6eef82bf819aaec23a0a5eb289cb0f3f91092c37eb7f0",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 3010560,
+ "unique_rows": 11760,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "table_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.016872901935130358,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3011364,
+ "backend_with_key_registry_bytes": 3173089,
+ "backend_with_datrie_key_index_bytes": 3106644,
+ "full_model_serving_artifact_bytes": 3467775,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.372394,
+ "warm_p50_ms": 0.0487495,
+ "warm_p95_ms": 0.05017045,
+ "warm_p99_ms": 0.05178485,
+ "samples_ms": [
+ 0.05252,
+ 0.049238,
+ 0.048749,
+ 0.050146,
+ 0.050077,
+ 0.050635,
+ 0.049727,
+ 0.049658,
+ 0.049239,
+ 0.04868,
+ 0.048749,
+ 0.048749,
+ 0.048749,
+ 0.048749,
+ 0.048749,
+ 0.04819,
+ 0.048749,
+ 0.049657,
+ 0.048261,
+ 0.048261,
+ 0.04875,
+ 0.04875,
+ 0.048261,
+ 0.048261,
+ 0.048191,
+ 0.049727,
+ 0.04826,
+ 0.049238,
+ 0.04826,
+ 0.049168,
+ 0.047771,
+ 0.049238,
+ 0.048749,
+ 0.049168,
+ 0.048679,
+ 0.049238,
+ 0.04868,
+ 0.049169,
+ 0.049168,
+ 0.049657
+ ],
+ "warm_rows_per_second": 10502671.8222751
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7448185687646842,
+ "character_perplexity": 5.724862712278213,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2908740,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 368,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 368,
+ "zlib_blocks": 368,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "table_sha256": "7fd6c0632f94d7e3cbf018cb6c5053182ccef84cb8d20764c97cde11eb40c13f",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2654925410170108,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2909946,
+ "backend_with_key_registry_bytes": 3071671,
+ "backend_with_datrie_key_index_bytes": 3005226,
+ "full_model_serving_artifact_bytes": 3366357,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.765056,
+ "warm_p50_ms": 0.058666,
+ "warm_p95_ms": 0.06246595,
+ "warm_p99_ms": 0.07769494999999998,
+ "samples_ms": [
+ 0.06146,
+ 0.060972,
+ 0.059574,
+ 0.059085,
+ 0.059575,
+ 0.058667,
+ 0.087092,
+ 0.060483,
+ 0.060971,
+ 0.062438,
+ 0.058177,
+ 0.05727,
+ 0.060483,
+ 0.060623,
+ 0.060483,
+ 0.0572,
+ 0.057759,
+ 0.062997,
+ 0.058667,
+ 0.057689,
+ 0.057759,
+ 0.058666,
+ 0.057689,
+ 0.05727,
+ 0.056851,
+ 0.056781,
+ 0.05727,
+ 0.059574,
+ 0.058667,
+ 0.058248,
+ 0.057689,
+ 0.056781,
+ 0.058666,
+ 0.059156,
+ 0.058248,
+ 0.058666,
+ 0.056711,
+ 0.056292,
+ 0.057689,
+ 0.05727
+ ],
+ "warm_rows_per_second": 8727371.90195343
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7448185687646842,
+ "character_perplexity": 5.724862712278213,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2908740,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 368,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 368,
+ "zlib_blocks": 368,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "table_sha256": "7fd6c0632f94d7e3cbf018cb6c5053182ccef84cb8d20764c97cde11eb40c13f",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3065303419716656,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2909945,
+ "backend_with_key_registry_bytes": 3071670,
+ "backend_with_datrie_key_index_bytes": 3005225,
+ "full_model_serving_artifact_bytes": 3366356,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.735233,
+ "warm_p50_ms": 0.0586315,
+ "warm_p95_ms": 0.06439385,
+ "warm_p99_ms": 0.07321912,
+ "samples_ms": [
+ 0.078013,
+ 0.065721,
+ 0.060483,
+ 0.060063,
+ 0.058666,
+ 0.058667,
+ 0.058248,
+ 0.058597,
+ 0.057689,
+ 0.057759,
+ 0.058597,
+ 0.058666,
+ 0.064324,
+ 0.061042,
+ 0.058667,
+ 0.057689,
+ 0.058178,
+ 0.058178,
+ 0.058178,
+ 0.060552,
+ 0.061879,
+ 0.061111,
+ 0.058666,
+ 0.058178,
+ 0.061461,
+ 0.060134,
+ 0.058178,
+ 0.058178,
+ 0.061042,
+ 0.057759,
+ 0.058178,
+ 0.057759,
+ 0.057689,
+ 0.057689,
+ 0.056362,
+ 0.060552,
+ 0.060064,
+ 0.058178,
+ 0.058667,
+ 0.056292
+ ],
+ "warm_rows_per_second": 8732507.26998286
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7448185687646842,
+ "character_perplexity": 5.724862712278213,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2908740,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 368,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 368,
+ "zlib_blocks": 368,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "table_sha256": "7fd6c0632f94d7e3cbf018cb6c5053182ccef84cb8d20764c97cde11eb40c13f",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.26189110218547285,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2909945,
+ "backend_with_key_registry_bytes": 3071670,
+ "backend_with_datrie_key_index_bytes": 3005225,
+ "full_model_serving_artifact_bytes": 3366356,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.714909,
+ "warm_p50_ms": 0.05727,
+ "warm_p95_ms": 0.05976374999999999,
+ "warm_p99_ms": 0.07620616,
+ "samples_ms": [
+ 0.06202,
+ 0.059645,
+ 0.059156,
+ 0.059644,
+ 0.059574,
+ 0.059644,
+ 0.059505,
+ 0.059226,
+ 0.058108,
+ 0.058667,
+ 0.056782,
+ 0.0572,
+ 0.058247,
+ 0.058178,
+ 0.0572,
+ 0.05727,
+ 0.058666,
+ 0.056781,
+ 0.056851,
+ 0.056781,
+ 0.056781,
+ 0.056362,
+ 0.056781,
+ 0.05727,
+ 0.05727,
+ 0.056711,
+ 0.05727,
+ 0.056712,
+ 0.0572,
+ 0.05727,
+ 0.056781,
+ 0.056781,
+ 0.056293,
+ 0.0572,
+ 0.05727,
+ 0.05727,
+ 0.085276,
+ 0.059086,
+ 0.056781,
+ 0.05727
+ ],
+ "warm_rows_per_second": 8940108.25912345
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7448185687646842,
+ "character_perplexity": 5.724862712278213,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2732409,
+ "unique_rows": 11022,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11022
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "6cf3900d9818e1125b3dc91dcd6f874c00edb8cfdc7a079205e0cffde4c1b6ad",
+ "table_sha256": "2567f4b0454d6af57ca2f55ba6cd49f3222af1b279da2463c2cc3c5b13d1728a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.28853103797882795,
+ "representatives_sha256": "5371f13a257c93dc8bd3ba977ad2b4ad230151b800443e84d16b0f013d279495",
+ "structurally_merged_rows": 738,
+ "backend_artifact_bytes": 2733614,
+ "backend_with_key_registry_bytes": 2895339,
+ "backend_with_datrie_key_index_bytes": 2828894,
+ "full_model_serving_artifact_bytes": 3190025,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9902119474867733,
+ "mean_cosine": 0.9999571484538186,
+ "p05_cosine": 0.9997077550982506,
+ "relative_l2_mean": 0.002177182218231527,
+ "relative_l2_p95": 0.0243536770022943,
+ "relative_l2_max": 0.14017569822550982,
+ "weighted_relative_l2_mean": 0.0008977103354020354,
+ "weighted_relative_frobenius": 0.0052940877733579465,
+ "weighted_mean_cosine": 0.9999852803166117,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9902119474867733,
+ "mean_cosine": 0.9999571484538186,
+ "p05_cosine": 0.9997077550982506,
+ "relative_l2_mean": 0.002177182218231527,
+ "relative_l2_p95": 0.0243536770022943,
+ "relative_l2_max": 0.14017569822550982,
+ "weighted_relative_l2_mean": 0.0008977103354020354,
+ "weighted_relative_frobenius": 0.0052940877733579465,
+ "weighted_mean_cosine": 0.9999852803166117,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.001969,
+ "warm_p50_ms": 0.05727,
+ "warm_p95_ms": 0.05971479999999999,
+ "warm_p99_ms": 0.06679269,
+ "samples_ms": [
+ 0.061041,
+ 0.059085,
+ 0.059645,
+ 0.058178,
+ 0.059574,
+ 0.059155,
+ 0.059156,
+ 0.058737,
+ 0.059085,
+ 0.059155,
+ 0.058667,
+ 0.058178,
+ 0.05727,
+ 0.056781,
+ 0.0572,
+ 0.07047,
+ 0.059644,
+ 0.058667,
+ 0.058666,
+ 0.05727,
+ 0.057201,
+ 0.056781,
+ 0.058247,
+ 0.057759,
+ 0.05727,
+ 0.05727,
+ 0.0572,
+ 0.056712,
+ 0.056781,
+ 0.056292,
+ 0.057201,
+ 0.05727,
+ 0.057201,
+ 0.056781,
+ 0.056851,
+ 0.0572,
+ 0.056781,
+ 0.0572,
+ 0.058108,
+ 0.05727
+ ],
+ "warm_rows_per_second": 8940108.25912345
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744775876674702,
+ "character_perplexity": 5.7246183111411995,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.269117868789314e-05,
+ "mean_prediction_kl_nats": 1.4920202650035391e-06,
+ "prediction_kl_p99_nats": 2.4167480924158577e-05,
+ "maximum_prediction_kl_nats": 0.005274976667000143,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29146748781204224,
+ "maximum_per_target_cross_entropy_increase": 0.03230166435241699
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.269117868789314e-05,
+ "mean_prediction_kl_nats": 1.4920202650035391e-06,
+ "prediction_kl_p99_nats": 2.4167480924158577e-05,
+ "maximum_prediction_kl_nats": 0.005274976667000143,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29146748781204224,
+ "maximum_per_target_cross_entropy_increase": 0.03230166435241699
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2732175,
+ "unique_rows": 11021,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11021
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "6546cea75e4ce2644b34c2e2427d96e5a68cb4262381e713192aee48b5976c70",
+ "table_sha256": "4c31862bd204934f3613922f9772e5ebd35dc5a689f8d7ad9d504a359c29a4fe",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2955526781734079,
+ "representatives_sha256": "19694405b05c56a8d71896197ae0b471d362baf069fc0566c431ad753cc50974",
+ "structurally_merged_rows": 739,
+ "backend_artifact_bytes": 2733380,
+ "backend_with_key_registry_bytes": 2895105,
+ "backend_with_datrie_key_index_bytes": 2828660,
+ "full_model_serving_artifact_bytes": 3189791,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9837543867790927,
+ "mean_cosine": 0.9999557670241228,
+ "p05_cosine": 0.9997070350940703,
+ "relative_l2_mean": 0.0021925403769333758,
+ "relative_l2_p95": 0.02447099841182412,
+ "relative_l2_max": 0.18061194633374375,
+ "weighted_relative_l2_mean": 0.0008977103354020354,
+ "weighted_relative_frobenius": 0.0052940877733579465,
+ "weighted_mean_cosine": 0.9999852803166117,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9837543867790927,
+ "mean_cosine": 0.9999557670241228,
+ "p05_cosine": 0.9997070350940703,
+ "relative_l2_mean": 0.0021925403769333758,
+ "relative_l2_p95": 0.02447099841182412,
+ "relative_l2_max": 0.18061194633374375,
+ "weighted_relative_l2_mean": 0.0008977103354020354,
+ "weighted_relative_frobenius": 0.0052940877733579465,
+ "weighted_mean_cosine": 0.9999852803166117,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.937574,
+ "warm_p50_ms": 0.0582125,
+ "warm_p95_ms": 0.0606928,
+ "warm_p99_ms": 0.06342505,
+ "samples_ms": [
+ 0.060623,
+ 0.060483,
+ 0.059226,
+ 0.058178,
+ 0.058667,
+ 0.060133,
+ 0.058666,
+ 0.058737,
+ 0.058667,
+ 0.058667,
+ 0.058178,
+ 0.057688,
+ 0.057689,
+ 0.057759,
+ 0.058178,
+ 0.057758,
+ 0.057689,
+ 0.058247,
+ 0.058667,
+ 0.058667,
+ 0.058666,
+ 0.058248,
+ 0.057689,
+ 0.062019,
+ 0.058597,
+ 0.064324,
+ 0.058178,
+ 0.056222,
+ 0.057759,
+ 0.058178,
+ 0.056781,
+ 0.058248,
+ 0.058178,
+ 0.059156,
+ 0.05727,
+ 0.057689,
+ 0.05727,
+ 0.05727,
+ 0.056781,
+ 0.059645
+ ],
+ "warm_rows_per_second": 8795361.820914753
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744775876674702,
+ "character_perplexity": 5.7246183111411995,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.269117868789314e-05,
+ "mean_prediction_kl_nats": 1.4920202650035391e-06,
+ "prediction_kl_p99_nats": 2.4167480924158577e-05,
+ "maximum_prediction_kl_nats": 0.005274976667000143,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29146748781204224,
+ "maximum_per_target_cross_entropy_increase": 0.03230166435241699
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.269117868789314e-05,
+ "mean_prediction_kl_nats": 1.4920202650035391e-06,
+ "prediction_kl_p99_nats": 2.4167480924158577e-05,
+ "maximum_prediction_kl_nats": 0.005274976667000143,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29146748781204224,
+ "maximum_per_target_cross_entropy_increase": 0.03230166435241699
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2731766,
+ "unique_rows": 11019,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11019
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "9cdfdee066702234e8f4c084e8f0eca32f8cdbadbc2444baced796bfb7506b98",
+ "table_sha256": "4416d44102413d97608dc33b22ce9eb461383ea9ee746b930dba00f5d13b4154",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.33247158606536686,
+ "representatives_sha256": "ef9dbb10d591cf7deb74509522f592834311f7886ae73e5fdb843b1be1d00ddc",
+ "structurally_merged_rows": 741,
+ "backend_artifact_bytes": 2732971,
+ "backend_with_key_registry_bytes": 2894696,
+ "backend_with_datrie_key_index_bytes": 2828251,
+ "full_model_serving_artifact_bytes": 3189382,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9732942287303055,
+ "mean_cosine": 0.9999489751159576,
+ "p05_cosine": 0.9997031993162947,
+ "relative_l2_mean": 0.002249751455277053,
+ "relative_l2_p95": 0.024490715699937732,
+ "relative_l2_max": 0.23161634472212844,
+ "weighted_relative_l2_mean": 0.0008977103354020354,
+ "weighted_relative_frobenius": 0.0052940877733579465,
+ "weighted_mean_cosine": 0.9999852803166117,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9732942287303055,
+ "mean_cosine": 0.9999489751159576,
+ "p05_cosine": 0.9997031993162947,
+ "relative_l2_mean": 0.002249751455277053,
+ "relative_l2_p95": 0.024490715699937732,
+ "relative_l2_max": 0.23161634472212844,
+ "weighted_relative_l2_mean": 0.0008977103354020354,
+ "weighted_relative_frobenius": 0.0052940877733579465,
+ "weighted_mean_cosine": 0.9999852803166117,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.272464,
+ "warm_p50_ms": 0.058667,
+ "warm_p95_ms": 0.06201884999999999,
+ "warm_p99_ms": 0.07753276999999999,
+ "samples_ms": [
+ 0.063346,
+ 0.060971,
+ 0.059225,
+ 0.059086,
+ 0.059156,
+ 0.058177,
+ 0.058667,
+ 0.058248,
+ 0.058178,
+ 0.058178,
+ 0.058247,
+ 0.058108,
+ 0.058666,
+ 0.058666,
+ 0.058667,
+ 0.058247,
+ 0.058178,
+ 0.058667,
+ 0.060552,
+ 0.059156,
+ 0.058667,
+ 0.058666,
+ 0.057689,
+ 0.058667,
+ 0.086603,
+ 0.059574,
+ 0.059225,
+ 0.056851,
+ 0.061949,
+ 0.060063,
+ 0.058667,
+ 0.058248,
+ 0.058178,
+ 0.058178,
+ 0.058178,
+ 0.05727,
+ 0.056292,
+ 0.061531,
+ 0.061112,
+ 0.060064
+ ],
+ "warm_rows_per_second": 8727223.14077761
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744775876674702,
+ "character_perplexity": 5.7246183111411995,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.269117868789314e-05,
+ "mean_prediction_kl_nats": 1.4920202650035391e-06,
+ "prediction_kl_p99_nats": 2.4167480924158577e-05,
+ "maximum_prediction_kl_nats": 0.005274976667000143,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29146748781204224,
+ "maximum_per_target_cross_entropy_increase": 0.03230166435241699
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.269117868789314e-05,
+ "mean_prediction_kl_nats": 1.4920202650035391e-06,
+ "prediction_kl_p99_nats": 2.4167480924158577e-05,
+ "maximum_prediction_kl_nats": 0.005274976667000143,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29146748781204224,
+ "maximum_per_target_cross_entropy_increase": 0.03230166435241699
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2731011,
+ "unique_rows": 11016,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11016
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "d9f4ff5c93b675a11e75a64012210089806f67f607a4f3e4e2e2e85387ad54db",
+ "table_sha256": "3527027d88a1b20f04a9b9fd658ee9b9eb440e71672b9b6515041a90a4c5b460",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25376729504205287,
+ "representatives_sha256": "e309956b4ee73e9916fb8c464f341e509a700946436d6e62b17fcf80ce54fdbc",
+ "structurally_merged_rows": 744,
+ "backend_artifact_bytes": 2732216,
+ "backend_with_key_registry_bytes": 2893941,
+ "backend_with_datrie_key_index_bytes": 2827496,
+ "full_model_serving_artifact_bytes": 3188627,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9601687946809417,
+ "mean_cosine": 0.999939420382366,
+ "p05_cosine": 0.9997003626652597,
+ "relative_l2_mean": 0.0023194666962265153,
+ "relative_l2_p95": 0.02470920431965932,
+ "relative_l2_max": 0.2817572783926916,
+ "weighted_relative_l2_mean": 0.0009328916835455479,
+ "weighted_relative_frobenius": 0.006194876167627727,
+ "weighted_mean_cosine": 0.9999803967664052,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9601687946809417,
+ "mean_cosine": 0.999939420382366,
+ "p05_cosine": 0.9997003626652597,
+ "relative_l2_mean": 0.0023194666962265153,
+ "relative_l2_p95": 0.02470920431965932,
+ "relative_l2_max": 0.2817572783926916,
+ "weighted_relative_l2_mean": 0.0009328916835455479,
+ "weighted_relative_frobenius": 0.006194876167627727,
+ "weighted_mean_cosine": 0.9999803967664052,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.239429,
+ "warm_p50_ms": 0.05727,
+ "warm_p95_ms": 0.061094199999999994,
+ "warm_p99_ms": 0.07811394999999999,
+ "samples_ms": [
+ 0.059156,
+ 0.060552,
+ 0.063416,
+ 0.060972,
+ 0.059715,
+ 0.058597,
+ 0.058178,
+ 0.056781,
+ 0.0572,
+ 0.056781,
+ 0.0572,
+ 0.0572,
+ 0.05727,
+ 0.0572,
+ 0.05727,
+ 0.056851,
+ 0.05727,
+ 0.05727,
+ 0.056362,
+ 0.05727,
+ 0.0572,
+ 0.056292,
+ 0.05727,
+ 0.056781,
+ 0.056712,
+ 0.058178,
+ 0.087511,
+ 0.060553,
+ 0.058666,
+ 0.056222,
+ 0.058178,
+ 0.055803,
+ 0.056292,
+ 0.056781,
+ 0.056711,
+ 0.0572,
+ 0.057759,
+ 0.057759,
+ 0.058737,
+ 0.057689
+ ],
+ "warm_rows_per_second": 8940108.25912345
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744914414661637,
+ "character_perplexity": 5.725411443176186,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 9.585049031779498e-05,
+ "mean_prediction_kl_nats": 2.047808473627732e-05,
+ "prediction_kl_p99_nats": 2.5371232452318965e-05,
+ "maximum_prediction_kl_nats": 0.07808918028685391,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.7477188110351562,
+ "maximum_per_target_cross_entropy_increase": 0.7326599359512329
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 9.585049031779498e-05,
+ "mean_prediction_kl_nats": 2.047808473627732e-05,
+ "prediction_kl_p99_nats": 2.5371232452318965e-05,
+ "maximum_prediction_kl_nats": 0.07808918028685391,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.7477188110351562,
+ "maximum_per_target_cross_entropy_increase": 0.7326599359512329
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2730517,
+ "unique_rows": 11014,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11014
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "07d3c6bde8799f757a6cd2c8b7fc5c6af26686f9470a270392ea00717ead9319",
+ "table_sha256": "47ea53506ce896d19ab1f81929ebf7c54d34616813ab228b077aa83d41289e87",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2493922938592732,
+ "representatives_sha256": "7107e68ec326b0530464a531e9230aa1faa8377c353b27c623ae49607fc34677",
+ "structurally_merged_rows": 746,
+ "backend_artifact_bytes": 2731722,
+ "backend_with_key_registry_bytes": 2893447,
+ "backend_with_datrie_key_index_bytes": 2827002,
+ "full_model_serving_artifact_bytes": 3188133,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9180803831583273,
+ "mean_cosine": 0.9999261414854912,
+ "p05_cosine": 0.9996995621760043,
+ "relative_l2_mean": 0.002386593403605294,
+ "relative_l2_p95": 0.024880453595176592,
+ "relative_l2_max": 0.4052832618653121,
+ "weighted_relative_l2_mean": 0.001579007164529692,
+ "weighted_relative_frobenius": 0.012000382636823912,
+ "weighted_mean_cosine": 0.9998503171798633,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9180803831583273,
+ "mean_cosine": 0.9999261414854912,
+ "p05_cosine": 0.9996995621760043,
+ "relative_l2_mean": 0.002386593403605294,
+ "relative_l2_p95": 0.024880453595176592,
+ "relative_l2_max": 0.4052832618653121,
+ "weighted_relative_l2_mean": 0.001579007164529692,
+ "weighted_relative_frobenius": 0.012000382636823912,
+ "weighted_mean_cosine": 0.9998503171798633,
+ "max_relative_norm_change": 0.010005118677483
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.146191,
+ "warm_p50_ms": 0.057689,
+ "warm_p95_ms": 0.060157449999999994,
+ "warm_p99_ms": 0.06807802999999998,
+ "samples_ms": [
+ 0.060133,
+ 0.060063,
+ 0.057689,
+ 0.057689,
+ 0.072845,
+ 0.059645,
+ 0.057759,
+ 0.057759,
+ 0.058177,
+ 0.0572,
+ 0.0572,
+ 0.05727,
+ 0.058178,
+ 0.058178,
+ 0.057758,
+ 0.05727,
+ 0.057689,
+ 0.057759,
+ 0.05727,
+ 0.056781,
+ 0.0572,
+ 0.0572,
+ 0.057759,
+ 0.057269,
+ 0.05727,
+ 0.056362,
+ 0.058108,
+ 0.056292,
+ 0.060622,
+ 0.059994,
+ 0.058247,
+ 0.058597,
+ 0.057758,
+ 0.056711,
+ 0.056711,
+ 0.056223,
+ 0.05727,
+ 0.058666,
+ 0.057689,
+ 0.05727
+ ],
+ "warm_rows_per_second": 8875175.510062577
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7452017749241024,
+ "character_perplexity": 5.727056935324424,
+ "next_character_accuracy": 0.5,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0003832795922784004,
+ "mean_prediction_kl_nats": 7.977040469276395e-05,
+ "prediction_kl_p99_nats": 0.0002471613945079618,
+ "maximum_prediction_kl_nats": 0.07808918028685391,
+ "top1_changed_fraction": 0.0005367274957828554,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.7477188110351562,
+ "maximum_per_target_cross_entropy_increase": 0.7326599359512329
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0003832795922784004,
+ "mean_prediction_kl_nats": 7.977040469276395e-05,
+ "prediction_kl_p99_nats": 0.0002471613945079618,
+ "maximum_prediction_kl_nats": 0.07808918028685391,
+ "top1_changed_fraction": 0.0005367274957828554,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.7477188110351562,
+ "maximum_per_target_cross_entropy_increase": 0.7326599359512329
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 128,
+ "storage": {
+ "artifact_bytes": 1505280,
+ "unique_rows": 11760,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "d87571fa9337131f33612c6b040cbda96d6836e255a7c1cac1b5b07b4f12bd42",
+ "table_sha256": "080d6efb61408847148fbcc3c917499b8c29966c124b53c22e9aeca49ffbf968",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.03281928994692862,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1506085,
+ "backend_with_key_registry_bytes": 1667810,
+ "backend_with_datrie_key_index_bytes": 1601365,
+ "full_model_serving_artifact_bytes": 1962496,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999589501113,
+ "mean_cosine": 0.9999999792496315,
+ "p05_cosine": 0.9999999715554034,
+ "relative_l2_mean": 0.00020645360269819725,
+ "relative_l2_p95": 0.00024229765914773267,
+ "relative_l2_max": 0.000297725575888126,
+ "weighted_relative_l2_mean": 0.0001896498664291317,
+ "weighted_relative_frobenius": 0.00020927156759669916,
+ "weighted_mean_cosine": 0.9999999809188298,
+ "max_relative_norm_change": 0.00017913432724310765
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.394115,
+ "warm_p50_ms": 0.13318750000000001,
+ "warm_p95_ms": 0.1390331,
+ "warm_p99_ms": 0.16223252,
+ "samples_ms": [
+ 0.135352,
+ 0.132489,
+ 0.133885,
+ 0.172229,
+ 0.137658,
+ 0.134305,
+ 0.132489,
+ 0.133397,
+ 0.132978,
+ 0.133955,
+ 0.132978,
+ 0.132419,
+ 0.133397,
+ 0.132489,
+ 0.133956,
+ 0.132419,
+ 0.138146,
+ 0.132489,
+ 0.132419,
+ 0.132489,
+ 0.132978,
+ 0.132489,
+ 0.133956,
+ 0.132489,
+ 0.138635,
+ 0.133886,
+ 0.131511,
+ 0.132909,
+ 0.133397,
+ 0.132489,
+ 0.132908,
+ 0.133467,
+ 0.146597,
+ 0.134794,
+ 0.133955,
+ 0.131581,
+ 0.132978,
+ 0.133816,
+ 0.134375,
+ 0.132489
+ ],
+ "warm_rows_per_second": 3844204.5987799154
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7448141552828185,
+ "character_perplexity": 5.724837445756205,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.413472126263009e-06,
+ "mean_prediction_kl_nats": 4.48062978350823e-09,
+ "prediction_kl_p99_nats": 3.0549903725941194e-08,
+ "maximum_prediction_kl_nats": 1.9802621999273275e-07,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001046895980834961,
+ "maximum_per_target_cross_entropy_increase": 0.0007162094116210938
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1491621,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 184,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 184,
+ "zlib_blocks": 184,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "d87571fa9337131f33612c6b040cbda96d6836e255a7c1cac1b5b07b4f12bd42",
+ "table_sha256": "6374e32276827cec1fe4c3e58baa23a84a6803254c713026972f2f0768e8c818",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.15144531498663127,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1492828,
+ "backend_with_key_registry_bytes": 1654553,
+ "backend_with_datrie_key_index_bytes": 1588108,
+ "full_model_serving_artifact_bytes": 1949239,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999589501113,
+ "mean_cosine": 0.9999999792496315,
+ "p05_cosine": 0.9999999715554034,
+ "relative_l2_mean": 0.00020645360269819725,
+ "relative_l2_p95": 0.00024229765914773267,
+ "relative_l2_max": 0.000297725575888126,
+ "weighted_relative_l2_mean": 0.0001896498664291317,
+ "weighted_relative_frobenius": 0.00020927156759669916,
+ "weighted_mean_cosine": 0.9999999809188298,
+ "max_relative_norm_change": 0.00017913432724310765
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.872925,
+ "warm_p50_ms": 0.14146350000000002,
+ "warm_p95_ms": 0.14627579999999998,
+ "warm_p99_ms": 0.16738487999999999,
+ "samples_ms": [
+ 0.144712,
+ 0.141848,
+ 0.141918,
+ 0.156095,
+ 0.140939,
+ 0.141988,
+ 0.14101,
+ 0.141498,
+ 0.141429,
+ 0.141918,
+ 0.14101,
+ 0.14094,
+ 0.140032,
+ 0.141079,
+ 0.141429,
+ 0.141499,
+ 0.141499,
+ 0.140521,
+ 0.141499,
+ 0.141918,
+ 0.141498,
+ 0.14094,
+ 0.142407,
+ 0.140032,
+ 0.14094,
+ 0.140521,
+ 0.14094,
+ 0.141918,
+ 0.141988,
+ 0.141498,
+ 0.174603,
+ 0.145759,
+ 0.141428,
+ 0.141987,
+ 0.141359,
+ 0.140032,
+ 0.139613,
+ 0.140102,
+ 0.141499,
+ 0.14094
+ ],
+ "warm_rows_per_second": 3619308.160762316
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7448141552828185,
+ "character_perplexity": 5.724837445756205,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.413472126263009e-06,
+ "mean_prediction_kl_nats": 4.48062978350823e-09,
+ "prediction_kl_p99_nats": 3.0549903725941194e-08,
+ "maximum_prediction_kl_nats": 1.9802621999273275e-07,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001046895980834961,
+ "maximum_per_target_cross_entropy_increase": 0.0007162094116210938
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1491621,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 184,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 184,
+ "zlib_blocks": 184,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "d87571fa9337131f33612c6b040cbda96d6836e255a7c1cac1b5b07b4f12bd42",
+ "table_sha256": "6374e32276827cec1fe4c3e58baa23a84a6803254c713026972f2f0768e8c818",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.19145602802745998,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1492827,
+ "backend_with_key_registry_bytes": 1654552,
+ "backend_with_datrie_key_index_bytes": 1588107,
+ "full_model_serving_artifact_bytes": 1949238,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999589501113,
+ "mean_cosine": 0.9999999792496315,
+ "p05_cosine": 0.9999999715554034,
+ "relative_l2_mean": 0.00020645360269819725,
+ "relative_l2_p95": 0.00024229765914773267,
+ "relative_l2_max": 0.000297725575888126,
+ "weighted_relative_l2_mean": 0.0001896498664291317,
+ "weighted_relative_frobenius": 0.00020927156759669916,
+ "weighted_mean_cosine": 0.9999999809188298,
+ "max_relative_norm_change": 0.00017913432724310765
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.83081,
+ "warm_p50_ms": 0.1403115,
+ "warm_p95_ms": 0.14467264999999999,
+ "warm_p99_ms": 0.16306738999999998,
+ "samples_ms": [
+ 0.144292,
+ 0.141429,
+ 0.141918,
+ 0.141498,
+ 0.14094,
+ 0.141499,
+ 0.141499,
+ 0.151905,
+ 0.141918,
+ 0.141498,
+ 0.14094,
+ 0.140521,
+ 0.14059,
+ 0.140032,
+ 0.140032,
+ 0.140032,
+ 0.140031,
+ 0.139613,
+ 0.140939,
+ 0.139543,
+ 0.140031,
+ 0.139613,
+ 0.139054,
+ 0.140032,
+ 0.139543,
+ 0.139543,
+ 0.141429,
+ 0.139543,
+ 0.139543,
+ 0.141429,
+ 0.139543,
+ 0.140032,
+ 0.140102,
+ 0.140032,
+ 0.139543,
+ 0.170204,
+ 0.141918,
+ 0.140521,
+ 0.139613,
+ 0.14101
+ ],
+ "warm_rows_per_second": 3649023.779234061
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7448141552828185,
+ "character_perplexity": 5.724837445756205,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.413472126263009e-06,
+ "mean_prediction_kl_nats": 4.48062978350823e-09,
+ "prediction_kl_p99_nats": 3.0549903725941194e-08,
+ "maximum_prediction_kl_nats": 1.9802621999273275e-07,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001046895980834961,
+ "maximum_per_target_cross_entropy_increase": 0.0007162094116210938
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1491621,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 184,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 184,
+ "zlib_blocks": 184,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "d87571fa9337131f33612c6b040cbda96d6836e255a7c1cac1b5b07b4f12bd42",
+ "table_sha256": "6374e32276827cec1fe4c3e58baa23a84a6803254c713026972f2f0768e8c818",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1503708059899509,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1492827,
+ "backend_with_key_registry_bytes": 1654552,
+ "backend_with_datrie_key_index_bytes": 1588107,
+ "full_model_serving_artifact_bytes": 1949238,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999589501113,
+ "mean_cosine": 0.9999999792496315,
+ "p05_cosine": 0.9999999715554034,
+ "relative_l2_mean": 0.00020645360269819725,
+ "relative_l2_p95": 0.00024229765914773267,
+ "relative_l2_max": 0.000297725575888126,
+ "weighted_relative_l2_mean": 0.0001896498664291317,
+ "weighted_relative_frobenius": 0.00020927156759669916,
+ "weighted_mean_cosine": 0.9999999809188298,
+ "max_relative_norm_change": 0.00017913432724310765
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.864404,
+ "warm_p50_ms": 0.1409745,
+ "warm_p95_ms": 0.14347580000000001,
+ "warm_p99_ms": 0.15073919,
+ "samples_ms": [
+ 0.145201,
+ 0.142826,
+ 0.140521,
+ 0.141009,
+ 0.14094,
+ 0.143385,
+ 0.141918,
+ 0.141079,
+ 0.141429,
+ 0.142476,
+ 0.140032,
+ 0.140032,
+ 0.141499,
+ 0.141498,
+ 0.141429,
+ 0.141429,
+ 0.14094,
+ 0.140032,
+ 0.14094,
+ 0.141429,
+ 0.141499,
+ 0.141429,
+ 0.140521,
+ 0.141918,
+ 0.140521,
+ 0.141498,
+ 0.15428,
+ 0.142407,
+ 0.14052,
+ 0.140521,
+ 0.14101,
+ 0.140032,
+ 0.140032,
+ 0.140521,
+ 0.140521,
+ 0.14052,
+ 0.140521,
+ 0.14052,
+ 0.140521,
+ 0.139543
+ ],
+ "warm_rows_per_second": 3631862.499955666
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7448141552828185,
+ "character_perplexity": 5.724837445756205,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.413472126263009e-06,
+ "mean_prediction_kl_nats": 4.48062978350823e-09,
+ "prediction_kl_p99_nats": 3.0549903725941194e-08,
+ "maximum_prediction_kl_nats": 1.9802621999273275e-07,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001046895980834961,
+ "maximum_per_target_cross_entropy_increase": 0.0007162094116210938
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404390,
+ "unique_rows": 11021,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11021
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "50a8e89da6a04132e68a9177c2801d3c8005d4de0c9d1bd4900df2d6599169ed",
+ "table_sha256": "65d6ac9b7c9a13ac811d321eb32c45152a51f7654e70c055337f4a1927e37bcc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14685038197785616,
+ "representatives_sha256": "9dae2e8d991c0156bb6ef113d07f5a8b76cf2d8848bd9336bc01b46acc0690f2",
+ "structurally_merged_rows": 739,
+ "backend_artifact_bytes": 1405595,
+ "backend_with_key_registry_bytes": 1567320,
+ "backend_with_datrie_key_index_bytes": 1500875,
+ "full_model_serving_artifact_bytes": 1862006,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9902142136992986,
+ "mean_cosine": 0.9999570241423484,
+ "p05_cosine": 0.9997069046578577,
+ "relative_l2_mean": 0.002374939974054794,
+ "relative_l2_p95": 0.02446510401433875,
+ "relative_l2_max": 0.14016502249394194,
+ "weighted_relative_l2_mean": 0.0010836185932586516,
+ "weighted_relative_frobenius": 0.005302751473250229,
+ "weighted_mean_cosine": 0.9999851985547283,
+ "max_relative_norm_change": 0.01006602117159563
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9902144854182734,
+ "mean_cosine": 0.9999570398592639,
+ "p05_cosine": 0.999706774862023,
+ "relative_l2_mean": 0.0021816109990323834,
+ "relative_l2_p95": 0.024439317221512102,
+ "relative_l2_max": 0.1401612936749213,
+ "weighted_relative_l2_mean": 0.0009004591305039904,
+ "weighted_relative_frobenius": 0.005299381939658822,
+ "weighted_mean_cosine": 0.9999852152415863,
+ "max_relative_norm_change": 0.010091674015100499
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.546905,
+ "warm_p50_ms": 0.141429,
+ "warm_p95_ms": 0.14389729999999998,
+ "warm_p99_ms": 0.16187839999999998,
+ "samples_ms": [
+ 0.145689,
+ 0.143733,
+ 0.143803,
+ 0.14094,
+ 0.14101,
+ 0.140451,
+ 0.141987,
+ 0.141429,
+ 0.141918,
+ 0.143244,
+ 0.141009,
+ 0.142337,
+ 0.141499,
+ 0.14094,
+ 0.140032,
+ 0.140521,
+ 0.141498,
+ 0.141429,
+ 0.14094,
+ 0.141009,
+ 0.14101,
+ 0.141429,
+ 0.141498,
+ 0.172229,
+ 0.143315,
+ 0.141429,
+ 0.141988,
+ 0.141987,
+ 0.14094,
+ 0.141499,
+ 0.140521,
+ 0.140521,
+ 0.14101,
+ 0.141358,
+ 0.14101,
+ 0.141429,
+ 0.141498,
+ 0.140032,
+ 0.14101,
+ 0.140939
+ ],
+ "warm_rows_per_second": 3620191.0499261115
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7447695886431893,
+ "character_perplexity": 5.724582314674034,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.8978921988153566e-05,
+ "mean_prediction_kl_nats": 1.5003648247710951e-06,
+ "prediction_kl_p99_nats": 2.4263089393849343e-05,
+ "maximum_prediction_kl_nats": 0.005281839625376909,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29156380891799927,
+ "maximum_per_target_cross_entropy_increase": 0.03223705291748047
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.4565646551129405e-05,
+ "mean_prediction_kl_nats": 1.4950874560512e-06,
+ "prediction_kl_p99_nats": 2.4156584127195344e-05,
+ "maximum_prediction_kl_nats": 0.005281004070749687,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29161351919174194,
+ "maximum_per_target_cross_entropy_increase": 0.0321347713470459
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404305,
+ "unique_rows": 11020,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11020
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "29825b4a87a28b2025ca51c27e3cae5b6a3bfcd0c05afdad9b26976522531df5",
+ "table_sha256": "6fe923c28cbf5215b28698d8f27a6bc756f6dcd846bf687e0a8148ea5c93b4a1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14199619903229177,
+ "representatives_sha256": "f0ec87af3587e337c1c4c8edb2557719b558517a81d43d2a606f47308b7cb7fe",
+ "structurally_merged_rows": 740,
+ "backend_artifact_bytes": 1405510,
+ "backend_with_key_registry_bytes": 1567235,
+ "backend_with_datrie_key_index_bytes": 1500790,
+ "full_model_serving_artifact_bytes": 1861921,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9837543408059701,
+ "mean_cosine": 0.9999556427102196,
+ "p05_cosine": 0.9997066138329546,
+ "relative_l2_mean": 0.0023902789999247576,
+ "relative_l2_p95": 0.024466222694217606,
+ "relative_l2_max": 0.1806133927282591,
+ "weighted_relative_l2_mean": 0.0010836185932586516,
+ "weighted_relative_frobenius": 0.005302751473250229,
+ "weighted_mean_cosine": 0.9999851985547283,
+ "max_relative_norm_change": 0.01006602117159563
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9837551288575086,
+ "mean_cosine": 0.9999556584926701,
+ "p05_cosine": 0.9997063715966455,
+ "relative_l2_mean": 0.002196970119194705,
+ "relative_l2_p95": 0.024462612396548088,
+ "relative_l2_max": 0.18062325310889688,
+ "weighted_relative_l2_mean": 0.0009004591305039904,
+ "weighted_relative_frobenius": 0.005299381939658822,
+ "weighted_mean_cosine": 0.9999852152415863,
+ "max_relative_norm_change": 0.010091674015100499
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.499623,
+ "warm_p50_ms": 0.14101,
+ "warm_p95_ms": 0.14609409999999998,
+ "warm_p99_ms": 0.16184239,
+ "samples_ms": [
+ 0.145689,
+ 0.143873,
+ 0.142476,
+ 0.142337,
+ 0.141918,
+ 0.142895,
+ 0.141009,
+ 0.140451,
+ 0.140521,
+ 0.153791,
+ 0.141988,
+ 0.14101,
+ 0.139543,
+ 0.14101,
+ 0.14052,
+ 0.14094,
+ 0.141498,
+ 0.141009,
+ 0.141499,
+ 0.14101,
+ 0.140032,
+ 0.140032,
+ 0.140032,
+ 0.141918,
+ 0.139612,
+ 0.14101,
+ 0.141499,
+ 0.141498,
+ 0.141429,
+ 0.140101,
+ 0.14101,
+ 0.14101,
+ 0.14094,
+ 0.140521,
+ 0.141428,
+ 0.140521,
+ 0.140521,
+ 0.16699,
+ 0.144711,
+ 0.141498
+ ],
+ "warm_rows_per_second": 3630948.159704985
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7447695886431893,
+ "character_perplexity": 5.724582314674034,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.8978921988153566e-05,
+ "mean_prediction_kl_nats": 1.5003648247710951e-06,
+ "prediction_kl_p99_nats": 2.4263089393849343e-05,
+ "maximum_prediction_kl_nats": 0.005281839625376909,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29156380891799927,
+ "maximum_per_target_cross_entropy_increase": 0.03223705291748047
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.4565646551129405e-05,
+ "mean_prediction_kl_nats": 1.4950874560512e-06,
+ "prediction_kl_p99_nats": 2.4156584127195344e-05,
+ "maximum_prediction_kl_nats": 0.005281004070749687,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29161351919174194,
+ "maximum_per_target_cross_entropy_increase": 0.0321347713470459
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404121,
+ "unique_rows": 11018,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11018
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "8537fda851a863a3a84029a9a35810e06ec5375cec16d553c57aea00962d9e0f",
+ "table_sha256": "f120fb1d0d68e8d285e8011efd6ac02134d293f62c57e01123df89b484ffaf92",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1425056920852512,
+ "representatives_sha256": "9c79b707e53fa8c6b68a2764aed4fc2e78871fde3072fd85b38371484c85b693",
+ "structurally_merged_rows": 742,
+ "backend_artifact_bytes": 1405326,
+ "backend_with_key_registry_bytes": 1567051,
+ "backend_with_datrie_key_index_bytes": 1500606,
+ "full_model_serving_artifact_bytes": 1861737,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9732970565929171,
+ "mean_cosine": 0.9999488516192019,
+ "p05_cosine": 0.9997014529554792,
+ "relative_l2_mean": 0.002447449892319762,
+ "relative_l2_p95": 0.024615140375736734,
+ "relative_l2_max": 0.23160677521428877,
+ "weighted_relative_l2_mean": 0.0010836185932586516,
+ "weighted_relative_frobenius": 0.005302751473250229,
+ "weighted_mean_cosine": 0.9999851985547283,
+ "max_relative_norm_change": 0.01006602117159563
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9732851261570991,
+ "mean_cosine": 0.9999488651117413,
+ "p05_cosine": 0.9997016279208365,
+ "relative_l2_mean": 0.002254190340158536,
+ "relative_l2_p95": 0.024584885109046436,
+ "relative_l2_max": 0.23163816052831235,
+ "weighted_relative_l2_mean": 0.0009004591305039904,
+ "weighted_relative_frobenius": 0.005299381939658822,
+ "weighted_mean_cosine": 0.9999852152415863,
+ "max_relative_norm_change": 0.010091674015100499
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.541178,
+ "warm_p50_ms": 0.1405205,
+ "warm_p95_ms": 0.14438294999999998,
+ "warm_p99_ms": 0.14938222999999998,
+ "samples_ms": [
+ 0.144362,
+ 0.141917,
+ 0.141498,
+ 0.141918,
+ 0.139613,
+ 0.141917,
+ 0.139543,
+ 0.139543,
+ 0.140591,
+ 0.140101,
+ 0.140032,
+ 0.140032,
+ 0.152324,
+ 0.144292,
+ 0.141917,
+ 0.140521,
+ 0.140451,
+ 0.139613,
+ 0.140101,
+ 0.140102,
+ 0.14094,
+ 0.142476,
+ 0.144781,
+ 0.142825,
+ 0.140521,
+ 0.14101,
+ 0.140101,
+ 0.140521,
+ 0.139474,
+ 0.140521,
+ 0.140032,
+ 0.14101,
+ 0.140032,
+ 0.140032,
+ 0.14052,
+ 0.140521,
+ 0.140451,
+ 0.139543,
+ 0.139612,
+ 0.140102
+ ],
+ "warm_rows_per_second": 3643596.485921983
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7447695886431893,
+ "character_perplexity": 5.724582314674034,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.8978921988153566e-05,
+ "mean_prediction_kl_nats": 1.5003648247710951e-06,
+ "prediction_kl_p99_nats": 2.4263089393849343e-05,
+ "maximum_prediction_kl_nats": 0.005281839625376909,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29156380891799927,
+ "maximum_per_target_cross_entropy_increase": 0.03223705291748047
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.4565646551129405e-05,
+ "mean_prediction_kl_nats": 1.4950874560512e-06,
+ "prediction_kl_p99_nats": 2.4156584127195344e-05,
+ "maximum_prediction_kl_nats": 0.005281004070749687,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.29161351919174194,
+ "maximum_per_target_cross_entropy_increase": 0.0321347713470459
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1403739,
+ "unique_rows": 11015,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11015
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "e51ee71e2228cb2bae4d140947b94f98089beb87d8b3da0238138235aae919ec",
+ "table_sha256": "367b1f82a4867aabb4b2fe45ee290ce4f2b35a47a9d747a475ae28d5dc4e452a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1405726228840649,
+ "representatives_sha256": "6bd3fe3deae00366f6cd75d1d9d6f11fba4db36859d0dd3ee02f2410ca6d3924",
+ "structurally_merged_rows": 745,
+ "backend_artifact_bytes": 1404944,
+ "backend_with_key_registry_bytes": 1566669,
+ "backend_with_datrie_key_index_bytes": 1500224,
+ "full_model_serving_artifact_bytes": 1861355,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.960155144819022,
+ "mean_cosine": 0.999939296270149,
+ "p05_cosine": 0.9997002578366734,
+ "relative_l2_mean": 0.0025171110293148212,
+ "relative_l2_p95": 0.024786348941689802,
+ "relative_l2_max": 0.28180178923424,
+ "weighted_relative_l2_mean": 0.0011187750817993856,
+ "weighted_relative_frobenius": 0.006202467465180739,
+ "weighted_mean_cosine": 0.9999803141270506,
+ "max_relative_norm_change": 0.01006602117159563
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9601560142271234,
+ "mean_cosine": 0.9999393090213079,
+ "p05_cosine": 0.9997002081246212,
+ "relative_l2_mean": 0.0023239118823211095,
+ "relative_l2_p95": 0.024743986042518068,
+ "relative_l2_max": 0.2818126270142,
+ "weighted_relative_l2_mean": 0.0009356451784137656,
+ "weighted_relative_frobenius": 0.006199542749492902,
+ "weighted_mean_cosine": 0.9999803306355224,
+ "max_relative_norm_change": 0.010091674015100499
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.450384,
+ "warm_p50_ms": 0.14149850000000003,
+ "warm_p95_ms": 0.14665644999999994,
+ "warm_p99_ms": 0.16485845999999998,
+ "samples_ms": [
+ 0.145759,
+ 0.143873,
+ 0.163708,
+ 0.145689,
+ 0.142406,
+ 0.141918,
+ 0.144362,
+ 0.143315,
+ 0.142826,
+ 0.145689,
+ 0.143873,
+ 0.143314,
+ 0.140451,
+ 0.14101,
+ 0.140521,
+ 0.14101,
+ 0.139612,
+ 0.141918,
+ 0.141499,
+ 0.141009,
+ 0.141009,
+ 0.139543,
+ 0.140032,
+ 0.141499,
+ 0.14094,
+ 0.14094,
+ 0.139613,
+ 0.140032,
+ 0.140032,
+ 0.165594,
+ 0.145759,
+ 0.141917,
+ 0.144362,
+ 0.14101,
+ 0.139543,
+ 0.141987,
+ 0.141429,
+ 0.141498,
+ 0.140521,
+ 0.14101
+ ],
+ "warm_rows_per_second": 3618412.9160379786
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7449081665379969,
+ "character_perplexity": 5.725375670209355,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 8.960178731304851e-05,
+ "mean_prediction_kl_nats": 2.0500548492147738e-05,
+ "prediction_kl_p99_nats": 2.5340378747971894e-05,
+ "maximum_prediction_kl_nats": 0.07817244889218078,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.7477798461914062,
+ "maximum_per_target_cross_entropy_increase": 0.7331861257553101
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 9.40156743748698e-05,
+ "mean_prediction_kl_nats": 2.0480700068925953e-05,
+ "prediction_kl_p99_nats": 2.5365348219468226e-05,
+ "maximum_prediction_kl_nats": 0.07807902524619609,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.7478551864624023,
+ "maximum_per_target_cross_entropy_increase": 0.7328281402587891
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1403425,
+ "unique_rows": 11013,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 172,
+ "zlib_blocks": 172,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11013
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "3a46124e28abdca5775ac021026a983e1e9350a898fc0c14c4cd53e040fba201",
+ "table_sha256": "28a8198f4dcf9ea36a8bd4fc23c634f230f93309d74d90de5fca62247f568e33",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1806480789091438,
+ "representatives_sha256": "0a681ee7bd1e11d04748c0880d3282b57af0a234af552328129aa424dfc43422",
+ "structurally_merged_rows": 747,
+ "backend_artifact_bytes": 1404630,
+ "backend_with_key_registry_bytes": 1566355,
+ "backend_with_datrie_key_index_bytes": 1499910,
+ "full_model_serving_artifact_bytes": 1861041,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9180787725868748,
+ "mean_cosine": 0.9999260172138794,
+ "p05_cosine": 0.999697042210066,
+ "relative_l2_mean": 0.002584209073013027,
+ "relative_l2_p95": 0.024874615513097185,
+ "relative_l2_max": 0.40529397587175664,
+ "weighted_relative_l2_mean": 0.0017645788534315207,
+ "weighted_relative_frobenius": 0.012004533806314922,
+ "weighted_mean_cosine": 0.9998502321939515,
+ "max_relative_norm_change": 0.01006602117159563
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9180784295526778,
+ "mean_cosine": 0.9999260305791812,
+ "p05_cosine": 0.9996978534310647,
+ "relative_l2_mean": 0.0023910388290498586,
+ "relative_l2_p95": 0.024873201335340264,
+ "relative_l2_max": 0.40530101614691605,
+ "weighted_relative_l2_mean": 0.0015817841165046895,
+ "weighted_relative_frobenius": 0.012002918545574105,
+ "weighted_mean_cosine": 0.9998502493311484,
+ "max_relative_norm_change": 0.010091674015100499
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.464074,
+ "warm_p50_ms": 0.1426505,
+ "warm_p95_ms": 0.1932726,
+ "warm_p99_ms": 0.21362054999999996,
+ "samples_ms": [
+ 0.143873,
+ 0.142895,
+ 0.142406,
+ 0.14101,
+ 0.140032,
+ 0.141498,
+ 0.140032,
+ 0.14094,
+ 0.140451,
+ 0.153791,
+ 0.184381,
+ 0.175092,
+ 0.188782,
+ 0.196324,
+ 0.183124,
+ 0.185429,
+ 0.188851,
+ 0.18927,
+ 0.224679,
+ 0.193112,
+ 0.18452,
+ 0.183544,
+ 0.179283,
+ 0.18075,
+ 0.187804,
+ 0.173137,
+ 0.144223,
+ 0.139124,
+ 0.139054,
+ 0.139124,
+ 0.139124,
+ 0.140031,
+ 0.140032,
+ 0.140521,
+ 0.139124,
+ 0.139124,
+ 0.139543,
+ 0.140521,
+ 0.139124,
+ 0.140032
+ ],
+ "warm_rows_per_second": 3589191.7658893587
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7451954546843569,
+ "character_perplexity": 5.7270207390659404,
+ "next_character_accuracy": 0.5,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00037695695009420585,
+ "mean_prediction_kl_nats": 7.977080127673468e-05,
+ "prediction_kl_p99_nats": 0.0002477179325109346,
+ "maximum_prediction_kl_nats": 0.07817244889218078,
+ "top1_changed_fraction": 0.0006134028523232633,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.7477798461914062,
+ "maximum_per_target_cross_entropy_increase": 0.7331861257553101
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00038137210539562576,
+ "mean_prediction_kl_nats": 7.974733484039327e-05,
+ "prediction_kl_p99_nats": 0.00024725658206455503,
+ "maximum_prediction_kl_nats": 0.07807902524619609,
+ "top1_changed_fraction": 0.0005367274957828554,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.7478551864624023,
+ "maximum_per_target_cross_entropy_increase": 0.7328281402587891
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 68,
+ "storage": {
+ "artifact_bytes": 799680,
+ "unique_rows": 11760,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "11d317c88faf40d248e5062c81409488c5bb22cad5013acb677fd2fa7b6038a7",
+ "table_sha256": "49e03d9e3f51e955fe94dc7e04708d1b097f5f5abbc78925511998693be3084e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.015065478160977364,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 800527,
+ "backend_with_key_registry_bytes": 962252,
+ "backend_with_datrie_key_index_bytes": 895807,
+ "full_model_serving_artifact_bytes": 1256938,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999473909108884,
+ "mean_cosine": 0.9999870639260287,
+ "p05_cosine": 0.9999756364724784,
+ "relative_l2_mean": 0.004980035071369554,
+ "relative_l2_p95": 0.007019963381426742,
+ "relative_l2_max": 0.01031362099466868,
+ "weighted_relative_l2_mean": 0.005052883220721706,
+ "weighted_relative_frobenius": 0.005508949576832854,
+ "weighted_mean_cosine": 0.9999857367458814,
+ "max_relative_norm_change": 0.0028045305141336367
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.352839,
+ "warm_p50_ms": 0.098407,
+ "warm_p95_ms": 0.10171345,
+ "warm_p99_ms": 0.11197704,
+ "samples_ms": [
+ 0.102178,
+ 0.099314,
+ 0.099314,
+ 0.100292,
+ 0.099384,
+ 0.098826,
+ 0.099315,
+ 0.098407,
+ 0.097498,
+ 0.097917,
+ 0.097917,
+ 0.097498,
+ 0.09701,
+ 0.09701,
+ 0.098895,
+ 0.097987,
+ 0.098476,
+ 0.098477,
+ 0.098896,
+ 0.097988,
+ 0.098406,
+ 0.098406,
+ 0.097917,
+ 0.098895,
+ 0.09659,
+ 0.097498,
+ 0.098336,
+ 0.097917,
+ 0.097429,
+ 0.097918,
+ 0.099315,
+ 0.098407,
+ 0.099385,
+ 0.098896,
+ 0.097918,
+ 0.10085,
+ 0.098407,
+ 0.099314,
+ 0.118242,
+ 0.101689
+ ],
+ "warm_rows_per_second": 5202881.908807301
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744925143976087,
+ "character_perplexity": 5.725472873245465,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00010658089074226226,
+ "mean_prediction_kl_nats": 3.530829573772787e-06,
+ "prediction_kl_p99_nats": 2.241467108712739e-05,
+ "maximum_prediction_kl_nats": 8.750250229179642e-05,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": 0.0005367274957828738,
+ "maximum_absolute_logit_change": 0.02865290641784668,
+ "maximum_per_target_cross_entropy_increase": 0.021081924438476562
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 893714,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 98,
+ "max_block_bytes": 8192,
+ "raw_blocks": 15,
+ "compressed_blocks": 83,
+ "zlib_blocks": 83,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "11d317c88faf40d248e5062c81409488c5bb22cad5013acb677fd2fa7b6038a7",
+ "table_sha256": "bc3d36b6926f94ea6c377e48cd3ed142b78e879cd954ae468fb87cc40de42f3c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08310471288859844,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 894960,
+ "backend_with_key_registry_bytes": 1056685,
+ "backend_with_datrie_key_index_bytes": 990240,
+ "full_model_serving_artifact_bytes": 1351371,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999473909108884,
+ "mean_cosine": 0.9999870639260287,
+ "p05_cosine": 0.9999756364724784,
+ "relative_l2_mean": 0.004980035071369554,
+ "relative_l2_p95": 0.007019963381426742,
+ "relative_l2_max": 0.01031362099466868,
+ "weighted_relative_l2_mean": 0.005052883220721706,
+ "weighted_relative_frobenius": 0.005508949576832854,
+ "weighted_mean_cosine": 0.9999857367458814,
+ "max_relative_norm_change": 0.0028045305141336367
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.767161,
+ "warm_p50_ms": 0.1112225,
+ "warm_p95_ms": 0.115049,
+ "warm_p99_ms": 0.11549008999999999,
+ "samples_ms": [
+ 0.115448,
+ 0.113562,
+ 0.112026,
+ 0.110629,
+ 0.11035,
+ 0.110629,
+ 0.110209,
+ 0.113073,
+ 0.111676,
+ 0.111258,
+ 0.110209,
+ 0.109301,
+ 0.111676,
+ 0.110768,
+ 0.111676,
+ 0.112584,
+ 0.115517,
+ 0.114539,
+ 0.111188,
+ 0.111607,
+ 0.113073,
+ 0.11014,
+ 0.110769,
+ 0.111607,
+ 0.111257,
+ 0.110699,
+ 0.109721,
+ 0.111257,
+ 0.111676,
+ 0.112096,
+ 0.110699,
+ 0.11021,
+ 0.110279,
+ 0.112584,
+ 0.115028,
+ 0.110699,
+ 0.11014,
+ 0.110768,
+ 0.110768,
+ 0.110209
+ ],
+ "warm_rows_per_second": 4603385.1064308025
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744925143976087,
+ "character_perplexity": 5.725472873245465,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00010658089074226226,
+ "mean_prediction_kl_nats": 3.530829573772787e-06,
+ "prediction_kl_p99_nats": 2.241467108712739e-05,
+ "maximum_prediction_kl_nats": 8.750250229179642e-05,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": 0.0005367274957828738,
+ "maximum_absolute_logit_change": 0.02865290641784668,
+ "maximum_per_target_cross_entropy_increase": 0.021081924438476562
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 893714,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 98,
+ "max_block_bytes": 8192,
+ "raw_blocks": 15,
+ "compressed_blocks": 83,
+ "zlib_blocks": 83,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "11d317c88faf40d248e5062c81409488c5bb22cad5013acb677fd2fa7b6038a7",
+ "table_sha256": "bc3d36b6926f94ea6c377e48cd3ed142b78e879cd954ae468fb87cc40de42f3c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.12294375686906278,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 894959,
+ "backend_with_key_registry_bytes": 1056684,
+ "backend_with_datrie_key_index_bytes": 990239,
+ "full_model_serving_artifact_bytes": 1351370,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999473909108884,
+ "mean_cosine": 0.9999870639260287,
+ "p05_cosine": 0.9999756364724784,
+ "relative_l2_mean": 0.004980035071369554,
+ "relative_l2_p95": 0.007019963381426742,
+ "relative_l2_max": 0.01031362099466868,
+ "weighted_relative_l2_mean": 0.005052883220721706,
+ "weighted_relative_frobenius": 0.005508949576832854,
+ "weighted_mean_cosine": 0.9999857367458814,
+ "max_relative_norm_change": 0.0028045305141336367
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.733567,
+ "warm_p50_ms": 0.1102795,
+ "warm_p95_ms": 0.11222789999999999,
+ "warm_p99_ms": 0.12965448999999998,
+ "samples_ms": [
+ 0.113423,
+ 0.112165,
+ 0.111117,
+ 0.111117,
+ 0.110279,
+ 0.111746,
+ 0.11021,
+ 0.11021,
+ 0.111747,
+ 0.112165,
+ 0.10979,
+ 0.109721,
+ 0.140032,
+ 0.111258,
+ 0.109302,
+ 0.11021,
+ 0.110698,
+ 0.110349,
+ 0.110628,
+ 0.11021,
+ 0.11028,
+ 0.110699,
+ 0.109721,
+ 0.110279,
+ 0.109232,
+ 0.109651,
+ 0.11035,
+ 0.110769,
+ 0.108743,
+ 0.10986,
+ 0.108394,
+ 0.11028,
+ 0.111676,
+ 0.109232,
+ 0.110209,
+ 0.11028,
+ 0.109302,
+ 0.11021,
+ 0.110768,
+ 0.10979
+ ],
+ "warm_rows_per_second": 4642748.652288049
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744925143976087,
+ "character_perplexity": 5.725472873245465,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00010658089074226226,
+ "mean_prediction_kl_nats": 3.530829573772787e-06,
+ "prediction_kl_p99_nats": 2.241467108712739e-05,
+ "maximum_prediction_kl_nats": 8.750250229179642e-05,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": 0.0005367274957828738,
+ "maximum_absolute_logit_change": 0.02865290641784668,
+ "maximum_per_target_cross_entropy_increase": 0.021081924438476562
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 893714,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 98,
+ "max_block_bytes": 8192,
+ "raw_blocks": 15,
+ "compressed_blocks": 83,
+ "zlib_blocks": 83,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "11d317c88faf40d248e5062c81409488c5bb22cad5013acb677fd2fa7b6038a7",
+ "table_sha256": "bc3d36b6926f94ea6c377e48cd3ed142b78e879cd954ae468fb87cc40de42f3c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08123408304527402,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 894959,
+ "backend_with_key_registry_bytes": 1056684,
+ "backend_with_datrie_key_index_bytes": 990239,
+ "full_model_serving_artifact_bytes": 1351370,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999473909108884,
+ "mean_cosine": 0.9999870639260287,
+ "p05_cosine": 0.9999756364724784,
+ "relative_l2_mean": 0.004980035071369554,
+ "relative_l2_p95": 0.007019963381426742,
+ "relative_l2_max": 0.01031362099466868,
+ "weighted_relative_l2_mean": 0.005052883220721706,
+ "weighted_relative_frobenius": 0.005508949576832854,
+ "weighted_mean_cosine": 0.9999857367458814,
+ "max_relative_norm_change": 0.0028045305141336367
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.708983,
+ "warm_p50_ms": 0.1106985,
+ "warm_p95_ms": 0.11400545,
+ "warm_p99_ms": 0.13964835999999997,
+ "samples_ms": [
+ 0.11447,
+ 0.113981,
+ 0.111187,
+ 0.110629,
+ 0.110769,
+ 0.112165,
+ 0.111676,
+ 0.111257,
+ 0.110698,
+ 0.110698,
+ 0.111187,
+ 0.110768,
+ 0.110768,
+ 0.11021,
+ 0.108953,
+ 0.10972,
+ 0.110698,
+ 0.109721,
+ 0.108882,
+ 0.155746,
+ 0.113562,
+ 0.110768,
+ 0.111676,
+ 0.110209,
+ 0.110279,
+ 0.110769,
+ 0.11021,
+ 0.111118,
+ 0.11028,
+ 0.109301,
+ 0.109791,
+ 0.110699,
+ 0.11021,
+ 0.109791,
+ 0.110279,
+ 0.112165,
+ 0.11021,
+ 0.110698,
+ 0.111676,
+ 0.111676
+ ],
+ "warm_rows_per_second": 4625175.5895518
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744925143976087,
+ "character_perplexity": 5.725472873245465,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00010658089074226226,
+ "mean_prediction_kl_nats": 3.530829573772787e-06,
+ "prediction_kl_p99_nats": 2.241467108712739e-05,
+ "maximum_prediction_kl_nats": 8.750250229179642e-05,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": 0.0005367274957828738,
+ "maximum_absolute_logit_change": 0.02865290641784668,
+ "maximum_per_target_cross_entropy_increase": 0.021081924438476562
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 843341,
+ "unique_rows": 11026,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 10,
+ "compressed_blocks": 82,
+ "zlib_blocks": 82,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11026
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "889ecd64b0c80353367f03cabfa6be39ca55699423581564d950fa8a65639032",
+ "table_sha256": "0e0317b5cc07cf1fc6c1de549fe20fdfd88ff2b5f632e6bcafc51903bca7fefa",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08627795591019094,
+ "representatives_sha256": "9dd523868af05e2ec81f6ada0eb4a1996ecb07459d057849fdece7eadb2c4760",
+ "structurally_merged_rows": 734,
+ "backend_artifact_bytes": 844585,
+ "backend_with_key_registry_bytes": 1006310,
+ "backend_with_datrie_key_index_bytes": 939865,
+ "full_model_serving_artifact_bytes": 1300996,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9925185716726764,
+ "mean_cosine": 0.9999452723868479,
+ "p05_cosine": 0.9997098827945491,
+ "relative_l2_mean": 0.006869948350346231,
+ "relative_l2_p95": 0.024423295005078234,
+ "relative_l2_max": 0.12213938394952205,
+ "weighted_relative_l2_mean": 0.005803078514374146,
+ "weighted_relative_frobenius": 0.007636228340675911,
+ "weighted_mean_cosine": 0.9999712258731059,
+ "max_relative_norm_change": 0.010522080816159867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9924329975314833,
+ "mean_cosine": 0.9999567840970592,
+ "p05_cosine": 0.9996946934776181,
+ "relative_l2_mean": 0.002197937842712327,
+ "relative_l2_p95": 0.024897731630796352,
+ "relative_l2_max": 0.12282019564119777,
+ "weighted_relative_l2_mean": 0.0009324724100068205,
+ "weighted_relative_frobenius": 0.005520060633454694,
+ "weighted_mean_cosine": 0.9999845157529222,
+ "max_relative_norm_change": 0.009943799209470403
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.77659,
+ "warm_p50_ms": 0.110279,
+ "warm_p95_ms": 0.11504995,
+ "warm_p99_ms": 0.13300074999999997,
+ "samples_ms": [
+ 0.115029,
+ 0.114051,
+ 0.111117,
+ 0.111676,
+ 0.110768,
+ 0.10979,
+ 0.111118,
+ 0.112095,
+ 0.111257,
+ 0.110769,
+ 0.11021,
+ 0.109302,
+ 0.110768,
+ 0.109232,
+ 0.109791,
+ 0.11028,
+ 0.144223,
+ 0.115448,
+ 0.109721,
+ 0.110628,
+ 0.110279,
+ 0.11021,
+ 0.109721,
+ 0.110279,
+ 0.110628,
+ 0.11014,
+ 0.110349,
+ 0.11021,
+ 0.109791,
+ 0.110279,
+ 0.111606,
+ 0.110209,
+ 0.110699,
+ 0.108883,
+ 0.11021,
+ 0.10986,
+ 0.108813,
+ 0.109232,
+ 0.11028,
+ 0.110279
+ ],
+ "warm_rows_per_second": 4642769.702300528
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.74488879921546,
+ "character_perplexity": 5.725264786085873,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 7.023291699165417e-05,
+ "mean_prediction_kl_nats": 4.729751282754875e-06,
+ "prediction_kl_p99_nats": 3.664258964740034e-05,
+ "maximum_prediction_kl_nats": 0.0006163941351551014,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": 0.0005367274957828738,
+ "maximum_absolute_logit_change": 0.29679757356643677,
+ "maximum_per_target_cross_entropy_increase": 0.03550541400909424
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -3.634410016403937e-05,
+ "mean_prediction_kl_nats": 1.19834438801275e-06,
+ "prediction_kl_p99_nats": 2.5357726543204313e-05,
+ "maximum_prediction_kl_nats": 0.0006348251018272824,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.294968843460083,
+ "maximum_per_target_cross_entropy_increase": 0.036371469497680664
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 843228,
+ "unique_rows": 11024,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 9,
+ "compressed_blocks": 83,
+ "zlib_blocks": 83,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11024
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "5c68f0c4c1115c4ac4f47468c0970427d5816b50837276b478fca3a30081d8e6",
+ "table_sha256": "c4d1471287f3aa2449f1f0868a87bf6930aa29cfea2ff6268b138258600617fe",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08620755490846932,
+ "representatives_sha256": "78b8ed7d64ca2bffb2aefe2ea471e41d72bdf8a8c5a4826ee733dae05607e74c",
+ "structurally_merged_rows": 736,
+ "backend_artifact_bytes": 844471,
+ "backend_with_key_registry_bytes": 1006196,
+ "backend_with_datrie_key_index_bytes": 939751,
+ "full_model_serving_artifact_bytes": 1300882,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9835029294138288,
+ "mean_cosine": 0.9999430210554684,
+ "p05_cosine": 0.9997076885379784,
+ "relative_l2_mean": 0.006896279924639333,
+ "relative_l2_p95": 0.02444953821625191,
+ "relative_l2_max": 0.1820996223771739,
+ "weighted_relative_l2_mean": 0.0058064973361122395,
+ "weighted_relative_frobenius": 0.007663621969295788,
+ "weighted_mean_cosine": 0.9999709715160588,
+ "max_relative_norm_change": 0.010522080816159867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9837323408538031,
+ "mean_cosine": 0.9999545361559533,
+ "p05_cosine": 0.99969453404189,
+ "relative_l2_mean": 0.00222548504458345,
+ "relative_l2_p95": 0.025013083634212202,
+ "relative_l2_max": 0.18096465905004575,
+ "weighted_relative_l2_mean": 0.0009361089003034648,
+ "weighted_relative_frobenius": 0.005558494820779118,
+ "weighted_mean_cosine": 0.9999842571601029,
+ "max_relative_norm_change": 0.009943799209470403
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.847967,
+ "warm_p50_ms": 0.109302,
+ "warm_p95_ms": 0.11183775,
+ "warm_p99_ms": 0.13465091999999998,
+ "samples_ms": [
+ 0.113562,
+ 0.109371,
+ 0.110279,
+ 0.110699,
+ 0.109791,
+ 0.148134,
+ 0.11021,
+ 0.109302,
+ 0.108394,
+ 0.108743,
+ 0.108813,
+ 0.108324,
+ 0.109721,
+ 0.108324,
+ 0.109721,
+ 0.109721,
+ 0.109301,
+ 0.111746,
+ 0.110698,
+ 0.109301,
+ 0.109302,
+ 0.109791,
+ 0.109302,
+ 0.11021,
+ 0.111747,
+ 0.11021,
+ 0.108812,
+ 0.109721,
+ 0.109232,
+ 0.108813,
+ 0.107835,
+ 0.109301,
+ 0.108743,
+ 0.11028,
+ 0.109301,
+ 0.108812,
+ 0.109301,
+ 0.108883,
+ 0.108743,
+ 0.111257
+ ],
+ "warm_rows_per_second": 4684269.27229145
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744884551868943,
+ "character_perplexity": 5.725240468954067,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.598528119194803e-05,
+ "mean_prediction_kl_nats": 5.179810257603944e-06,
+ "prediction_kl_p99_nats": 3.688835137538909e-05,
+ "maximum_prediction_kl_nats": 0.005859044073035739,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": 0.0005367274957828738,
+ "maximum_absolute_logit_change": 0.29679757356643677,
+ "maximum_per_target_cross_entropy_increase": 0.03550541400909424
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.059128329518424e-05,
+ "mean_prediction_kl_nats": 1.6181806715206679e-06,
+ "prediction_kl_p99_nats": 2.5458426726578934e-05,
+ "maximum_prediction_kl_nats": 0.005460578185880949,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.294968843460083,
+ "maximum_per_target_cross_entropy_increase": 0.036371469497680664
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 843074,
+ "unique_rows": 11022,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 12,
+ "compressed_blocks": 80,
+ "zlib_blocks": 80,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11022
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "577aa68acb8cdba2e986daa24ca1390815aa8ab5fe978503dc17cb6c0ded05c9",
+ "table_sha256": "f28ea920fb725358777bd78c0a51d5b70ee961f1451c19a290e7e865ca36c721",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.11374809592962265,
+ "representatives_sha256": "b537133c6bfce4f4984e9e774b5715cf35b391ecde6bb0204224736b9ec2c9f9",
+ "structurally_merged_rows": 738,
+ "backend_artifact_bytes": 844318,
+ "backend_with_key_registry_bytes": 1006043,
+ "backend_with_datrie_key_index_bytes": 939598,
+ "full_model_serving_artifact_bytes": 1300729,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9731854466670903,
+ "mean_cosine": 0.9999362098877692,
+ "p05_cosine": 0.9997062041798425,
+ "relative_l2_mean": 0.006952655238500744,
+ "relative_l2_p95": 0.02454468943200593,
+ "relative_l2_max": 0.2319535707428676,
+ "weighted_relative_l2_mean": 0.0058064973361122395,
+ "weighted_relative_frobenius": 0.007663621969295788,
+ "weighted_mean_cosine": 0.9999709715160588,
+ "max_relative_norm_change": 0.010522080816159867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.972878135926894,
+ "mean_cosine": 0.9999477237904446,
+ "p05_cosine": 0.9996918254970836,
+ "relative_l2_mean": 0.002282572680306364,
+ "relative_l2_p95": 0.025176145704010097,
+ "relative_l2_max": 0.23306861619888342,
+ "weighted_relative_l2_mean": 0.0009361089003034648,
+ "weighted_relative_frobenius": 0.005558494820779118,
+ "weighted_mean_cosine": 0.9999842571601029,
+ "max_relative_norm_change": 0.009943799209470403
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.755777,
+ "warm_p50_ms": 0.1112225,
+ "warm_p95_ms": 0.1141007,
+ "warm_p99_ms": 0.13974522999999997,
+ "samples_ms": [
+ 0.116356,
+ 0.113982,
+ 0.111746,
+ 0.112654,
+ 0.111746,
+ 0.111676,
+ 0.111606,
+ 0.112095,
+ 0.111537,
+ 0.110279,
+ 0.111606,
+ 0.111676,
+ 0.108743,
+ 0.111258,
+ 0.109791,
+ 0.111118,
+ 0.110768,
+ 0.110698,
+ 0.109232,
+ 0.111258,
+ 0.108813,
+ 0.110768,
+ 0.111257,
+ 0.111676,
+ 0.110699,
+ 0.112166,
+ 0.154699,
+ 0.112584,
+ 0.111188,
+ 0.109232,
+ 0.111187,
+ 0.110698,
+ 0.109302,
+ 0.11021,
+ 0.109372,
+ 0.10972,
+ 0.109721,
+ 0.111257,
+ 0.110699,
+ 0.111607
+ ],
+ "warm_rows_per_second": 4603385.1064308025
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.744884551868943,
+ "character_perplexity": 5.725240468954067,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.598528119194803e-05,
+ "mean_prediction_kl_nats": 5.179810257603944e-06,
+ "prediction_kl_p99_nats": 3.688835137538909e-05,
+ "maximum_prediction_kl_nats": 0.005859044073035739,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": 0.0005367274957828738,
+ "maximum_absolute_logit_change": 0.29679757356643677,
+ "maximum_per_target_cross_entropy_increase": 0.03550541400909424
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.059128329518424e-05,
+ "mean_prediction_kl_nats": 1.6181806715206679e-06,
+ "prediction_kl_p99_nats": 2.5458426726578934e-05,
+ "maximum_prediction_kl_nats": 0.005460578185880949,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.294968843460083,
+ "maximum_per_target_cross_entropy_increase": 0.036371469497680664
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842896,
+ "unique_rows": 11019,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 11,
+ "compressed_blocks": 81,
+ "zlib_blocks": 81,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11019
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "95dc83bc2653e9337145412df1324dc2d312e84415fddbf1dff367f5d4e5f095",
+ "table_sha256": "810f7e3c3217dacae23a615b7c5f1be4e0203bb144f3708d6981a8fa6fd29936",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08859019307419658,
+ "representatives_sha256": "f227517c766358e150595bb5e8325380441dcc07039c2bd4f4b6e83899b70c5e",
+ "structurally_merged_rows": 741,
+ "backend_artifact_bytes": 844140,
+ "backend_with_key_registry_bytes": 1005865,
+ "backend_with_datrie_key_index_bytes": 939420,
+ "full_model_serving_artifact_bytes": 1300551,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9598427504538848,
+ "mean_cosine": 0.9999266341928937,
+ "p05_cosine": 0.9997015199376899,
+ "relative_l2_mean": 0.00702112551067649,
+ "relative_l2_p95": 0.0247034349911432,
+ "relative_l2_max": 0.2828580077137856,
+ "weighted_relative_l2_mean": 0.005841040036016813,
+ "weighted_relative_frobenius": 0.008315220355991155,
+ "weighted_mean_cosine": 0.9999660661244957,
+ "max_relative_norm_change": 0.010522080816159867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9600234276278778,
+ "mean_cosine": 0.9999381728045064,
+ "p05_cosine": 0.9996858536608259,
+ "relative_l2_mean": 0.0023522664632986278,
+ "relative_l2_p95": 0.02526999824170825,
+ "relative_l2_max": 0.28214583754347966,
+ "weighted_relative_l2_mean": 0.0009713035943111449,
+ "weighted_relative_frobenius": 0.006424690083692396,
+ "weighted_mean_cosine": 0.999979367336869,
+ "max_relative_norm_change": 0.009943799209470403
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.748723,
+ "warm_p50_ms": 0.1106985,
+ "warm_p95_ms": 0.11589139999999999,
+ "warm_p99_ms": 0.16287847999999994,
+ "samples_ms": [
+ 0.116355,
+ 0.11405,
+ 0.11447,
+ 0.111676,
+ 0.110279,
+ 0.113562,
+ 0.113073,
+ 0.112585,
+ 0.112585,
+ 0.110699,
+ 0.111188,
+ 0.109791,
+ 0.111117,
+ 0.110768,
+ 0.109232,
+ 0.109721,
+ 0.112166,
+ 0.111606,
+ 0.192623,
+ 0.115867,
+ 0.111188,
+ 0.11028,
+ 0.110209,
+ 0.109302,
+ 0.109232,
+ 0.111118,
+ 0.109301,
+ 0.109721,
+ 0.109791,
+ 0.110698,
+ 0.110139,
+ 0.109861,
+ 0.110699,
+ 0.109721,
+ 0.111187,
+ 0.109232,
+ 0.109372,
+ 0.10972,
+ 0.109372,
+ 0.109302
+ ],
+ "warm_rows_per_second": 4625175.5895518
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7450230403743143,
+ "character_perplexity": 5.726033403854421,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00020449251537473282,
+ "mean_prediction_kl_nats": 2.428676828904074e-05,
+ "prediction_kl_p99_nats": 3.8123776147374744e-05,
+ "maximum_prediction_kl_nats": 0.07868376115915458,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": 0.0005367274957828738,
+ "maximum_absolute_logit_change": 0.7481122016906738,
+ "maximum_per_target_cross_entropy_increase": 0.7353435754776001
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 9.790119023641044e-05,
+ "mean_prediction_kl_nats": 2.0551677530587976e-05,
+ "prediction_kl_p99_nats": 2.6455327187997673e-05,
+ "maximum_prediction_kl_nats": 0.07804923014879808,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.7464032173156738,
+ "maximum_per_target_cross_entropy_increase": 0.732738733291626
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842765,
+ "unique_rows": 11017,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 11,
+ "compressed_blocks": 81,
+ "zlib_blocks": 81,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11017
+ },
+ "source_float32_sha256": "74ba1cd907605debe6e87c0b632bf10ec6a40efceb5d71912f3cfe5977051323",
+ "decoded_float32_sha256": "e16cb157f87f733131edb7b59df971e344233e298557ad846424f4f42f9dabd7",
+ "table_sha256": "6fe5e4305ed3902d9e6b90311421e59b5f21efa6168d83d881f362e53725e750",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.18645608704537153,
+ "representatives_sha256": "784ce25c32c23db2424eabaa2626daad205fc934da2e7f23d562b2abf3b6433c",
+ "structurally_merged_rows": 743,
+ "backend_artifact_bytes": 844009,
+ "backend_with_key_registry_bytes": 1005734,
+ "backend_with_datrie_key_index_bytes": 939289,
+ "full_model_serving_artifact_bytes": 1300420,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9178180295155871,
+ "mean_cosine": 0.9999133394146261,
+ "p05_cosine": 0.9997003036972144,
+ "relative_l2_mean": 0.007087242341082681,
+ "relative_l2_p95": 0.024850056293793433,
+ "relative_l2_max": 0.4061799028640071,
+ "weighted_relative_l2_mean": 0.0064790225927562044,
+ "weighted_relative_frobenius": 0.013234738509292052,
+ "weighted_mean_cosine": 0.9998356390480357,
+ "max_relative_norm_change": 0.010522080816159867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9172858244959732,
+ "mean_cosine": 0.9999248077935126,
+ "p05_cosine": 0.9996848710539707,
+ "relative_l2_mean": 0.002419604315695432,
+ "relative_l2_p95": 0.025287800604829137,
+ "relative_l2_max": 0.40764289911274443,
+ "weighted_relative_l2_mean": 0.0016208584560864932,
+ "weighted_relative_frobenius": 0.01216213538365954,
+ "weighted_mean_cosine": 0.9998481026659187,
+ "max_relative_norm_change": 0.009943799209470403
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.763739,
+ "warm_p50_ms": 0.112095,
+ "warm_p95_ms": 0.12583645,
+ "warm_p99_ms": 0.14118436,
+ "samples_ms": [
+ 0.11454,
+ 0.112654,
+ 0.111746,
+ 0.113562,
+ 0.14527,
+ 0.116845,
+ 0.114469,
+ 0.111676,
+ 0.11447,
+ 0.111607,
+ 0.125365,
+ 0.112585,
+ 0.112514,
+ 0.112095,
+ 0.112095,
+ 0.112095,
+ 0.112654,
+ 0.112654,
+ 0.113982,
+ 0.112654,
+ 0.111257,
+ 0.109791,
+ 0.11014,
+ 0.111257,
+ 0.109301,
+ 0.112584,
+ 0.112095,
+ 0.110768,
+ 0.111746,
+ 0.111676,
+ 0.111117,
+ 0.112095,
+ 0.109721,
+ 0.111746,
+ 0.110209,
+ 0.111117,
+ 0.110628,
+ 0.110279,
+ 0.113143,
+ 0.134794
+ ],
+ "warm_rows_per_second": 4567554.306614925
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7453099604119993,
+ "character_perplexity": 5.727676553289417,
+ "next_character_accuracy": 0.5005367274957828,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0004915123999688301,
+ "mean_prediction_kl_nats": 8.339455953100242e-05,
+ "prediction_kl_p99_nats": 0.00028306938107107763,
+ "maximum_prediction_kl_nats": 0.07868376115915458,
+ "top1_changed_fraction": 0.0023002606962122373,
+ "accuracy_change": 0.0006900782088636315,
+ "maximum_absolute_logit_change": 0.7481122016906738,
+ "maximum_per_target_cross_entropy_increase": 0.7353435754776001
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.000384890487255598,
+ "mean_prediction_kl_nats": 7.9830180664046e-05,
+ "prediction_kl_p99_nats": 0.00027854762476799084,
+ "maximum_prediction_kl_nats": 0.07804923014879808,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.7464032173156738,
+ "maximum_per_target_cross_entropy_increase": 0.732738733291626
+ }
+ }
+ ],
+ "baseline_quality": {
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7448185687646842,
+ "character_perplexity": 5.724862712278213,
+ "next_character_accuracy": 0.4998466492869192,
+ "unknown_targets": 0
+ },
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6103526363822893,
+ "character_perplexity": 5.00457571217865,
+ "next_character_accuracy": 0.5303227785018784,
+ "unknown_targets": 0
+ }
+ },
+ "selection": {
+ "fp32": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 2893447,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.610478337983415,
+ "character_perplexity": 5.005204834898665,
+ "next_character_accuracy": 0.5303227785018784,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00012570950190301332,
+ "mean_prediction_kl_nats": 3.522476674748964e-05,
+ "prediction_kl_p99_nats": 3.073411681134878e-05,
+ "maximum_prediction_kl_nats": 0.069073305132374,
+ "top1_changed_fraction": 0.0007666947788085563,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6207461357116699,
+ "maximum_per_target_cross_entropy_increase": 0.810072660446167
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -4.4151146362623543e-05,
+ 0.00031408959841323583
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "fp16": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 1566355,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6104803438537196,
+ "character_perplexity": 5.0052148747004805,
+ "next_character_accuracy": 0.5303227785018784,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00012771562637668983,
+ "mean_prediction_kl_nats": 3.5253858489440136e-05,
+ "prediction_kl_p99_nats": 3.0714668360973565e-05,
+ "maximum_prediction_kl_nats": 0.0691435144903208,
+ "top1_changed_fraction": 0.0007666947788085563,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6211075782775879,
+ "maximum_per_target_cross_entropy_increase": 0.8102507591247559
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -4.18414171096396e-05,
+ 0.00031688913937407293
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "int8": {
+ "feasible": true,
+ "name": "int8-raw",
+ "backend_with_key_registry_bytes": 962252,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6103997491236468,
+ "character_perplexity": 5.004811497013974,
+ "next_character_accuracy": 0.5307061258912826,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 4.711385118005514e-05,
+ "mean_prediction_kl_nats": 3.3553609696799543e-06,
+ "prediction_kl_p99_nats": 2.355900870385665e-05,
+ "maximum_prediction_kl_nats": 7.72629051648873e-05,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": 0.00038334738940426494,
+ "maximum_absolute_logit_change": 0.029384613037109375,
+ "maximum_per_target_cross_entropy_increase": 0.021229267120361328
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 8.000077854203593e-07,
+ 9.973376735015557e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ }
+ },
+ "fixed_quantization_test_controls": {
+ "fp16": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6103547072455107,
+ "character_perplexity": 5.0045860759811625,
+ "next_character_accuracy": 0.5303227785018784,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": 2.0708653656909176e-06,
+ "mean_prediction_kl_nats": 4.295727044007571e-09,
+ "prediction_kl_p99_nats": 2.9273537964819035e-08,
+ "maximum_prediction_kl_nats": 1.6115332053182775e-07,
+ "top1_changed_fraction": 7.666947788085564e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0011279582977294922,
+ "maximum_per_target_cross_entropy_increase": 0.0007696151733398438
+ }
+ },
+ "int8": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6103997491236468,
+ "character_perplexity": 5.004811497013974,
+ "next_character_accuracy": 0.5307061258912826,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": 4.711385118005514e-05,
+ "mean_prediction_kl_nats": 3.3553609696799543e-06,
+ "prediction_kl_p99_nats": 2.355900870385665e-05,
+ "maximum_prediction_kl_nats": 7.72629051648873e-05,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": 0.00038334738940426494,
+ "maximum_absolute_logit_change": 0.029384613037109375,
+ "maximum_per_target_cross_entropy_increase": 0.021229267120361328
+ }
+ }
+ },
+ "checkpoint_sha256": "3576a59df8b4c48fb5cac466445e0472fd53af1450d4c7ffe30d612cf90a6214",
+ "registry_sha256": "7fa792a7a13db17dc47320b6a65018697d338beda3b6328e9775d68716636b54"
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 7,
+ "rows": 11760,
+ "dim": 64,
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "source_float32_bytes": 3010560,
+ "key_registry_bytes": 161725,
+ "unchanged_backbone_artifact_bytes": 294686,
+ "alternative_datrie_key_index_bytes": 95280,
+ "query_rows": 512,
+ "query_ids_sha256": "8ba8e1ab159a822d9159f679166b91955ebc405920cda85ccf24f213c3eff2ad",
+ "frequency_trace_sha256": "73ddaf0153da6d907ac6eef82bf819aaec23a0a5eb289cb0f3f91092c37eb7f0",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 3010560,
+ "unique_rows": 11760,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "table_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.014276340138167143,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3011364,
+ "backend_with_key_registry_bytes": 3173089,
+ "backend_with_datrie_key_index_bytes": 3106644,
+ "full_model_serving_artifact_bytes": 3467775,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.394604,
+ "warm_p50_ms": 0.049727,
+ "warm_p95_ms": 0.05261179999999999,
+ "warm_p99_ms": 0.06417813,
+ "samples_ms": [
+ 0.052102,
+ 0.052521,
+ 0.051124,
+ 0.049238,
+ 0.049238,
+ 0.049657,
+ 0.050146,
+ 0.050146,
+ 0.050216,
+ 0.050146,
+ 0.050146,
+ 0.049657,
+ 0.049727,
+ 0.050146,
+ 0.049657,
+ 0.049727,
+ 0.050146,
+ 0.049727,
+ 0.049657,
+ 0.049657,
+ 0.049727,
+ 0.054337,
+ 0.050146,
+ 0.049657,
+ 0.049727,
+ 0.050146,
+ 0.050146,
+ 0.049727,
+ 0.049657,
+ 0.049657,
+ 0.049168,
+ 0.050216,
+ 0.049658,
+ 0.050147,
+ 0.050146,
+ 0.049727,
+ 0.049727,
+ 0.049727,
+ 0.07047,
+ 0.050146
+ ],
+ "warm_rows_per_second": 10296217.346713053
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7619878874351227,
+ "character_perplexity": 5.824003357774225,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2908719,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 368,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 368,
+ "zlib_blocks": 368,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "table_sha256": "a8a6088cda5b1406c3b425bc807b56a4eb283ab757ee8817c0438802b3677b0e",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2650056080892682,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2909925,
+ "backend_with_key_registry_bytes": 3071650,
+ "backend_with_datrie_key_index_bytes": 3005205,
+ "full_model_serving_artifact_bytes": 3366336,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 12.715526,
+ "warm_p50_ms": 0.058247,
+ "warm_p95_ms": 0.07766719999999999,
+ "warm_p99_ms": 0.08123956,
+ "samples_ms": [
+ 0.077524,
+ 0.081784,
+ 0.077105,
+ 0.077104,
+ 0.073333,
+ 0.060552,
+ 0.058736,
+ 0.058667,
+ 0.058178,
+ 0.058247,
+ 0.058178,
+ 0.058247,
+ 0.056362,
+ 0.058667,
+ 0.058248,
+ 0.05727,
+ 0.05727,
+ 0.0572,
+ 0.0572,
+ 0.057759,
+ 0.05727,
+ 0.0572,
+ 0.056292,
+ 0.056711,
+ 0.058178,
+ 0.058178,
+ 0.058667,
+ 0.058178,
+ 0.058666,
+ 0.058178,
+ 0.058667,
+ 0.056781,
+ 0.057758,
+ 0.056292,
+ 0.07473,
+ 0.080388,
+ 0.071937,
+ 0.066698,
+ 0.063766,
+ 0.058667
+ ],
+ "warm_rows_per_second": 8790152.282520989
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7619878874351227,
+ "character_perplexity": 5.824003357774225,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2908719,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 368,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 368,
+ "zlib_blocks": 368,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "table_sha256": "a8a6088cda5b1406c3b425bc807b56a4eb283ab757ee8817c0438802b3677b0e",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3114413758739829,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2909924,
+ "backend_with_key_registry_bytes": 3071649,
+ "backend_with_datrie_key_index_bytes": 3005204,
+ "full_model_serving_artifact_bytes": 3366335,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.74508,
+ "warm_p50_ms": 0.058667,
+ "warm_p95_ms": 0.062165499999999985,
+ "warm_p99_ms": 0.07782507999999999,
+ "samples_ms": [
+ 0.066279,
+ 0.06153,
+ 0.060063,
+ 0.058667,
+ 0.059505,
+ 0.060622,
+ 0.059086,
+ 0.059644,
+ 0.061949,
+ 0.059156,
+ 0.059156,
+ 0.056712,
+ 0.058178,
+ 0.058177,
+ 0.058667,
+ 0.057689,
+ 0.058667,
+ 0.0572,
+ 0.05727,
+ 0.085207,
+ 0.060972,
+ 0.058667,
+ 0.059993,
+ 0.05727,
+ 0.057269,
+ 0.056781,
+ 0.056781,
+ 0.058667,
+ 0.0572,
+ 0.057759,
+ 0.058667,
+ 0.059156,
+ 0.056292,
+ 0.058667,
+ 0.0572,
+ 0.05727,
+ 0.056362,
+ 0.056851,
+ 0.058667,
+ 0.058178
+ ],
+ "warm_rows_per_second": 8727223.14077761
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7619878874351227,
+ "character_perplexity": 5.824003357774225,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2908719,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 368,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 368,
+ "zlib_blocks": 368,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "table_sha256": "a8a6088cda5b1406c3b425bc807b56a4eb283ab757ee8817c0438802b3677b0e",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2655402428936213,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2909924,
+ "backend_with_key_registry_bytes": 3071649,
+ "backend_with_datrie_key_index_bytes": 3005204,
+ "full_model_serving_artifact_bytes": 3366335,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.742845,
+ "warm_p50_ms": 0.058178,
+ "warm_p95_ms": 0.06009095,
+ "warm_p99_ms": 0.062028049999999994,
+ "samples_ms": [
+ 0.060622,
+ 0.060063,
+ 0.058248,
+ 0.057688,
+ 0.058178,
+ 0.059086,
+ 0.058667,
+ 0.058248,
+ 0.057759,
+ 0.057688,
+ 0.058667,
+ 0.062927,
+ 0.057759,
+ 0.058737,
+ 0.058667,
+ 0.058248,
+ 0.058666,
+ 0.058178,
+ 0.058736,
+ 0.056781,
+ 0.056362,
+ 0.057759,
+ 0.05727,
+ 0.05727,
+ 0.05727,
+ 0.058178,
+ 0.058248,
+ 0.059505,
+ 0.059644,
+ 0.058177,
+ 0.057689,
+ 0.05727,
+ 0.056781,
+ 0.057689,
+ 0.058248,
+ 0.0572,
+ 0.057689,
+ 0.058248,
+ 0.058178,
+ 0.057759
+ ],
+ "warm_rows_per_second": 8800577.537900925
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7619878874351227,
+ "character_perplexity": 5.824003357774225,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2732976,
+ "unique_rows": 11025,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11025
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "6f3cad212c8b92053408e3f1413f511ac4d98f1f102a24c84b584af64aaeb28b",
+ "table_sha256": "3a0405ee037427e6a197d0743c9770fb58b43c2f4e69b0852b3f80a3b08c4fe5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.26928883814252913,
+ "representatives_sha256": "0433240c0af194a070351d5d57755a2eb04ea4365346eabc9150ccbb46b1c357",
+ "structurally_merged_rows": 735,
+ "backend_artifact_bytes": 2734181,
+ "backend_with_key_registry_bytes": 2895906,
+ "backend_with_datrie_key_index_bytes": 2829461,
+ "full_model_serving_artifact_bytes": 3190592,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9932284489604623,
+ "mean_cosine": 0.9999581097104241,
+ "p05_cosine": 0.9996986518093584,
+ "relative_l2_mean": 0.002162167872962852,
+ "relative_l2_p95": 0.0248147106845416,
+ "relative_l2_max": 0.11643283329987608,
+ "weighted_relative_l2_mean": 0.000902934593800698,
+ "weighted_relative_frobenius": 0.005299696904874133,
+ "weighted_mean_cosine": 0.9999857134575801,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9932284489604623,
+ "mean_cosine": 0.9999581097104241,
+ "p05_cosine": 0.9996986518093584,
+ "relative_l2_mean": 0.002162167872962852,
+ "relative_l2_p95": 0.0248147106845416,
+ "relative_l2_max": 0.11643283329987608,
+ "weighted_relative_l2_mean": 0.000902934593800698,
+ "weighted_relative_frobenius": 0.005299696904874133,
+ "weighted_mean_cosine": 0.9999857134575801,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.070133,
+ "warm_p50_ms": 0.057758,
+ "warm_p95_ms": 0.06153005,
+ "warm_p99_ms": 0.06268085,
+ "samples_ms": [
+ 0.06153,
+ 0.063416,
+ 0.060063,
+ 0.061041,
+ 0.059225,
+ 0.059645,
+ 0.059715,
+ 0.058177,
+ 0.058248,
+ 0.058248,
+ 0.05727,
+ 0.058177,
+ 0.057689,
+ 0.057689,
+ 0.059156,
+ 0.057689,
+ 0.057689,
+ 0.061531,
+ 0.058178,
+ 0.057759,
+ 0.058178,
+ 0.057689,
+ 0.05727,
+ 0.057689,
+ 0.05727,
+ 0.057759,
+ 0.05727,
+ 0.057689,
+ 0.057758,
+ 0.057689,
+ 0.057689,
+ 0.05727,
+ 0.057689,
+ 0.057759,
+ 0.05727,
+ 0.0572,
+ 0.057759,
+ 0.057689,
+ 0.057689,
+ 0.057758
+ ],
+ "warm_rows_per_second": 8864572.87302192
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7620156589705396,
+ "character_perplexity": 5.8241651015356695,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.7771921049442663e-05,
+ "mean_prediction_kl_nats": 1.8640273193708572e-06,
+ "prediction_kl_p99_nats": 3.06553759923407e-05,
+ "maximum_prediction_kl_nats": 0.009303638315643778,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.30404186248779297,
+ "maximum_per_target_cross_entropy_increase": 0.04256582260131836
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 2.7771921049442663e-05,
+ "mean_prediction_kl_nats": 1.8640273193708572e-06,
+ "prediction_kl_p99_nats": 3.06553759923407e-05,
+ "maximum_prediction_kl_nats": 0.009303638315643778,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.30404186248779297,
+ "maximum_per_target_cross_entropy_increase": 0.04256582260131836
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2732564,
+ "unique_rows": 11023,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11023
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "dc4a2442d9603c206a29f14840ac47b6e1c4bcca8b25696099674cd5dbd13711",
+ "table_sha256": "066e95a608e2f17caa6ead04461b729bc5e7312d05c1e9f11cac5dae2a6444b7",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.26060099294409156,
+ "representatives_sha256": "b9022b09ef1ceb36d6ae184adf95f607090945b983bc83f60a1538cb45bc4baa",
+ "structurally_merged_rows": 737,
+ "backend_artifact_bytes": 2733769,
+ "backend_with_key_registry_bytes": 2895494,
+ "backend_with_datrie_key_index_bytes": 2829049,
+ "full_model_serving_artifact_bytes": 3190180,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9815472208001749,
+ "mean_cosine": 0.9999555118921756,
+ "p05_cosine": 0.9996977313037,
+ "relative_l2_mean": 0.0021918473872563965,
+ "relative_l2_p95": 0.024869959194056858,
+ "relative_l2_max": 0.1928935956994601,
+ "weighted_relative_l2_mean": 0.0009108762785184122,
+ "weighted_relative_frobenius": 0.005387920989374522,
+ "weighted_mean_cosine": 0.9999850981342973,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9815472208001749,
+ "mean_cosine": 0.9999555118921756,
+ "p05_cosine": 0.9996977313037,
+ "relative_l2_mean": 0.0021918473872563965,
+ "relative_l2_p95": 0.024869959194056858,
+ "relative_l2_max": 0.1928935956994601,
+ "weighted_relative_l2_mean": 0.0009108762785184122,
+ "weighted_relative_frobenius": 0.005387920989374522,
+ "weighted_mean_cosine": 0.9999850981342973,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.792724,
+ "warm_p50_ms": 0.058143,
+ "warm_p95_ms": 0.060133,
+ "warm_p99_ms": 0.06064418,
+ "samples_ms": [
+ 0.060971,
+ 0.060133,
+ 0.059644,
+ 0.059644,
+ 0.057689,
+ 0.059644,
+ 0.059574,
+ 0.059645,
+ 0.059086,
+ 0.059155,
+ 0.058108,
+ 0.05727,
+ 0.057759,
+ 0.060064,
+ 0.060133,
+ 0.057689,
+ 0.057759,
+ 0.057689,
+ 0.057689,
+ 0.058178,
+ 0.057689,
+ 0.058667,
+ 0.05727,
+ 0.058666,
+ 0.057759,
+ 0.05727,
+ 0.059575,
+ 0.057758,
+ 0.058248,
+ 0.057758,
+ 0.060133,
+ 0.057689,
+ 0.057759,
+ 0.058178,
+ 0.05727,
+ 0.058667,
+ 0.057689,
+ 0.057759,
+ 0.059645,
+ 0.057689
+ ],
+ "warm_rows_per_second": 8805875.169839878
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7620191870194526,
+ "character_perplexity": 5.824185649511272,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 3.130007416692424e-05,
+ "mean_prediction_kl_nats": 3.677169019773703e-06,
+ "prediction_kl_p99_nats": 3.134725482268693e-05,
+ "maximum_prediction_kl_nats": 0.01202603974033005,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.30404186248779297,
+ "maximum_per_target_cross_entropy_increase": 0.10781413316726685
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 3.130007416692424e-05,
+ "mean_prediction_kl_nats": 3.677169019773703e-06,
+ "prediction_kl_p99_nats": 3.134725482268693e-05,
+ "maximum_prediction_kl_nats": 0.01202603974033005,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.30404186248779297,
+ "maximum_per_target_cross_entropy_increase": 0.10781413316726685
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2732564,
+ "unique_rows": 11023,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11023
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "dc4a2442d9603c206a29f14840ac47b6e1c4bcca8b25696099674cd5dbd13711",
+ "table_sha256": "066e95a608e2f17caa6ead04461b729bc5e7312d05c1e9f11cac5dae2a6444b7",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2508778888732195,
+ "representatives_sha256": "b9022b09ef1ceb36d6ae184adf95f607090945b983bc83f60a1538cb45bc4baa",
+ "structurally_merged_rows": 737,
+ "backend_artifact_bytes": 2733769,
+ "backend_with_key_registry_bytes": 2895494,
+ "backend_with_datrie_key_index_bytes": 2829049,
+ "full_model_serving_artifact_bytes": 3190180,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9815472208001749,
+ "mean_cosine": 0.9999555118921756,
+ "p05_cosine": 0.9996977313037,
+ "relative_l2_mean": 0.0021918473872563965,
+ "relative_l2_p95": 0.024869959194056858,
+ "relative_l2_max": 0.1928935956994601,
+ "weighted_relative_l2_mean": 0.0009108762785184122,
+ "weighted_relative_frobenius": 0.005387920989374522,
+ "weighted_mean_cosine": 0.9999850981342973,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9815472208001749,
+ "mean_cosine": 0.9999555118921756,
+ "p05_cosine": 0.9996977313037,
+ "relative_l2_mean": 0.0021918473872563965,
+ "relative_l2_p95": 0.024869959194056858,
+ "relative_l2_max": 0.1928935956994601,
+ "weighted_relative_l2_mean": 0.0009108762785184122,
+ "weighted_relative_frobenius": 0.005387920989374522,
+ "weighted_mean_cosine": 0.9999850981342973,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 10.77198,
+ "warm_p50_ms": 0.057759,
+ "warm_p95_ms": 0.06009095,
+ "warm_p99_ms": 0.06113318,
+ "samples_ms": [
+ 0.06146,
+ 0.060063,
+ 0.057759,
+ 0.0572,
+ 0.0572,
+ 0.058178,
+ 0.058667,
+ 0.057689,
+ 0.057689,
+ 0.0572,
+ 0.058667,
+ 0.058247,
+ 0.056292,
+ 0.058667,
+ 0.056781,
+ 0.058178,
+ 0.056781,
+ 0.059086,
+ 0.057689,
+ 0.058667,
+ 0.057759,
+ 0.056781,
+ 0.059225,
+ 0.059155,
+ 0.0572,
+ 0.057758,
+ 0.05727,
+ 0.057689,
+ 0.057689,
+ 0.058178,
+ 0.056781,
+ 0.058178,
+ 0.060622,
+ 0.059645,
+ 0.057759,
+ 0.05727,
+ 0.058667,
+ 0.05734,
+ 0.057759,
+ 0.05727
+ ],
+ "warm_rows_per_second": 8864419.397842761
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7620191870194526,
+ "character_perplexity": 5.824185649511272,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 3.130007416692424e-05,
+ "mean_prediction_kl_nats": 3.677169019773703e-06,
+ "prediction_kl_p99_nats": 3.134725482268693e-05,
+ "maximum_prediction_kl_nats": 0.01202603974033005,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.30404186248779297,
+ "maximum_per_target_cross_entropy_increase": 0.10781413316726685
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 3.130007416692424e-05,
+ "mean_prediction_kl_nats": 3.677169019773703e-06,
+ "prediction_kl_p99_nats": 3.134725482268693e-05,
+ "maximum_prediction_kl_nats": 0.01202603974033005,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.30404186248779297,
+ "maximum_per_target_cross_entropy_increase": 0.10781413316726685
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2732050,
+ "unique_rows": 11021,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11021
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "4928f06da055cb0b4ad7ff511b6eeb90faf48ccdd7ca720e699eca8c64cb0997",
+ "table_sha256": "dfa5967e59333b01d1f76f69eb18ef1459343bbfcf338cbf085910c6baea1239",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25085365399718285,
+ "representatives_sha256": "23ac79b3f25b1f4c3454b3eaf09d347cbd138e5036dab7825f9678041e80a60e",
+ "structurally_merged_rows": 739,
+ "backend_artifact_bytes": 2733255,
+ "backend_with_key_registry_bytes": 2894980,
+ "backend_with_datrie_key_index_bytes": 2828535,
+ "full_model_serving_artifact_bytes": 3189666,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9536948201645477,
+ "mean_cosine": 0.999947709268633,
+ "p05_cosine": 0.9996967567902134,
+ "relative_l2_mean": 0.00224327120065354,
+ "relative_l2_p95": 0.024912690144640124,
+ "relative_l2_max": 0.30424701629479983,
+ "weighted_relative_l2_mean": 0.0011273360959741232,
+ "weighted_relative_frobenius": 0.009895820273657955,
+ "weighted_mean_cosine": 0.999952168096359,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9536948201645477,
+ "mean_cosine": 0.999947709268633,
+ "p05_cosine": 0.9996967567902134,
+ "relative_l2_mean": 0.00224327120065354,
+ "relative_l2_p95": 0.024912690144640124,
+ "relative_l2_max": 0.30424701629479983,
+ "weighted_relative_l2_mean": 0.0011273360959741232,
+ "weighted_relative_frobenius": 0.009895820273657955,
+ "weighted_mean_cosine": 0.999952168096359,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.107848,
+ "warm_p50_ms": 0.058213,
+ "warm_p95_ms": 0.0643064,
+ "warm_p99_ms": 0.07859450999999999,
+ "samples_ms": [
+ 0.063905,
+ 0.060063,
+ 0.058178,
+ 0.058667,
+ 0.058177,
+ 0.064254,
+ 0.065302,
+ 0.060552,
+ 0.060133,
+ 0.060064,
+ 0.059156,
+ 0.058667,
+ 0.058178,
+ 0.058248,
+ 0.056781,
+ 0.057689,
+ 0.057759,
+ 0.087093,
+ 0.059994,
+ 0.059156,
+ 0.059086,
+ 0.059156,
+ 0.059156,
+ 0.058178,
+ 0.058248,
+ 0.057759,
+ 0.058248,
+ 0.057689,
+ 0.057689,
+ 0.058177,
+ 0.0572,
+ 0.057689,
+ 0.055315,
+ 0.05727,
+ 0.058667,
+ 0.057759,
+ 0.057689,
+ 0.058178,
+ 0.057759,
+ 0.0572
+ ],
+ "warm_rows_per_second": 8795286.276261317
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7621428741898844,
+ "character_perplexity": 5.824906071106866,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00015499876582933343,
+ "mean_prediction_kl_nats": 9.021660495853367e-05,
+ "prediction_kl_p99_nats": 4.863982362564446e-05,
+ "maximum_prediction_kl_nats": 0.05025749620497935,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6694626808166504,
+ "maximum_per_target_cross_entropy_increase": 0.3763465881347656
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00015499876582933343,
+ "mean_prediction_kl_nats": 9.021660495853367e-05,
+ "prediction_kl_p99_nats": 4.863982362564446e-05,
+ "maximum_prediction_kl_nats": 0.05025749620497935,
+ "top1_changed_fraction": 0.00015335071308081582,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6694626808166504,
+ "maximum_per_target_cross_entropy_increase": 0.3763465881347656
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2730917,
+ "unique_rows": 11016,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11016
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "22c8fe6614ed4ff860170ad45c38208546df2ad07b0c98445a408a3ef57f2d17",
+ "table_sha256": "4ff1286f6601bdcc973c4fb202d12254a203df7053fd9d7a0fa7ef67bb2518cc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25930648297071457,
+ "representatives_sha256": "6bd47a0656069790de5e168cb0bc8157803c0809d0f911ce4fc16653eca85836",
+ "structurally_merged_rows": 744,
+ "backend_artifact_bytes": 2732122,
+ "backend_with_key_registry_bytes": 2893847,
+ "backend_with_datrie_key_index_bytes": 2827402,
+ "full_model_serving_artifact_bytes": 3188533,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9015189941457709,
+ "mean_cosine": 0.9999128903526319,
+ "p05_cosine": 0.9996942024133612,
+ "relative_l2_mean": 0.0024144631442916398,
+ "relative_l2_p95": 0.024995065662336325,
+ "relative_l2_max": 0.44265425326457186,
+ "weighted_relative_l2_mean": 0.0014611725216560735,
+ "weighted_relative_frobenius": 0.01452628211499881,
+ "weighted_mean_cosine": 0.9998874771283787,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9015189941457709,
+ "mean_cosine": 0.9999128903526319,
+ "p05_cosine": 0.9996942024133612,
+ "relative_l2_mean": 0.0024144631442916398,
+ "relative_l2_p95": 0.024995065662336325,
+ "relative_l2_max": 0.44265425326457186,
+ "weighted_relative_l2_mean": 0.0014611725216560735,
+ "weighted_relative_frobenius": 0.01452628211499881,
+ "weighted_mean_cosine": 0.9998874771283787,
+ "max_relative_norm_change": 0.009859081402261516
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.326451,
+ "warm_p50_ms": 0.0582125,
+ "warm_p95_ms": 0.06102435,
+ "warm_p99_ms": 0.06346775,
+ "samples_ms": [
+ 0.064394,
+ 0.059645,
+ 0.060552,
+ 0.058247,
+ 0.058667,
+ 0.059645,
+ 0.059085,
+ 0.058667,
+ 0.060972,
+ 0.060134,
+ 0.058667,
+ 0.057759,
+ 0.057689,
+ 0.059086,
+ 0.056781,
+ 0.058247,
+ 0.058178,
+ 0.058666,
+ 0.05734,
+ 0.05727,
+ 0.057759,
+ 0.057689,
+ 0.057689,
+ 0.05727,
+ 0.057689,
+ 0.057689,
+ 0.057759,
+ 0.057689,
+ 0.057759,
+ 0.057828,
+ 0.057689,
+ 0.058667,
+ 0.058737,
+ 0.058667,
+ 0.057689,
+ 0.057689,
+ 0.056711,
+ 0.062019,
+ 0.059575,
+ 0.059644
+ ],
+ "warm_rows_per_second": 8795361.820914753
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7622545889350498,
+ "character_perplexity": 5.825556835353516,
+ "next_character_accuracy": 0.5002300260696212,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0002667370679342048,
+ "mean_prediction_kl_nats": 0.00011683976641027171,
+ "prediction_kl_p99_nats": 9.319469902709709e-05,
+ "maximum_prediction_kl_nats": 0.05025749620497935,
+ "top1_changed_fraction": 0.0005367274957828554,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.8649499416351318,
+ "maximum_per_target_cross_entropy_increase": 0.3763465881347656
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0002667370679342048,
+ "mean_prediction_kl_nats": 0.00011683976641027171,
+ "prediction_kl_p99_nats": 9.319469902709709e-05,
+ "maximum_prediction_kl_nats": 0.05025749620497935,
+ "top1_changed_fraction": 0.0005367274957828554,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.8649499416351318,
+ "maximum_per_target_cross_entropy_increase": 0.3763465881347656
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 128,
+ "storage": {
+ "artifact_bytes": 1505280,
+ "unique_rows": 11760,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "811798cfb33aa60ba8fe68fa7bdc76afdb78f8ca17a89d0774d37291da0f6e18",
+ "table_sha256": "be3e6cb3faabcf8e7951e8696b5142f010d1678db8d4b7364a8077e60007e139",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.018310656072571874,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1506085,
+ "backend_with_key_registry_bytes": 1667810,
+ "backend_with_datrie_key_index_bytes": 1601365,
+ "full_model_serving_artifact_bytes": 1962496,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.999999955330057,
+ "mean_cosine": 0.9999999792978743,
+ "p05_cosine": 0.9999999716056523,
+ "relative_l2_mean": 0.00020619931955426777,
+ "relative_l2_p95": 0.00024277821700190904,
+ "relative_l2_max": 0.00030034986302671096,
+ "weighted_relative_l2_mean": 0.00018740057164396058,
+ "weighted_relative_frobenius": 0.00020735037194932163,
+ "weighted_mean_cosine": 0.9999999813082834,
+ "max_relative_norm_change": 0.00016211508208365433
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.590508,
+ "warm_p50_ms": 0.13646999999999998,
+ "warm_p95_ms": 0.20729335,
+ "warm_p99_ms": 0.22529658,
+ "samples_ms": [
+ 0.202889,
+ 0.216648,
+ 0.206801,
+ 0.198699,
+ 0.199118,
+ 0.193949,
+ 0.198279,
+ 0.194927,
+ 0.199607,
+ 0.195835,
+ 0.192134,
+ 0.158959,
+ 0.136261,
+ 0.13242,
+ 0.135772,
+ 0.134794,
+ 0.132,
+ 0.134374,
+ 0.133816,
+ 0.133397,
+ 0.132908,
+ 0.14527,
+ 0.191644,
+ 0.195905,
+ 0.230826,
+ 0.200654,
+ 0.194928,
+ 0.157562,
+ 0.134863,
+ 0.133956,
+ 0.133467,
+ 0.136679,
+ 0.132978,
+ 0.133816,
+ 0.132978,
+ 0.131512,
+ 0.132489,
+ 0.131511,
+ 0.131511,
+ 0.131022
+ ],
+ "warm_rows_per_second": 3751740.3092254717
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761986084105392,
+ "character_perplexity": 5.823992855185288,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.8033281047462318e-06,
+ "mean_prediction_kl_nats": 4.335405924479014e-09,
+ "prediction_kl_p99_nats": 3.1074343989937316e-08,
+ "maximum_prediction_kl_nats": 5.961903623836042e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010232925415039062,
+ "maximum_per_target_cross_entropy_increase": 0.000675201416015625
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1491617,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 184,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 184,
+ "zlib_blocks": 184,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "811798cfb33aa60ba8fe68fa7bdc76afdb78f8ca17a89d0774d37291da0f6e18",
+ "table_sha256": "eb91629c3bc916b563bf24f086a779b9d4628c56b45efffbc009d24980259264",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14671796304173768,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1492824,
+ "backend_with_key_registry_bytes": 1654549,
+ "backend_with_datrie_key_index_bytes": 1588104,
+ "full_model_serving_artifact_bytes": 1949235,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.999999955330057,
+ "mean_cosine": 0.9999999792978743,
+ "p05_cosine": 0.9999999716056523,
+ "relative_l2_mean": 0.00020619931955426777,
+ "relative_l2_p95": 0.00024277821700190904,
+ "relative_l2_max": 0.00030034986302671096,
+ "weighted_relative_l2_mean": 0.00018740057164396058,
+ "weighted_relative_frobenius": 0.00020735037194932163,
+ "weighted_mean_cosine": 0.9999999813082834,
+ "max_relative_norm_change": 0.00016211508208365433
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.94109,
+ "warm_p50_ms": 0.14191749999999997,
+ "warm_p95_ms": 0.1476904,
+ "warm_p99_ms": 0.15549723999999998,
+ "samples_ms": [
+ 0.148553,
+ 0.143454,
+ 0.146247,
+ 0.144362,
+ 0.144711,
+ 0.146597,
+ 0.143733,
+ 0.142895,
+ 0.141429,
+ 0.14108,
+ 0.141847,
+ 0.141987,
+ 0.141918,
+ 0.141918,
+ 0.141847,
+ 0.146667,
+ 0.140032,
+ 0.141428,
+ 0.142407,
+ 0.140032,
+ 0.142336,
+ 0.159937,
+ 0.147645,
+ 0.142406,
+ 0.141917,
+ 0.141987,
+ 0.141848,
+ 0.141918,
+ 0.140031,
+ 0.141429,
+ 0.140032,
+ 0.14094,
+ 0.14101,
+ 0.141009,
+ 0.141429,
+ 0.140032,
+ 0.14101,
+ 0.14094,
+ 0.142407,
+ 0.141428
+ ],
+ "warm_rows_per_second": 3607729.843042613
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761986084105392,
+ "character_perplexity": 5.823992855185288,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.8033281047462318e-06,
+ "mean_prediction_kl_nats": 4.335405924479014e-09,
+ "prediction_kl_p99_nats": 3.1074343989937316e-08,
+ "maximum_prediction_kl_nats": 5.961903623836042e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010232925415039062,
+ "maximum_per_target_cross_entropy_increase": 0.000675201416015625
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1491617,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 184,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 184,
+ "zlib_blocks": 184,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "811798cfb33aa60ba8fe68fa7bdc76afdb78f8ca17a89d0774d37291da0f6e18",
+ "table_sha256": "eb91629c3bc916b563bf24f086a779b9d4628c56b45efffbc009d24980259264",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.18984828097745776,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1492823,
+ "backend_with_key_registry_bytes": 1654548,
+ "backend_with_datrie_key_index_bytes": 1588103,
+ "full_model_serving_artifact_bytes": 1949234,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.999999955330057,
+ "mean_cosine": 0.9999999792978743,
+ "p05_cosine": 0.9999999716056523,
+ "relative_l2_mean": 0.00020619931955426777,
+ "relative_l2_p95": 0.00024277821700190904,
+ "relative_l2_max": 0.00030034986302671096,
+ "weighted_relative_l2_mean": 0.00018740057164396058,
+ "weighted_relative_frobenius": 0.00020735037194932163,
+ "weighted_mean_cosine": 0.9999999813082834,
+ "max_relative_norm_change": 0.00016211508208365433
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.877116,
+ "warm_p50_ms": 0.14149850000000003,
+ "warm_p95_ms": 0.14473545000000002,
+ "warm_p99_ms": 0.16253985999999998,
+ "samples_ms": [
+ 0.1452,
+ 0.143733,
+ 0.142895,
+ 0.142476,
+ 0.142337,
+ 0.141499,
+ 0.141498,
+ 0.141499,
+ 0.141499,
+ 0.141429,
+ 0.14101,
+ 0.141359,
+ 0.141499,
+ 0.141009,
+ 0.14101,
+ 0.141429,
+ 0.142336,
+ 0.141498,
+ 0.143315,
+ 0.144711,
+ 0.144362,
+ 0.142406,
+ 0.142336,
+ 0.14101,
+ 0.142407,
+ 0.173626,
+ 0.142826,
+ 0.141498,
+ 0.140591,
+ 0.140032,
+ 0.140032,
+ 0.141988,
+ 0.141429,
+ 0.141009,
+ 0.141429,
+ 0.141917,
+ 0.140101,
+ 0.140521,
+ 0.140032,
+ 0.139543
+ ],
+ "warm_rows_per_second": 3618412.9160379786
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761986084105392,
+ "character_perplexity": 5.823992855185288,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.8033281047462318e-06,
+ "mean_prediction_kl_nats": 4.335405924479014e-09,
+ "prediction_kl_p99_nats": 3.1074343989937316e-08,
+ "maximum_prediction_kl_nats": 5.961903623836042e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010232925415039062,
+ "maximum_per_target_cross_entropy_increase": 0.000675201416015625
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1491617,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 184,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 184,
+ "zlib_blocks": 184,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "811798cfb33aa60ba8fe68fa7bdc76afdb78f8ca17a89d0774d37291da0f6e18",
+ "table_sha256": "eb91629c3bc916b563bf24f086a779b9d4628c56b45efffbc009d24980259264",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.15167432511225343,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1492823,
+ "backend_with_key_registry_bytes": 1654548,
+ "backend_with_datrie_key_index_bytes": 1588103,
+ "full_model_serving_artifact_bytes": 1949234,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.999999955330057,
+ "mean_cosine": 0.9999999792978743,
+ "p05_cosine": 0.9999999716056523,
+ "relative_l2_mean": 0.00020619931955426777,
+ "relative_l2_p95": 0.00024277821700190904,
+ "relative_l2_max": 0.00030034986302671096,
+ "weighted_relative_l2_mean": 0.00018740057164396058,
+ "weighted_relative_frobenius": 0.00020735037194932163,
+ "weighted_mean_cosine": 0.9999999813082834,
+ "max_relative_norm_change": 0.00016211508208365433
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.848271,
+ "warm_p50_ms": 0.14237149999999998,
+ "warm_p95_ms": 0.169924,
+ "warm_p99_ms": 0.17363035999999998,
+ "samples_ms": [
+ 0.144362,
+ 0.176,
+ 0.155606,
+ 0.147155,
+ 0.144781,
+ 0.143803,
+ 0.143803,
+ 0.144292,
+ 0.142337,
+ 0.143245,
+ 0.145759,
+ 0.144781,
+ 0.169924,
+ 0.141498,
+ 0.140521,
+ 0.14094,
+ 0.140939,
+ 0.141498,
+ 0.140032,
+ 0.143873,
+ 0.143803,
+ 0.144222,
+ 0.141987,
+ 0.141917,
+ 0.14094,
+ 0.141918,
+ 0.169924,
+ 0.14101,
+ 0.16315,
+ 0.144223,
+ 0.142896,
+ 0.14101,
+ 0.141498,
+ 0.141498,
+ 0.140032,
+ 0.141848,
+ 0.142406,
+ 0.141917,
+ 0.14094,
+ 0.140032
+ ],
+ "warm_rows_per_second": 3596225.368138989
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761986084105392,
+ "character_perplexity": 5.823992855185288,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.8033281047462318e-06,
+ "mean_prediction_kl_nats": 4.335405924479014e-09,
+ "prediction_kl_p99_nats": 3.1074343989937316e-08,
+ "maximum_prediction_kl_nats": 5.961903623836042e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010232925415039062,
+ "maximum_per_target_cross_entropy_increase": 0.000675201416015625
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404829,
+ "unique_rows": 11025,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11025
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "aae891e910365c9838a8e14ff62bc518e1387413e3136272c19c395acdcd1d62",
+ "table_sha256": "f10854c90a71b5b38d2dc288a44794aff49beb0abcf2c13e8058bcad7e440637",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14365919097326696,
+ "representatives_sha256": "0433240c0af194a070351d5d57755a2eb04ea4365346eabc9150ccbb46b1c357",
+ "structurally_merged_rows": 735,
+ "backend_artifact_bytes": 1406034,
+ "backend_with_key_registry_bytes": 1567759,
+ "backend_with_datrie_key_index_bytes": 1501314,
+ "full_model_serving_artifact_bytes": 1862445,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.993229021916863,
+ "mean_cosine": 0.9999580896496946,
+ "p05_cosine": 0.9996981100397673,
+ "relative_l2_mean": 0.0023555824252269805,
+ "relative_l2_p95": 0.02481648552558157,
+ "relative_l2_max": 0.1164297624166434,
+ "weighted_relative_l2_mean": 0.0010839693066042207,
+ "weighted_relative_frobenius": 0.0053040528702744776,
+ "weighted_mean_cosine": 0.9999856948084294,
+ "max_relative_norm_change": 0.009869613651210883
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9932305961542881,
+ "mean_cosine": 0.9999581091146609,
+ "p05_cosine": 0.9996977228528464,
+ "relative_l2_mean": 0.0021622629544317713,
+ "relative_l2_p95": 0.024824343791554238,
+ "relative_l2_max": 0.11641462754844928,
+ "weighted_relative_l2_mean": 0.0009031974001573273,
+ "weighted_relative_frobenius": 0.005301355563665608,
+ "weighted_mean_cosine": 0.9999857061304772,
+ "max_relative_norm_change": 0.009947623318429478
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.592791,
+ "warm_p50_ms": 0.1414285,
+ "warm_p95_ms": 0.14527,
+ "warm_p99_ms": 0.15131998,
+ "samples_ms": [
+ 0.14527,
+ 0.143314,
+ 0.141499,
+ 0.141499,
+ 0.14101,
+ 0.14527,
+ 0.142337,
+ 0.142407,
+ 0.141918,
+ 0.142476,
+ 0.141917,
+ 0.140032,
+ 0.141429,
+ 0.141428,
+ 0.139543,
+ 0.141917,
+ 0.140032,
+ 0.141848,
+ 0.14052,
+ 0.14101,
+ 0.141429,
+ 0.144292,
+ 0.143315,
+ 0.140521,
+ 0.141429,
+ 0.155188,
+ 0.142896,
+ 0.14101,
+ 0.141009,
+ 0.140521,
+ 0.140521,
+ 0.14101,
+ 0.139124,
+ 0.141429,
+ 0.140521,
+ 0.140451,
+ 0.140102,
+ 0.140031,
+ 0.140032,
+ 0.141428
+ ],
+ "warm_rows_per_second": 3620203.848587802
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7620133024290583,
+ "character_perplexity": 5.824151376665186,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.541531689919542e-05,
+ "mean_prediction_kl_nats": 1.8676947177272281e-06,
+ "prediction_kl_p99_nats": 3.056987213103323e-05,
+ "maximum_prediction_kl_nats": 0.009301476677892186,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.3039989471435547,
+ "maximum_per_target_cross_entropy_increase": 0.04271221160888672
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 2.7218694088177672e-05,
+ "mean_prediction_kl_nats": 1.8641467086508255e-06,
+ "prediction_kl_p99_nats": 3.073360547172181e-05,
+ "maximum_prediction_kl_nats": 0.00930052162812792,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.3040299415588379,
+ "maximum_per_target_cross_entropy_increase": 0.04257345199584961
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404621,
+ "unique_rows": 11023,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11023
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "90e0b3db1e82134bb99d2f655fea869dbdbd1b964085839484dc2fe5e4c8873c",
+ "table_sha256": "d5fb9ea6ff92fffe220e9a68f328f35920e64190e49e455cbb51888e90481c3f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.18668558495119214,
+ "representatives_sha256": "b9022b09ef1ceb36d6ae184adf95f607090945b983bc83f60a1538cb45bc4baa",
+ "structurally_merged_rows": 737,
+ "backend_artifact_bytes": 1405826,
+ "backend_with_key_registry_bytes": 1567551,
+ "backend_with_datrie_key_index_bytes": 1501106,
+ "full_model_serving_artifact_bytes": 1862237,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9815473821674361,
+ "mean_cosine": 0.9999554916802988,
+ "p05_cosine": 0.9996976348833175,
+ "relative_l2_mean": 0.002385236134563276,
+ "relative_l2_p95": 0.024851234547092083,
+ "relative_l2_max": 0.1928889038834104,
+ "weighted_relative_l2_mean": 0.0010919041739036977,
+ "weighted_relative_frobenius": 0.005392229889626232,
+ "weighted_mean_cosine": 0.999985079385915,
+ "max_relative_norm_change": 0.009869613651210883
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9815473620654889,
+ "mean_cosine": 0.9999555112196998,
+ "p05_cosine": 0.9996973251586957,
+ "relative_l2_mean": 0.002191942915397932,
+ "relative_l2_p95": 0.024867639489037825,
+ "relative_l2_max": 0.1928818136964326,
+ "weighted_relative_l2_mean": 0.0009111399513267083,
+ "weighted_relative_frobenius": 0.005389569694692704,
+ "weighted_mean_cosine": 0.9999850907541233,
+ "max_relative_norm_change": 0.009947623318429478
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.582804,
+ "warm_p50_ms": 0.1409745,
+ "warm_p95_ms": 0.14529795,
+ "warm_p99_ms": 0.15183567,
+ "samples_ms": [
+ 0.145201,
+ 0.142896,
+ 0.142826,
+ 0.142477,
+ 0.141009,
+ 0.141498,
+ 0.141918,
+ 0.140521,
+ 0.14052,
+ 0.145829,
+ 0.14094,
+ 0.140521,
+ 0.141009,
+ 0.14094,
+ 0.14101,
+ 0.139963,
+ 0.14527,
+ 0.14094,
+ 0.139543,
+ 0.140032,
+ 0.140032,
+ 0.155676,
+ 0.140451,
+ 0.140591,
+ 0.140521,
+ 0.14101,
+ 0.140032,
+ 0.140102,
+ 0.141498,
+ 0.14101,
+ 0.14101,
+ 0.138635,
+ 0.14101,
+ 0.140451,
+ 0.14101,
+ 0.140031,
+ 0.142825,
+ 0.142895,
+ 0.140102,
+ 0.14052
+ ],
+ "warm_rows_per_second": 3631862.499955666
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7620168032906287,
+ "character_perplexity": 5.824171766248611,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.8916273573369703e-05,
+ "mean_prediction_kl_nats": 3.681654640875653e-06,
+ "prediction_kl_p99_nats": 3.130291623387823e-05,
+ "maximum_prediction_kl_nats": 0.012031577285817373,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.3039989471435547,
+ "maximum_per_target_cross_entropy_increase": 0.10785055160522461
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 3.071965707573691e-05,
+ "mean_prediction_kl_nats": 3.6776569666511883e-06,
+ "prediction_kl_p99_nats": 3.151438021132739e-05,
+ "maximum_prediction_kl_nats": 0.012028617322816076,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.3040299415588379,
+ "maximum_per_target_cross_entropy_increase": 0.10784733295440674
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404621,
+ "unique_rows": 11023,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11023
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "90e0b3db1e82134bb99d2f655fea869dbdbd1b964085839484dc2fe5e4c8873c",
+ "table_sha256": "d5fb9ea6ff92fffe220e9a68f328f35920e64190e49e455cbb51888e90481c3f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1453714920207858,
+ "representatives_sha256": "b9022b09ef1ceb36d6ae184adf95f607090945b983bc83f60a1538cb45bc4baa",
+ "structurally_merged_rows": 737,
+ "backend_artifact_bytes": 1405826,
+ "backend_with_key_registry_bytes": 1567551,
+ "backend_with_datrie_key_index_bytes": 1501106,
+ "full_model_serving_artifact_bytes": 1862237,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9815473821674361,
+ "mean_cosine": 0.9999554916802988,
+ "p05_cosine": 0.9996976348833175,
+ "relative_l2_mean": 0.002385236134563276,
+ "relative_l2_p95": 0.024851234547092083,
+ "relative_l2_max": 0.1928889038834104,
+ "weighted_relative_l2_mean": 0.0010919041739036977,
+ "weighted_relative_frobenius": 0.005392229889626232,
+ "weighted_mean_cosine": 0.999985079385915,
+ "max_relative_norm_change": 0.009869613651210883
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9815473620654889,
+ "mean_cosine": 0.9999555112196998,
+ "p05_cosine": 0.9996973251586957,
+ "relative_l2_mean": 0.002191942915397932,
+ "relative_l2_p95": 0.024867639489037825,
+ "relative_l2_max": 0.1928818136964326,
+ "weighted_relative_l2_mean": 0.0009111399513267083,
+ "weighted_relative_frobenius": 0.005389569694692704,
+ "weighted_mean_cosine": 0.9999850907541233,
+ "max_relative_norm_change": 0.009947623318429478
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.555426,
+ "warm_p50_ms": 0.1419875,
+ "warm_p95_ms": 0.17034305,
+ "warm_p99_ms": 0.17204773,
+ "samples_ms": [
+ 0.150438,
+ 0.173137,
+ 0.144781,
+ 0.143734,
+ 0.142407,
+ 0.142895,
+ 0.142895,
+ 0.14101,
+ 0.142337,
+ 0.141499,
+ 0.141918,
+ 0.14052,
+ 0.140521,
+ 0.141429,
+ 0.141987,
+ 0.141917,
+ 0.147575,
+ 0.142407,
+ 0.170344,
+ 0.169854,
+ 0.170343,
+ 0.159936,
+ 0.141988,
+ 0.141429,
+ 0.141987,
+ 0.142406,
+ 0.141499,
+ 0.14101,
+ 0.144851,
+ 0.142966,
+ 0.142896,
+ 0.142896,
+ 0.141499,
+ 0.140031,
+ 0.140451,
+ 0.141009,
+ 0.140032,
+ 0.14101,
+ 0.140521,
+ 0.141499
+ ],
+ "warm_rows_per_second": 3605951.2281010654
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7620168032906287,
+ "character_perplexity": 5.824171766248611,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.8916273573369703e-05,
+ "mean_prediction_kl_nats": 3.681654640875653e-06,
+ "prediction_kl_p99_nats": 3.130291623387823e-05,
+ "maximum_prediction_kl_nats": 0.012031577285817373,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.3039989471435547,
+ "maximum_per_target_cross_entropy_increase": 0.10785055160522461
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 3.071965707573691e-05,
+ "mean_prediction_kl_nats": 3.6776569666511883e-06,
+ "prediction_kl_p99_nats": 3.151438021132739e-05,
+ "maximum_prediction_kl_nats": 0.012028617322816076,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.3040299415588379,
+ "maximum_per_target_cross_entropy_increase": 0.10784733295440674
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404406,
+ "unique_rows": 11021,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11021
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "00951198c4cc8a6dfcefc09b2f4af63b6e6751a341269d9d242ee5e5cd8eeb0f",
+ "table_sha256": "b72e110bfb4b26d659114a5ded8020f9e70d729b2a8caa0897e3daa7c77a34bc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14290357707068324,
+ "representatives_sha256": "23ac79b3f25b1f4c3454b3eaf09d347cbd138e5036dab7825f9678041e80a60e",
+ "structurally_merged_rows": 739,
+ "backend_artifact_bytes": 1405611,
+ "backend_with_key_registry_bytes": 1567336,
+ "backend_with_datrie_key_index_bytes": 1500891,
+ "full_model_serving_artifact_bytes": 1862022,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9536879264202704,
+ "mean_cosine": 0.999947688741713,
+ "p05_cosine": 0.9996970572637678,
+ "relative_l2_mean": 0.0024366258444878178,
+ "relative_l2_p95": 0.024877595078460064,
+ "relative_l2_max": 0.30427105702833357,
+ "weighted_relative_l2_mean": 0.0013082334405640028,
+ "weighted_relative_frobenius": 0.009898667711519395,
+ "weighted_mean_cosine": 0.9999521449646063,
+ "max_relative_norm_change": 0.009869613651210883
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9536887235445123,
+ "mean_cosine": 0.9999477084445584,
+ "p05_cosine": 0.9996969521525462,
+ "relative_l2_mean": 0.002243366937749199,
+ "relative_l2_p95": 0.024871331284442407,
+ "relative_l2_max": 0.30426727296020073,
+ "weighted_relative_l2_mean": 0.001127612257645129,
+ "weighted_relative_frobenius": 0.009897187065153733,
+ "weighted_mean_cosine": 0.9999521569043672,
+ "max_relative_norm_change": 0.009947623318429478
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.442353,
+ "warm_p50_ms": 0.14149850000000003,
+ "warm_p95_ms": 0.1699727,
+ "warm_p99_ms": 0.17423063,
+ "samples_ms": [
+ 0.144711,
+ 0.142896,
+ 0.142896,
+ 0.14101,
+ 0.140451,
+ 0.14101,
+ 0.141499,
+ 0.141429,
+ 0.157003,
+ 0.141917,
+ 0.14101,
+ 0.141918,
+ 0.141987,
+ 0.140521,
+ 0.169854,
+ 0.169295,
+ 0.175511,
+ 0.168946,
+ 0.141429,
+ 0.140032,
+ 0.141917,
+ 0.141498,
+ 0.140032,
+ 0.145689,
+ 0.140032,
+ 0.14108,
+ 0.140939,
+ 0.140521,
+ 0.142476,
+ 0.140521,
+ 0.141499,
+ 0.14094,
+ 0.140521,
+ 0.140032,
+ 0.141429,
+ 0.172228,
+ 0.145689,
+ 0.141987,
+ 0.141918,
+ 0.14101
+ ],
+ "warm_rows_per_second": 3618412.9160379786
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7621404760933819,
+ "character_perplexity": 5.824892102436738,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00015260030050063522,
+ "mean_prediction_kl_nats": 9.020268807501378e-05,
+ "prediction_kl_p99_nats": 4.803773480834617e-05,
+ "maximum_prediction_kl_nats": 0.05024508373924575,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6693630218505859,
+ "maximum_per_target_cross_entropy_increase": 0.37635135650634766
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00015440390704628653,
+ "mean_prediction_kl_nats": 9.021075152072548e-05,
+ "prediction_kl_p99_nats": 4.868768343371962e-05,
+ "maximum_prediction_kl_nats": 0.050253411308513185,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6694545745849609,
+ "maximum_per_target_cross_entropy_increase": 0.376340389251709
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1403813,
+ "unique_rows": 11016,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11016
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "1ffc72ca16ebb2263679347db4ab599758452a04990d3088f40760ef5fec0ac2",
+ "table_sha256": "7ebf6dbde9b37f79bc2f68357cd39a312fd52e6308a9e135fced9ee15413d0a3",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1452479420695454,
+ "representatives_sha256": "6bd47a0656069790de5e168cb0bc8157803c0809d0f911ce4fc16653eca85836",
+ "structurally_merged_rows": 744,
+ "backend_artifact_bytes": 1405018,
+ "backend_with_key_registry_bytes": 1566743,
+ "backend_with_datrie_key_index_bytes": 1500298,
+ "full_model_serving_artifact_bytes": 1861429,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9015172971537725,
+ "mean_cosine": 0.9999128697481101,
+ "p05_cosine": 0.999693868495509,
+ "relative_l2_mean": 0.0026077337396000037,
+ "relative_l2_p95": 0.025014643768802226,
+ "relative_l2_max": 0.44264593861029344,
+ "weighted_relative_l2_mean": 0.0016419059260299174,
+ "weighted_relative_frobenius": 0.014528266044215006,
+ "weighted_mean_cosine": 0.9998874538114033,
+ "max_relative_norm_change": 0.009869613651210883
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9014989543453286,
+ "mean_cosine": 0.9999128887472181,
+ "p05_cosine": 0.9996941556050969,
+ "relative_l2_mean": 0.002414558381037749,
+ "relative_l2_p95": 0.02500559539624438,
+ "relative_l2_max": 0.44268190571327065,
+ "weighted_relative_l2_mean": 0.0014614366736284347,
+ "weighted_relative_frobenius": 0.014527090261736098,
+ "weighted_mean_cosine": 0.9998874688052568,
+ "max_relative_norm_change": 0.009947623318429478
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.497178,
+ "warm_p50_ms": 0.1424065,
+ "warm_p95_ms": 0.14750865,
+ "warm_p99_ms": 0.16694762999999999,
+ "samples_ms": [
+ 0.144711,
+ 0.144712,
+ 0.142477,
+ 0.142895,
+ 0.142407,
+ 0.143384,
+ 0.154209,
+ 0.144222,
+ 0.142406,
+ 0.142476,
+ 0.141848,
+ 0.142407,
+ 0.142826,
+ 0.142476,
+ 0.141428,
+ 0.141429,
+ 0.147156,
+ 0.145201,
+ 0.142407,
+ 0.142966,
+ 0.141429,
+ 0.14527,
+ 0.144362,
+ 0.143804,
+ 0.142337,
+ 0.141499,
+ 0.141009,
+ 0.14101,
+ 0.14101,
+ 0.141917,
+ 0.141987,
+ 0.141988,
+ 0.140102,
+ 0.141499,
+ 0.175092,
+ 0.141917,
+ 0.141918,
+ 0.14101,
+ 0.141987,
+ 0.142336
+ ],
+ "warm_rows_per_second": 3595341.5047768187
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7622522200388164,
+ "character_perplexity": 5.825543035230217,
+ "next_character_accuracy": 0.5002300260696212,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00026436754263474687,
+ "mean_prediction_kl_nats": 0.00011682734223863031,
+ "prediction_kl_p99_nats": 9.370032000681665e-05,
+ "maximum_prediction_kl_nats": 0.05024508373924575,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.8651859760284424,
+ "maximum_per_target_cross_entropy_increase": 0.37635135650634766
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0002661713507337706,
+ "mean_prediction_kl_nats": 0.00011683772936588893,
+ "prediction_kl_p99_nats": 9.312245935059851e-05,
+ "maximum_prediction_kl_nats": 0.050253411308513185,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.8652682304382324,
+ "maximum_per_target_cross_entropy_increase": 0.376340389251709
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 68,
+ "storage": {
+ "artifact_bytes": 799680,
+ "unique_rows": 11760,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "f6ffab41a644dd67e026935d7b50595b09e149dee71f6d058748771dac0f90df",
+ "table_sha256": "286e59cb1634e5510ae31256135a585ad96dc5281b739140ec86c3f850c402a8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.015558068873360753,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 800527,
+ "backend_with_key_registry_bytes": 962252,
+ "backend_with_datrie_key_index_bytes": 895807,
+ "full_model_serving_artifact_bytes": 1256938,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999355787040166,
+ "mean_cosine": 0.9999869086187041,
+ "p05_cosine": 0.9999750882928043,
+ "relative_l2_mean": 0.00500744562114531,
+ "relative_l2_p95": 0.00710717717765626,
+ "relative_l2_max": 0.011400289643805562,
+ "weighted_relative_l2_mean": 0.005094142639764323,
+ "weighted_relative_frobenius": 0.005507144397623611,
+ "weighted_mean_cosine": 0.999985464118255,
+ "max_relative_norm_change": 0.0030795821881461597
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.335029,
+ "warm_p50_ms": 0.0974985,
+ "warm_p95_ms": 0.10225135,
+ "warm_p99_ms": 0.13308359999999997,
+ "samples_ms": [
+ 0.10127,
+ 0.098895,
+ 0.09701,
+ 0.096521,
+ 0.097428,
+ 0.097918,
+ 0.097848,
+ 0.097499,
+ 0.151905,
+ 0.102178,
+ 0.097918,
+ 0.09701,
+ 0.09694,
+ 0.097498,
+ 0.098825,
+ 0.097428,
+ 0.099384,
+ 0.097987,
+ 0.103645,
+ 0.10127,
+ 0.09694,
+ 0.100712,
+ 0.09701,
+ 0.09659,
+ 0.097498,
+ 0.098406,
+ 0.096451,
+ 0.097918,
+ 0.096521,
+ 0.096451,
+ 0.095543,
+ 0.09652,
+ 0.096521,
+ 0.096102,
+ 0.096521,
+ 0.097498,
+ 0.098825,
+ 0.098895,
+ 0.097917,
+ 0.099314
+ ],
+ "warm_rows_per_second": 5251362.8414796125
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761994707803187,
+ "character_perplexity": 5.824043079756192,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.82039132299268e-06,
+ "mean_prediction_kl_nats": 3.6558254564872534e-06,
+ "prediction_kl_p99_nats": 2.1573732474216604e-05,
+ "maximum_prediction_kl_nats": 9.125894225908877e-05,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.029454708099365234,
+ "maximum_per_target_cross_entropy_increase": 0.02509140968322754
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 893386,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 98,
+ "max_block_bytes": 8192,
+ "raw_blocks": 12,
+ "compressed_blocks": 86,
+ "zlib_blocks": 86,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "f6ffab41a644dd67e026935d7b50595b09e149dee71f6d058748771dac0f90df",
+ "table_sha256": "7abea1c680e92a405165f5a1849259c285226ac51a1c93562204d21ecce8820c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08383755898103118,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 894632,
+ "backend_with_key_registry_bytes": 1056357,
+ "backend_with_datrie_key_index_bytes": 989912,
+ "full_model_serving_artifact_bytes": 1351043,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999355787040166,
+ "mean_cosine": 0.9999869086187041,
+ "p05_cosine": 0.9999750882928043,
+ "relative_l2_mean": 0.00500744562114531,
+ "relative_l2_p95": 0.00710717717765626,
+ "relative_l2_max": 0.011400289643805562,
+ "weighted_relative_l2_mean": 0.005094142639764323,
+ "weighted_relative_frobenius": 0.005507144397623611,
+ "weighted_mean_cosine": 0.999985464118255,
+ "max_relative_norm_change": 0.0030795821881461597
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.955942,
+ "warm_p50_ms": 0.11028,
+ "warm_p95_ms": 0.11426389999999999,
+ "warm_p99_ms": 0.13382515999999997,
+ "samples_ms": [
+ 0.119639,
+ 0.113981,
+ 0.112095,
+ 0.112515,
+ 0.108394,
+ 0.111746,
+ 0.112165,
+ 0.111257,
+ 0.112515,
+ 0.112515,
+ 0.109721,
+ 0.111188,
+ 0.110279,
+ 0.110209,
+ 0.109791,
+ 0.111188,
+ 0.109302,
+ 0.108324,
+ 0.109721,
+ 0.111188,
+ 0.11028,
+ 0.112655,
+ 0.107835,
+ 0.142895,
+ 0.113562,
+ 0.10979,
+ 0.111117,
+ 0.109371,
+ 0.110699,
+ 0.109302,
+ 0.107905,
+ 0.110209,
+ 0.110699,
+ 0.10979,
+ 0.11028,
+ 0.110629,
+ 0.10986,
+ 0.110209,
+ 0.109721,
+ 0.11028
+ ],
+ "warm_rows_per_second": 4642727.602466449
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761994707803187,
+ "character_perplexity": 5.824043079756192,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.82039132299268e-06,
+ "mean_prediction_kl_nats": 3.6558254564872534e-06,
+ "prediction_kl_p99_nats": 2.1573732474216604e-05,
+ "maximum_prediction_kl_nats": 9.125894225908877e-05,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.029454708099365234,
+ "maximum_per_target_cross_entropy_increase": 0.02509140968322754
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 893386,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 98,
+ "max_block_bytes": 8192,
+ "raw_blocks": 12,
+ "compressed_blocks": 86,
+ "zlib_blocks": 86,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "f6ffab41a644dd67e026935d7b50595b09e149dee71f6d058748771dac0f90df",
+ "table_sha256": "7abea1c680e92a405165f5a1849259c285226ac51a1c93562204d21ecce8820c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.12420411384664476,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 894631,
+ "backend_with_key_registry_bytes": 1056356,
+ "backend_with_datrie_key_index_bytes": 989911,
+ "full_model_serving_artifact_bytes": 1351042,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999355787040166,
+ "mean_cosine": 0.9999869086187041,
+ "p05_cosine": 0.9999750882928043,
+ "relative_l2_mean": 0.00500744562114531,
+ "relative_l2_p95": 0.00710717717765626,
+ "relative_l2_max": 0.011400289643805562,
+ "weighted_relative_l2_mean": 0.005094142639764323,
+ "weighted_relative_frobenius": 0.005507144397623611,
+ "weighted_mean_cosine": 0.999985464118255,
+ "max_relative_norm_change": 0.0030795821881461597
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.91753,
+ "warm_p50_ms": 0.110489,
+ "warm_p95_ms": 0.11328269999999999,
+ "warm_p99_ms": 0.12083590999999999,
+ "samples_ms": [
+ 0.115937,
+ 0.112515,
+ 0.113073,
+ 0.110699,
+ 0.11028,
+ 0.113143,
+ 0.110698,
+ 0.108813,
+ 0.110769,
+ 0.109302,
+ 0.112095,
+ 0.123968,
+ 0.111117,
+ 0.111117,
+ 0.110768,
+ 0.109302,
+ 0.111118,
+ 0.10979,
+ 0.109232,
+ 0.11028,
+ 0.11021,
+ 0.109232,
+ 0.109302,
+ 0.11021,
+ 0.111187,
+ 0.111187,
+ 0.111187,
+ 0.111117,
+ 0.110698,
+ 0.10979,
+ 0.109721,
+ 0.108882,
+ 0.108743,
+ 0.110699,
+ 0.108813,
+ 0.10972,
+ 0.109791,
+ 0.11021,
+ 0.113074,
+ 0.11021
+ ],
+ "warm_rows_per_second": 4633945.46063409
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761994707803187,
+ "character_perplexity": 5.824043079756192,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.82039132299268e-06,
+ "mean_prediction_kl_nats": 3.6558254564872534e-06,
+ "prediction_kl_p99_nats": 2.1573732474216604e-05,
+ "maximum_prediction_kl_nats": 9.125894225908877e-05,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.029454708099365234,
+ "maximum_per_target_cross_entropy_increase": 0.02509140968322754
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 893386,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 98,
+ "max_block_bytes": 8192,
+ "raw_blocks": 12,
+ "compressed_blocks": 86,
+ "zlib_blocks": 86,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "f6ffab41a644dd67e026935d7b50595b09e149dee71f6d058748771dac0f90df",
+ "table_sha256": "7abea1c680e92a405165f5a1849259c285226ac51a1c93562204d21ecce8820c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08165040798485279,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 894631,
+ "backend_with_key_registry_bytes": 1056356,
+ "backend_with_datrie_key_index_bytes": 989911,
+ "full_model_serving_artifact_bytes": 1351042,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999355787040166,
+ "mean_cosine": 0.9999869086187041,
+ "p05_cosine": 0.9999750882928043,
+ "relative_l2_mean": 0.00500744562114531,
+ "relative_l2_p95": 0.00710717717765626,
+ "relative_l2_max": 0.011400289643805562,
+ "weighted_relative_l2_mean": 0.005094142639764323,
+ "weighted_relative_frobenius": 0.005507144397623611,
+ "weighted_mean_cosine": 0.999985464118255,
+ "max_relative_norm_change": 0.0030795821881461597
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.886311,
+ "warm_p50_ms": 0.111118,
+ "warm_p95_ms": 0.1136074,
+ "warm_p99_ms": 0.12426904,
+ "samples_ms": [
+ 0.113562,
+ 0.113143,
+ 0.112235,
+ 0.111746,
+ 0.111118,
+ 0.111677,
+ 0.111537,
+ 0.11028,
+ 0.111677,
+ 0.110699,
+ 0.112026,
+ 0.10979,
+ 0.111257,
+ 0.110629,
+ 0.109791,
+ 0.130534,
+ 0.11028,
+ 0.111606,
+ 0.110209,
+ 0.10979,
+ 0.108324,
+ 0.110279,
+ 0.111746,
+ 0.110698,
+ 0.11021,
+ 0.108813,
+ 0.10965,
+ 0.110349,
+ 0.111118,
+ 0.111118,
+ 0.111118,
+ 0.111746,
+ 0.111188,
+ 0.111188,
+ 0.111607,
+ 0.111607,
+ 0.11028,
+ 0.110209,
+ 0.10972,
+ 0.11447
+ ],
+ "warm_rows_per_second": 4607714.321712054
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761994707803187,
+ "character_perplexity": 5.824043079756192,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.82039132299268e-06,
+ "mean_prediction_kl_nats": 3.6558254564872534e-06,
+ "prediction_kl_p99_nats": 2.1573732474216604e-05,
+ "maximum_prediction_kl_nats": 9.125894225908877e-05,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.029454708099365234,
+ "maximum_per_target_cross_entropy_increase": 0.02509140968322754
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 843028,
+ "unique_rows": 11025,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 10,
+ "compressed_blocks": 82,
+ "zlib_blocks": 82,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11025
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "baddebcdadaa5e9ed9bd10c9957011f70f2e1711d3e0a79c782e80febbef68fd",
+ "table_sha256": "500510c305ae5f3d6b1008b379263ae6994195ab4eb0998621e26fb69f412ffe",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08750807098113,
+ "representatives_sha256": "7b7c1e628ffdfed37d646fe62d039db827254573cf58796ef56099be67fe76a6",
+ "structurally_merged_rows": 735,
+ "backend_artifact_bytes": 844272,
+ "backend_with_key_registry_bytes": 1005997,
+ "backend_with_datrie_key_index_bytes": 939552,
+ "full_model_serving_artifact_bytes": 1300683,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9932693001510837,
+ "mean_cosine": 0.9999451072569152,
+ "p05_cosine": 0.9996889849616917,
+ "relative_l2_mean": 0.00690180114140664,
+ "relative_l2_p95": 0.025131177274422562,
+ "relative_l2_max": 0.11602073490844073,
+ "weighted_relative_l2_mean": 0.0058481960078201126,
+ "weighted_relative_frobenius": 0.007637971226019644,
+ "weighted_mean_cosine": 0.9999712554419257,
+ "max_relative_norm_change": 0.010412935979967038
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9932108978475287,
+ "mean_cosine": 0.9999568063857526,
+ "p05_cosine": 0.9996803623187015,
+ "relative_l2_mean": 0.0022034061731780164,
+ "relative_l2_p95": 0.025527870272695073,
+ "relative_l2_max": 0.11652798676048805,
+ "weighted_relative_l2_mean": 0.0009280876904894124,
+ "weighted_relative_frobenius": 0.005479980336186081,
+ "weighted_mean_cosine": 0.9999850011202908,
+ "max_relative_norm_change": 0.009931715881154553
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.818774,
+ "warm_p50_ms": 0.109511,
+ "warm_p95_ms": 0.11263625,
+ "warm_p99_ms": 0.12160494999999999,
+ "samples_ms": [
+ 0.114959,
+ 0.110698,
+ 0.110279,
+ 0.110209,
+ 0.10972,
+ 0.111606,
+ 0.112514,
+ 0.10986,
+ 0.11021,
+ 0.11021,
+ 0.109302,
+ 0.111118,
+ 0.125854,
+ 0.109302,
+ 0.109721,
+ 0.109232,
+ 0.108883,
+ 0.108813,
+ 0.108813,
+ 0.108324,
+ 0.108813,
+ 0.109302,
+ 0.110698,
+ 0.109232,
+ 0.109302,
+ 0.109301,
+ 0.110768,
+ 0.109721,
+ 0.109302,
+ 0.110768,
+ 0.107905,
+ 0.109301,
+ 0.108813,
+ 0.111117,
+ 0.111257,
+ 0.109791,
+ 0.108813,
+ 0.109302,
+ 0.108743,
+ 0.109301
+ ],
+ "warm_rows_per_second": 4675329.418962479
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7619870145379721,
+ "character_perplexity": 5.823998274020508,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -8.728967694862381e-07,
+ "mean_prediction_kl_nats": 5.528766134507814e-06,
+ "prediction_kl_p99_nats": 4.193367377456218e-05,
+ "maximum_prediction_kl_nats": 0.00917243119023907,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.30211687088012695,
+ "maximum_per_target_cross_entropy_increase": 0.04476737976074219
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -7.69323562160144e-06,
+ "mean_prediction_kl_nats": 1.885823733372095e-06,
+ "prediction_kl_p99_nats": 3.158589785402744e-05,
+ "maximum_prediction_kl_nats": 0.009249863092956145,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.30378103256225586,
+ "maximum_per_target_cross_entropy_increase": 0.041460514068603516
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842894,
+ "unique_rows": 11023,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 11,
+ "compressed_blocks": 81,
+ "zlib_blocks": 81,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11023
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "cab1f9f7e8f4959b0926979e6921bc461c4e345d8e9cfb1581dcd2010c1c7fac",
+ "table_sha256": "033187898fa39c8d68eb9a9fc3110f8bb7c02a7ab434d4575ced7ddad6a1987b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08755814679898322,
+ "representatives_sha256": "6981e9856c2b8822a5be05de16e0ce8076b5640fe3c60b327f3c856fc8916fd0",
+ "structurally_merged_rows": 737,
+ "backend_artifact_bytes": 844138,
+ "backend_with_key_registry_bytes": 1005863,
+ "backend_with_datrie_key_index_bytes": 939418,
+ "full_model_serving_artifact_bytes": 1300549,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9813272660500127,
+ "mean_cosine": 0.9999425033140585,
+ "p05_cosine": 0.9996876099233718,
+ "relative_l2_mean": 0.006930306178244006,
+ "relative_l2_p95": 0.02523025606963592,
+ "relative_l2_max": 0.19399595707235132,
+ "weighted_relative_l2_mean": 0.00585568583280781,
+ "weighted_relative_frobenius": 0.007698795623847522,
+ "weighted_mean_cosine": 0.9999706467512589,
+ "max_relative_norm_change": 0.010412935979967038
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9810029673989541,
+ "mean_cosine": 0.9999541660152366,
+ "p05_cosine": 0.9996789564990175,
+ "relative_l2_mean": 0.0022332823861725636,
+ "relative_l2_p95": 0.025565944541053185,
+ "relative_l2_max": 0.1955265801156805,
+ "weighted_relative_l2_mean": 0.0009360131086985215,
+ "weighted_relative_frobenius": 0.005565058598246341,
+ "weighted_mean_cosine": 0.9999843880267955,
+ "max_relative_norm_change": 0.009931715881154553
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.669174,
+ "warm_p50_ms": 0.11171149999999999,
+ "warm_p95_ms": 0.11365279999999998,
+ "warm_p99_ms": 0.12291881999999998,
+ "samples_ms": [
+ 0.115378,
+ 0.113003,
+ 0.112165,
+ 0.111677,
+ 0.111258,
+ 0.112654,
+ 0.111676,
+ 0.112654,
+ 0.112165,
+ 0.109721,
+ 0.110698,
+ 0.110279,
+ 0.12774,
+ 0.111607,
+ 0.110699,
+ 0.110768,
+ 0.112654,
+ 0.113003,
+ 0.113074,
+ 0.110629,
+ 0.110699,
+ 0.112654,
+ 0.112585,
+ 0.113492,
+ 0.112165,
+ 0.113562,
+ 0.112585,
+ 0.113073,
+ 0.111746,
+ 0.111676,
+ 0.111606,
+ 0.112514,
+ 0.110209,
+ 0.10979,
+ 0.111188,
+ 0.110629,
+ 0.111677,
+ 0.111258,
+ 0.112235,
+ 0.11021
+ ],
+ "warm_rows_per_second": 4583234.492420208
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761990321590758,
+ "character_perplexity": 5.824017534322072,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.4341585978161666e-06,
+ "mean_prediction_kl_nats": 7.263649353478686e-06,
+ "prediction_kl_p99_nats": 4.307788164271621e-05,
+ "maximum_prediction_kl_nats": 0.011499851084857954,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.30211687088012695,
+ "maximum_per_target_cross_entropy_increase": 0.10541251301765442
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.386202809556931e-06,
+ "mean_prediction_kl_nats": 3.6478026720210987e-06,
+ "prediction_kl_p99_nats": 3.1849514343602146e-05,
+ "maximum_prediction_kl_nats": 0.01170144035072026,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.30378103256225586,
+ "maximum_per_target_cross_entropy_increase": 0.1062164306640625
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842894,
+ "unique_rows": 11023,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 11,
+ "compressed_blocks": 81,
+ "zlib_blocks": 81,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11023
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "cab1f9f7e8f4959b0926979e6921bc461c4e345d8e9cfb1581dcd2010c1c7fac",
+ "table_sha256": "033187898fa39c8d68eb9a9fc3110f8bb7c02a7ab434d4575ced7ddad6a1987b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08710683300159872,
+ "representatives_sha256": "6981e9856c2b8822a5be05de16e0ce8076b5640fe3c60b327f3c856fc8916fd0",
+ "structurally_merged_rows": 737,
+ "backend_artifact_bytes": 844138,
+ "backend_with_key_registry_bytes": 1005863,
+ "backend_with_datrie_key_index_bytes": 939418,
+ "full_model_serving_artifact_bytes": 1300549,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9813272660500127,
+ "mean_cosine": 0.9999425033140585,
+ "p05_cosine": 0.9996876099233718,
+ "relative_l2_mean": 0.006930306178244006,
+ "relative_l2_p95": 0.02523025606963592,
+ "relative_l2_max": 0.19399595707235132,
+ "weighted_relative_l2_mean": 0.00585568583280781,
+ "weighted_relative_frobenius": 0.007698795623847522,
+ "weighted_mean_cosine": 0.9999706467512589,
+ "max_relative_norm_change": 0.010412935979967038
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9810029673989541,
+ "mean_cosine": 0.9999541660152366,
+ "p05_cosine": 0.9996789564990175,
+ "relative_l2_mean": 0.0022332823861725636,
+ "relative_l2_p95": 0.025565944541053185,
+ "relative_l2_max": 0.1955265801156805,
+ "weighted_relative_l2_mean": 0.0009360131086985215,
+ "weighted_relative_frobenius": 0.005565058598246341,
+ "weighted_mean_cosine": 0.9999843880267955,
+ "max_relative_norm_change": 0.009931715881154553
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.67923,
+ "warm_p50_ms": 0.111153,
+ "warm_p95_ms": 0.11682735,
+ "warm_p99_ms": 0.13251994999999997,
+ "samples_ms": [
+ 0.117822,
+ 0.116775,
+ 0.112654,
+ 0.112654,
+ 0.116355,
+ 0.112025,
+ 0.110768,
+ 0.112235,
+ 0.112165,
+ 0.112654,
+ 0.110628,
+ 0.110349,
+ 0.11021,
+ 0.11014,
+ 0.10986,
+ 0.111257,
+ 0.110698,
+ 0.11014,
+ 0.110769,
+ 0.111188,
+ 0.111676,
+ 0.116355,
+ 0.11028,
+ 0.111257,
+ 0.110699,
+ 0.111118,
+ 0.111677,
+ 0.112654,
+ 0.111188,
+ 0.11021,
+ 0.11014,
+ 0.108882,
+ 0.110699,
+ 0.109861,
+ 0.110698,
+ 0.141917,
+ 0.115378,
+ 0.112095,
+ 0.110699,
+ 0.11028
+ ],
+ "warm_rows_per_second": 4606263.438683616
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.761990321590758,
+ "character_perplexity": 5.824017534322072,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.4341585978161666e-06,
+ "mean_prediction_kl_nats": 7.263649353478686e-06,
+ "prediction_kl_p99_nats": 4.307788164271621e-05,
+ "maximum_prediction_kl_nats": 0.011499851084857954,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.30211687088012695,
+ "maximum_per_target_cross_entropy_increase": 0.10541251301765442
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.386202809556931e-06,
+ "mean_prediction_kl_nats": 3.6478026720210987e-06,
+ "prediction_kl_p99_nats": 3.1849514343602146e-05,
+ "maximum_prediction_kl_nats": 0.01170144035072026,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.30378103256225586,
+ "maximum_per_target_cross_entropy_increase": 0.1062164306640625
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842756,
+ "unique_rows": 11021,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 10,
+ "compressed_blocks": 82,
+ "zlib_blocks": 82,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11021
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "ae6184a6148d23a1de6ec0fb9daeaad970c95ee9945259fe81029e0c54871e7f",
+ "table_sha256": "a72fd3463fc255f7b9558e25c7372ac1667c0c1e194ec0e99f438e49d8eff6ba",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08703964482992887,
+ "representatives_sha256": "29fc918a0f2f40e380f0476e35c73f352c984792f4332f4b2ed9d42ed215d0d6",
+ "structurally_merged_rows": 739,
+ "backend_artifact_bytes": 844000,
+ "backend_with_key_registry_bytes": 1005725,
+ "backend_with_datrie_key_index_bytes": 939280,
+ "full_model_serving_artifact_bytes": 1300411,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9536868525209022,
+ "mean_cosine": 0.9999346955272335,
+ "p05_cosine": 0.9996859107993521,
+ "relative_l2_mean": 0.006980950515669446,
+ "relative_l2_p95": 0.02526545387447921,
+ "relative_l2_max": 0.3043503712082752,
+ "weighted_relative_l2_mean": 0.0060684378003306744,
+ "weighted_relative_frobenius": 0.011322443046301875,
+ "weighted_mean_cosine": 0.9999377175024109,
+ "max_relative_norm_change": 0.010412935979967038
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9536637403364369,
+ "mean_cosine": 0.9999463655955405,
+ "p05_cosine": 0.9996754540293509,
+ "relative_l2_mean": 0.002284693356695273,
+ "relative_l2_p95": 0.025726232346550162,
+ "relative_l2_max": 0.3043045745324806,
+ "weighted_relative_l2_mean": 0.0011525003754382704,
+ "weighted_relative_frobenius": 0.009999323492696645,
+ "weighted_mean_cosine": 0.9999514403371849,
+ "max_relative_norm_change": 0.009931715881154553
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.776101,
+ "warm_p50_ms": 0.110698,
+ "warm_p95_ms": 0.116844,
+ "warm_p99_ms": 0.12408714,
+ "samples_ms": [
+ 0.115029,
+ 0.116844,
+ 0.113562,
+ 0.111188,
+ 0.110279,
+ 0.11028,
+ 0.110139,
+ 0.128718,
+ 0.109372,
+ 0.111187,
+ 0.116844,
+ 0.110768,
+ 0.110698,
+ 0.11028,
+ 0.111257,
+ 0.109721,
+ 0.110699,
+ 0.110698,
+ 0.10979,
+ 0.109651,
+ 0.109861,
+ 0.110698,
+ 0.111187,
+ 0.10979,
+ 0.109721,
+ 0.109721,
+ 0.111187,
+ 0.111257,
+ 0.113981,
+ 0.111118,
+ 0.111607,
+ 0.109721,
+ 0.10979,
+ 0.11021,
+ 0.111187,
+ 0.110768,
+ 0.11021,
+ 0.112584,
+ 0.11021,
+ 0.109302
+ ],
+ "warm_rows_per_second": 4625196.480514553
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7621143751519313,
+ "character_perplexity": 5.82474006925313,
+ "next_character_accuracy": 0.5003833767827021,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0001264957167171801,
+ "mean_prediction_kl_nats": 9.39694650236695e-05,
+ "prediction_kl_p99_nats": 5.843497417067707e-05,
+ "maximum_prediction_kl_nats": 0.05044044031322442,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.6701536178588867,
+ "maximum_per_target_cross_entropy_increase": 0.3766803741455078
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00011967450916720352,
+ "mean_prediction_kl_nats": 9.010682943788051e-05,
+ "prediction_kl_p99_nats": 4.83673399799821e-05,
+ "maximum_prediction_kl_nats": 0.050230559880163644,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.668968677520752,
+ "maximum_per_target_cross_entropy_increase": 0.37651634216308594
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842391,
+ "unique_rows": 11015,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 11,
+ "compressed_blocks": 81,
+ "zlib_blocks": 81,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11015
+ },
+ "source_float32_sha256": "04a26ae7fd213c12d1608cab148bca035b427fba2b336ac8e1f3f738a79bbe5b",
+ "decoded_float32_sha256": "db1bc93a2b33fe8bf16905e9c9c6352e5a5ef58010129e3203f46ec0141d9a36",
+ "table_sha256": "8a947de03d30fb5b4640ee6f2f5a7b57819c767f5b2fc00a13e7328245cf9323",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.10442022397182882,
+ "representatives_sha256": "7df31a8d8f1072a211185674b8ae54ed2d4d8ab575dbd298cf13d13e659db635",
+ "structurally_merged_rows": 745,
+ "backend_artifact_bytes": 843635,
+ "backend_with_key_registry_bytes": 1005360,
+ "backend_with_datrie_key_index_bytes": 938915,
+ "full_model_serving_artifact_bytes": 1300046,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9013648960972993,
+ "mean_cosine": 0.9998944903338037,
+ "p05_cosine": 0.9996827477502208,
+ "relative_l2_mean": 0.007179385016144764,
+ "relative_l2_p95": 0.025540107620820828,
+ "relative_l2_max": 0.44293560563358964,
+ "weighted_relative_l2_mean": 0.006433061655660153,
+ "weighted_relative_frobenius": 0.015764260127780892,
+ "weighted_mean_cosine": 0.9998666365352867,
+ "max_relative_norm_change": 0.010412935979967038
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9019021061884321,
+ "mean_cosine": 0.9999062014743357,
+ "p05_cosine": 0.9996731657436946,
+ "relative_l2_mean": 0.002485952204778026,
+ "relative_l2_p95": 0.025836427923357233,
+ "relative_l2_max": 0.44148202456311975,
+ "weighted_relative_l2_mean": 0.0015226214218138445,
+ "weighted_relative_frobenius": 0.014842333695256017,
+ "weighted_mean_cosine": 0.9998802509623124,
+ "max_relative_norm_change": 0.009968830078026136
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.746348,
+ "warm_p50_ms": 0.1116765,
+ "warm_p95_ms": 0.1151198,
+ "warm_p99_ms": 0.13588858999999998,
+ "samples_ms": [
+ 0.115029,
+ 0.114959,
+ 0.114469,
+ 0.11454,
+ 0.112584,
+ 0.112584,
+ 0.113143,
+ 0.112165,
+ 0.113981,
+ 0.112654,
+ 0.111746,
+ 0.111188,
+ 0.112514,
+ 0.110699,
+ 0.110769,
+ 0.110699,
+ 0.110629,
+ 0.112096,
+ 0.111188,
+ 0.111188,
+ 0.116845,
+ 0.110629,
+ 0.112584,
+ 0.112585,
+ 0.110769,
+ 0.148064,
+ 0.112166,
+ 0.111677,
+ 0.114051,
+ 0.110209,
+ 0.111676,
+ 0.111536,
+ 0.110629,
+ 0.110349,
+ 0.11021,
+ 0.11021,
+ 0.110279,
+ 0.111187,
+ 0.110628,
+ 0.110699
+ ],
+ "warm_rows_per_second": 4584670.90211459
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7621814550112926,
+ "character_perplexity": 5.825130805102898,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0001935863115821644,
+ "mean_prediction_kl_nats": 0.00013970315832643,
+ "prediction_kl_p99_nats": 0.00010846144734936316,
+ "maximum_prediction_kl_nats": 0.0683686249485541,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 1.8703510761260986,
+ "maximum_per_target_cross_entropy_increase": 0.3766803741455078
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00018676464645119763,
+ "mean_prediction_kl_nats": 0.00013571551772811213,
+ "prediction_kl_p99_nats": 0.00010863712816719506,
+ "maximum_prediction_kl_nats": 0.06812395448033506,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.8682162761688232,
+ "maximum_per_target_cross_entropy_increase": 0.37651634216308594
+ }
+ }
+ ],
+ "baseline_quality": {
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7619878874351227,
+ "character_perplexity": 5.824003357774225,
+ "next_character_accuracy": 0.5003067014261616,
+ "unknown_targets": 0
+ },
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6155188252771413,
+ "character_perplexity": 5.030497195633726,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ }
+ },
+ "selection": {
+ "fp32": {
+ "feasible": true,
+ "name": "fp32-joint-0.95",
+ "backend_with_key_registry_bytes": 2894980,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6159370315575023,
+ "character_perplexity": 5.0326014211237675,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00041829374079927037,
+ "mean_prediction_kl_nats": 0.0001224659868776353,
+ "prediction_kl_p99_nats": 6.492612482710993e-05,
+ "maximum_prediction_kl_nats": 0.049823547914828695,
+ "top1_changed_fraction": 0.0002300084336425669,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6587400436401367,
+ "maximum_per_target_cross_entropy_increase": 0.3685579299926758
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 9.471540831570154e-05,
+ 0.0007101388272276771
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "fp16": {
+ "feasible": true,
+ "name": "fp16-joint-0.95",
+ "backend_with_key_registry_bytes": 1567336,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6159358449097074,
+ "character_perplexity": 5.032595449201931,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.0004171065973414183,
+ "mean_prediction_kl_nats": 0.0001224485145529409,
+ "prediction_kl_p99_nats": 6.468627058898175e-05,
+ "maximum_prediction_kl_nats": 0.049807480276882234,
+ "top1_changed_fraction": 0.0002300084336425669,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6586532592773438,
+ "maximum_per_target_cross_entropy_increase": 0.3685433864593506
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 9.510922932824986e-05,
+ 0.0007089831271132332
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "int8": {
+ "feasible": true,
+ "name": "int8-raw",
+ "backend_with_key_registry_bytes": 962252,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.615536524896122,
+ "character_perplexity": 5.0305862343053445,
+ "next_character_accuracy": 0.5290960668557847,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 1.7699775619881114e-05,
+ "mean_prediction_kl_nats": 3.5440430999465437e-06,
+ "prediction_kl_p99_nats": 2.424724186036613e-05,
+ "maximum_prediction_kl_nats": 0.00011723431737023591,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": -0.00015333895576175038,
+ "maximum_absolute_logit_change": 0.02951216697692871,
+ "maximum_per_target_cross_entropy_increase": 0.0291440486907959
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -3.6594241267184036e-05,
+ 7.115336931680388e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ }
+ },
+ "fixed_quantization_test_controls": {
+ "fp16": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6155178001573416,
+ "character_perplexity": 5.030492038774091,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -1.0251192743875722e-06,
+ "mean_prediction_kl_nats": 4.051686320711723e-09,
+ "prediction_kl_p99_nats": 2.4765548518800662e-08,
+ "maximum_prediction_kl_nats": 7.780124816940553e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0011723041534423828,
+ "maximum_per_target_cross_entropy_increase": 0.0008974075317382812
+ }
+ },
+ "int8": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.615536524896122,
+ "character_perplexity": 5.0305862343053445,
+ "next_character_accuracy": 0.5290960668557847,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": 1.7699775619881114e-05,
+ "mean_prediction_kl_nats": 3.5440430999465437e-06,
+ "prediction_kl_p99_nats": 2.424724186036613e-05,
+ "maximum_prediction_kl_nats": 0.00011723431737023591,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": -0.00015333895576175038,
+ "maximum_absolute_logit_change": 0.02951216697692871,
+ "maximum_per_target_cross_entropy_increase": 0.0291440486907959
+ }
+ }
+ },
+ "checkpoint_sha256": "2a6ed05f8dd35bbd2851b5fa21fc34024df681e447fa8e426f5a8888bd13a567",
+ "registry_sha256": "7fa792a7a13db17dc47320b6a65018697d338beda3b6328e9775d68716636b54"
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 11,
+ "rows": 11760,
+ "dim": 64,
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "source_float32_bytes": 3010560,
+ "key_registry_bytes": 161725,
+ "unchanged_backbone_artifact_bytes": 294686,
+ "alternative_datrie_key_index_bytes": 95280,
+ "query_rows": 512,
+ "query_ids_sha256": "8ba8e1ab159a822d9159f679166b91955ebc405920cda85ccf24f213c3eff2ad",
+ "frequency_trace_sha256": "73ddaf0153da6d907ac6eef82bf819aaec23a0a5eb289cb0f3f91092c37eb7f0",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 3010560,
+ "unique_rows": 11760,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "table_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.014116892823949456,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3011364,
+ "backend_with_key_registry_bytes": 3173089,
+ "backend_with_datrie_key_index_bytes": 3106644,
+ "full_model_serving_artifact_bytes": 3467775,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.380356,
+ "warm_p50_ms": 0.048749,
+ "warm_p95_ms": 0.05079829999999999,
+ "warm_p99_ms": 0.06405067999999998,
+ "samples_ms": [
+ 0.050704,
+ 0.049657,
+ 0.049238,
+ 0.050635,
+ 0.050566,
+ 0.048261,
+ 0.048261,
+ 0.071378,
+ 0.04868,
+ 0.048749,
+ 0.049238,
+ 0.049238,
+ 0.048679,
+ 0.048679,
+ 0.048749,
+ 0.049727,
+ 0.05259,
+ 0.047422,
+ 0.047353,
+ 0.047772,
+ 0.047772,
+ 0.047772,
+ 0.047772,
+ 0.047771,
+ 0.049658,
+ 0.04875,
+ 0.04875,
+ 0.048331,
+ 0.04868,
+ 0.048191,
+ 0.049168,
+ 0.048749,
+ 0.048749,
+ 0.049168,
+ 0.048749,
+ 0.049168,
+ 0.04875,
+ 0.048679,
+ 0.04868,
+ 0.04826
+ ],
+ "warm_rows_per_second": 10502779.544195777
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519091112373895,
+ "character_perplexity": 5.765599346234589,
+ "next_character_accuracy": 0.5006900782088637,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2908412,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 368,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 368,
+ "zlib_blocks": 368,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "table_sha256": "7e93ec1a1a2a6f90dacb21b8baea14bcb657151f929b151963e2c31ab687088b",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2582893830258399,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2909618,
+ "backend_with_key_registry_bytes": 3071343,
+ "backend_with_datrie_key_index_bytes": 3004898,
+ "full_model_serving_artifact_bytes": 3366029,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.772598,
+ "warm_p50_ms": 0.056327,
+ "warm_p95_ms": 0.059204849999999996,
+ "warm_p99_ms": 0.06822830999999999,
+ "samples_ms": [
+ 0.060133,
+ 0.059156,
+ 0.058666,
+ 0.057759,
+ 0.055803,
+ 0.058177,
+ 0.057689,
+ 0.058178,
+ 0.056851,
+ 0.05727,
+ 0.056781,
+ 0.056781,
+ 0.055873,
+ 0.056292,
+ 0.056711,
+ 0.056292,
+ 0.056223,
+ 0.056711,
+ 0.056292,
+ 0.055385,
+ 0.055874,
+ 0.073404,
+ 0.055315,
+ 0.056292,
+ 0.05727,
+ 0.055804,
+ 0.055803,
+ 0.057689,
+ 0.056292,
+ 0.056292,
+ 0.056781,
+ 0.055803,
+ 0.055803,
+ 0.056292,
+ 0.056781,
+ 0.056292,
+ 0.056223,
+ 0.055734,
+ 0.056362,
+ 0.056362
+ ],
+ "warm_rows_per_second": 9089779.324302731
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519091112373895,
+ "character_perplexity": 5.765599346234589,
+ "next_character_accuracy": 0.5006900782088637,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2908412,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 368,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 368,
+ "zlib_blocks": 368,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "table_sha256": "7e93ec1a1a2a6f90dacb21b8baea14bcb657151f929b151963e2c31ab687088b",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3099501240067184,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2909617,
+ "backend_with_key_registry_bytes": 3071342,
+ "backend_with_datrie_key_index_bytes": 3004897,
+ "full_model_serving_artifact_bytes": 3366028,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.709182,
+ "warm_p50_ms": 0.05727,
+ "warm_p95_ms": 0.059179449999999995,
+ "warm_p99_ms": 0.05964461,
+ "samples_ms": [
+ 0.059645,
+ 0.059155,
+ 0.058667,
+ 0.058667,
+ 0.059155,
+ 0.058178,
+ 0.059644,
+ 0.058178,
+ 0.05727,
+ 0.056851,
+ 0.057759,
+ 0.057689,
+ 0.056781,
+ 0.057689,
+ 0.056711,
+ 0.05727,
+ 0.056781,
+ 0.057689,
+ 0.05727,
+ 0.057688,
+ 0.056292,
+ 0.05727,
+ 0.057759,
+ 0.056781,
+ 0.056781,
+ 0.056781,
+ 0.058737,
+ 0.056292,
+ 0.057689,
+ 0.057758,
+ 0.056781,
+ 0.055873,
+ 0.057269,
+ 0.05727,
+ 0.056781,
+ 0.056292,
+ 0.0572,
+ 0.056781,
+ 0.057759,
+ 0.057689
+ ],
+ "warm_rows_per_second": 8940108.25912345
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519091112373895,
+ "character_perplexity": 5.765599346234589,
+ "next_character_accuracy": 0.5006900782088637,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2908412,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 368,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 368,
+ "zlib_blocks": 368,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "table_sha256": "7e93ec1a1a2a6f90dacb21b8baea14bcb657151f929b151963e2c31ab687088b",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.34943046304397285,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2909617,
+ "backend_with_key_registry_bytes": 3071342,
+ "backend_with_datrie_key_index_bytes": 3004897,
+ "full_model_serving_artifact_bytes": 3366028,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.762191,
+ "warm_p50_ms": 0.058876,
+ "warm_p95_ms": 0.06073724999999999,
+ "warm_p99_ms": 0.08201328999999997,
+ "samples_ms": [
+ 0.062927,
+ 0.060622,
+ 0.059575,
+ 0.058667,
+ 0.059993,
+ 0.058667,
+ 0.058597,
+ 0.058247,
+ 0.058667,
+ 0.058597,
+ 0.059086,
+ 0.059575,
+ 0.058178,
+ 0.059225,
+ 0.059156,
+ 0.059645,
+ 0.059086,
+ 0.058667,
+ 0.057689,
+ 0.058667,
+ 0.057759,
+ 0.059085,
+ 0.057689,
+ 0.059226,
+ 0.059155,
+ 0.094216,
+ 0.060622,
+ 0.059085,
+ 0.059156,
+ 0.058667,
+ 0.059155,
+ 0.057759,
+ 0.058178,
+ 0.058666,
+ 0.059156,
+ 0.058178,
+ 0.057759,
+ 0.058248,
+ 0.059156,
+ 0.05727
+ ],
+ "warm_rows_per_second": 8696242.951287452
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519091112373895,
+ "character_perplexity": 5.765599346234589,
+ "next_character_accuracy": 0.5006900782088637,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2731820,
+ "unique_rows": 11021,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11021
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "f0f3a94725b746f29df1fec1fabd3306d43db9fb0792c7a68bc4530d60b530c3",
+ "table_sha256": "9aa7bbd6f4c34211b5b934797f20c9a017278737c90ae6212e8b47ce378ec213",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2539064190350473,
+ "representatives_sha256": "b2a2c586d832f72a82e6bed54951d102fcd31cb52e09568675a2d5bb129adde2",
+ "structurally_merged_rows": 739,
+ "backend_artifact_bytes": 2733025,
+ "backend_with_key_registry_bytes": 2894750,
+ "backend_with_datrie_key_index_bytes": 2828305,
+ "full_model_serving_artifact_bytes": 3189436,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9936638928519256,
+ "mean_cosine": 0.9999572416826517,
+ "p05_cosine": 0.9996970082566761,
+ "relative_l2_mean": 0.002194145516018723,
+ "relative_l2_p95": 0.02484940231566842,
+ "relative_l2_max": 0.11261831137802518,
+ "weighted_relative_l2_mean": 0.0009363406367065997,
+ "weighted_relative_frobenius": 0.005649047524990139,
+ "weighted_mean_cosine": 0.99998411650889,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9936638928519256,
+ "mean_cosine": 0.9999572416826517,
+ "p05_cosine": 0.9996970082566761,
+ "relative_l2_mean": 0.002194145516018723,
+ "relative_l2_p95": 0.02484940231566842,
+ "relative_l2_max": 0.11261831137802518,
+ "weighted_relative_l2_mean": 0.0009363406367065997,
+ "weighted_relative_frobenius": 0.005649047524990139,
+ "weighted_mean_cosine": 0.99998411650889,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.45461,
+ "warm_p50_ms": 0.059156,
+ "warm_p95_ms": 0.06339834999999999,
+ "warm_p99_ms": 0.08134977999999998,
+ "samples_ms": [
+ 0.064393,
+ 0.06153,
+ 0.059086,
+ 0.059156,
+ 0.057689,
+ 0.059575,
+ 0.059156,
+ 0.058248,
+ 0.056781,
+ 0.059225,
+ 0.057689,
+ 0.058667,
+ 0.058247,
+ 0.058178,
+ 0.062508,
+ 0.060623,
+ 0.060553,
+ 0.060623,
+ 0.060552,
+ 0.061949,
+ 0.060553,
+ 0.058666,
+ 0.063346,
+ 0.058177,
+ 0.058667,
+ 0.057759,
+ 0.062019,
+ 0.060133,
+ 0.060552,
+ 0.058247,
+ 0.092191,
+ 0.062997,
+ 0.060134,
+ 0.058597,
+ 0.057758,
+ 0.058178,
+ 0.059644,
+ 0.058666,
+ 0.058108,
+ 0.058177
+ ],
+ "warm_rows_per_second": 8655081.47947799
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519929477123368,
+ "character_perplexity": 5.766082734022218,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 8.383998932280922e-05,
+ "mean_prediction_kl_nats": 2.770631969897495e-06,
+ "prediction_kl_p99_nats": 4.4623592160388185e-05,
+ "maximum_prediction_kl_nats": 0.00425984633843475,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.24453763663768768,
+ "maximum_per_target_cross_entropy_increase": 0.12383627891540527
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 8.383998932280922e-05,
+ "mean_prediction_kl_nats": 2.770631969897495e-06,
+ "prediction_kl_p99_nats": 4.4623592160388185e-05,
+ "maximum_prediction_kl_nats": 0.00425984633843475,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.24453763663768768,
+ "maximum_per_target_cross_entropy_increase": 0.12383627891540527
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2731340,
+ "unique_rows": 11019,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11019
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "bd0b4a1a6b410139110dc36c42d89a8fe7133faf72c00a8b8e465bfbb070043b",
+ "table_sha256": "a632f5033a2e071283a882dfd952bc1e215658cee10447df37ed9e9cc6e1db68",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24768152995966375,
+ "representatives_sha256": "42f9123f512fff55a21e537090dfe83b149f298c3a78d81be88beb17535f0836",
+ "structurally_merged_rows": 741,
+ "backend_artifact_bytes": 2732545,
+ "backend_with_key_registry_bytes": 2894270,
+ "backend_with_datrie_key_index_bytes": 2827825,
+ "full_model_serving_artifact_bytes": 3188956,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9836618144825191,
+ "mean_cosine": 0.9999548743456357,
+ "p05_cosine": 0.9996957661170864,
+ "relative_l2_mean": 0.002222446327007907,
+ "relative_l2_p95": 0.02487716326434536,
+ "relative_l2_max": 0.18064942308355184,
+ "weighted_relative_l2_mean": 0.0009556901082547855,
+ "weighted_relative_frobenius": 0.006003689527503709,
+ "weighted_mean_cosine": 0.9999826539700699,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9836618144825191,
+ "mean_cosine": 0.9999548743456357,
+ "p05_cosine": 0.9996957661170864,
+ "relative_l2_mean": 0.002222446327007907,
+ "relative_l2_p95": 0.02487716326434536,
+ "relative_l2_max": 0.18064942308355184,
+ "weighted_relative_l2_mean": 0.0009556901082547855,
+ "weighted_relative_frobenius": 0.006003689527503709,
+ "weighted_mean_cosine": 0.9999826539700699,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.235169,
+ "warm_p50_ms": 0.057759,
+ "warm_p95_ms": 0.06262319999999999,
+ "warm_p99_ms": 0.08036211999999998,
+ "samples_ms": [
+ 0.064812,
+ 0.060622,
+ 0.058597,
+ 0.062508,
+ 0.060064,
+ 0.060064,
+ 0.059645,
+ 0.059155,
+ 0.058667,
+ 0.058667,
+ 0.057758,
+ 0.057689,
+ 0.057689,
+ 0.057759,
+ 0.058177,
+ 0.056222,
+ 0.057269,
+ 0.05727,
+ 0.0572,
+ 0.05727,
+ 0.061879,
+ 0.05727,
+ 0.05727,
+ 0.05727,
+ 0.05727,
+ 0.05727,
+ 0.058666,
+ 0.058178,
+ 0.090304,
+ 0.058178,
+ 0.057689,
+ 0.056851,
+ 0.057758,
+ 0.057759,
+ 0.057689,
+ 0.057759,
+ 0.057759,
+ 0.057689,
+ 0.058178,
+ 0.0572
+ ],
+ "warm_rows_per_second": 8864419.397842761
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519840589324267,
+ "character_perplexity": 5.766031480809641,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 7.495050368583347e-05,
+ "mean_prediction_kl_nats": 2.9385354645919115e-06,
+ "prediction_kl_p99_nats": 4.492838028181003e-05,
+ "maximum_prediction_kl_nats": 0.00425984633843475,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.27823472023010254,
+ "maximum_per_target_cross_entropy_increase": 0.12383627891540527
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 7.495050368583347e-05,
+ "mean_prediction_kl_nats": 2.9385354645919115e-06,
+ "prediction_kl_p99_nats": 4.492838028181003e-05,
+ "maximum_prediction_kl_nats": 0.00425984633843475,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.27823472023010254,
+ "maximum_per_target_cross_entropy_increase": 0.12383627891540527
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2730652,
+ "unique_rows": 11016,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11016
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "78a46ec0ebb7935888f785beb23076c3208dbc1a9afa7da0f03aad268f4f5dc1",
+ "table_sha256": "20aedeb52dbafcf13f45a377bbcc139e1409d2f1788ee829e23e52ebbb15e0b4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2505855329800397,
+ "representatives_sha256": "a274b4e9c81850d6bc824846ad33d3d9de69d4b8e93ea38ace03538c3a26d240",
+ "structurally_merged_rows": 744,
+ "backend_artifact_bytes": 2731857,
+ "backend_with_key_registry_bytes": 2893582,
+ "backend_with_datrie_key_index_bytes": 2827137,
+ "full_model_serving_artifact_bytes": 3188268,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9715875739388654,
+ "mean_cosine": 0.9999462280900236,
+ "p05_cosine": 0.9996883188648951,
+ "relative_l2_mean": 0.0022973633980691386,
+ "relative_l2_p95": 0.02503481431327582,
+ "relative_l2_max": 0.23930394652713793,
+ "weighted_relative_l2_mean": 0.001033890052483249,
+ "weighted_relative_frobenius": 0.008099872832815666,
+ "weighted_mean_cosine": 0.999973591051415,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9715875739388654,
+ "mean_cosine": 0.9999462280900236,
+ "p05_cosine": 0.9996883188648951,
+ "relative_l2_mean": 0.0022973633980691386,
+ "relative_l2_p95": 0.02503481431327582,
+ "relative_l2_max": 0.23930394652713793,
+ "weighted_relative_l2_mean": 0.001033890052483249,
+ "weighted_relative_frobenius": 0.008099872832815666,
+ "weighted_mean_cosine": 0.999973591051415,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.04115,
+ "warm_p50_ms": 0.058178,
+ "warm_p95_ms": 0.06108639999999999,
+ "warm_p99_ms": 0.06288657,
+ "samples_ms": [
+ 0.061949,
+ 0.060482,
+ 0.060133,
+ 0.058667,
+ 0.058667,
+ 0.058666,
+ 0.059156,
+ 0.058597,
+ 0.063486,
+ 0.057689,
+ 0.057759,
+ 0.05727,
+ 0.057689,
+ 0.058178,
+ 0.057689,
+ 0.057689,
+ 0.061041,
+ 0.060063,
+ 0.057689,
+ 0.058667,
+ 0.058178,
+ 0.05727,
+ 0.057689,
+ 0.057689,
+ 0.058666,
+ 0.059086,
+ 0.058667,
+ 0.0572,
+ 0.057689,
+ 0.057689,
+ 0.058178,
+ 0.058178,
+ 0.058667,
+ 0.058178,
+ 0.057689,
+ 0.057688,
+ 0.05727,
+ 0.055803,
+ 0.058178,
+ 0.058178
+ ],
+ "warm_rows_per_second": 8800577.537900925
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7520013117528885,
+ "character_perplexity": 5.76613096197372,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 9.22047660971588e-05,
+ "mean_prediction_kl_nats": 4.803345681097564e-06,
+ "prediction_kl_p99_nats": 5.8310881644520935e-05,
+ "maximum_prediction_kl_nats": 0.00425984633843475,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.595242977142334,
+ "maximum_per_target_cross_entropy_increase": 0.29039692878723145
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 9.22047660971588e-05,
+ "mean_prediction_kl_nats": 4.803345681097564e-06,
+ "prediction_kl_p99_nats": 5.8310881644520935e-05,
+ "maximum_prediction_kl_nats": 0.00425984633843475,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.595242977142334,
+ "maximum_per_target_cross_entropy_increase": 0.29039692878723145
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2729646,
+ "unique_rows": 11012,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 345,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 345,
+ "zlib_blocks": 345,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11012
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "eac53495e3e81d55c8e8b309eaa412d79df265e5aa6dd1c737112b3086a284a3",
+ "table_sha256": "144b6021639f0c3807015461bf170b763a30b62940c441846d36479e2556ce6a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2529222150333226,
+ "representatives_sha256": "16c486f4c3258df64b1f3e8f70b8b977e76c8da0cadff578ce6b2e5bbe67a062",
+ "structurally_merged_rows": 748,
+ "backend_artifact_bytes": 2730851,
+ "backend_with_key_registry_bytes": 2892576,
+ "backend_with_datrie_key_index_bytes": 2826131,
+ "full_model_serving_artifact_bytes": 3187262,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9541012187652339,
+ "mean_cosine": 0.9999330769840775,
+ "p05_cosine": 0.9996863061040018,
+ "relative_l2_mean": 0.0023918135538841395,
+ "relative_l2_p95": 0.0251782915387511,
+ "relative_l2_max": 0.30426032785720214,
+ "weighted_relative_l2_mean": 0.001083013765408763,
+ "weighted_relative_frobenius": 0.009491379452074398,
+ "weighted_mean_cosine": 0.9999667978349295,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9541012187652339,
+ "mean_cosine": 0.9999330769840775,
+ "p05_cosine": 0.9996863061040018,
+ "relative_l2_mean": 0.0023918135538841395,
+ "relative_l2_p95": 0.0251782915387511,
+ "relative_l2_max": 0.30426032785720214,
+ "weighted_relative_l2_mean": 0.001083013765408763,
+ "weighted_relative_frobenius": 0.009491379452074398,
+ "weighted_mean_cosine": 0.9999667978349295,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.328756,
+ "warm_p50_ms": 0.057689,
+ "warm_p95_ms": 0.0596413,
+ "warm_p99_ms": 0.06878280999999999,
+ "samples_ms": [
+ 0.060901,
+ 0.059225,
+ 0.057689,
+ 0.057759,
+ 0.058178,
+ 0.059574,
+ 0.059575,
+ 0.057689,
+ 0.058666,
+ 0.058248,
+ 0.057759,
+ 0.057759,
+ 0.057201,
+ 0.0572,
+ 0.05727,
+ 0.056781,
+ 0.05727,
+ 0.056781,
+ 0.05727,
+ 0.0572,
+ 0.05727,
+ 0.0572,
+ 0.05727,
+ 0.05727,
+ 0.05727,
+ 0.05727,
+ 0.05727,
+ 0.058178,
+ 0.057759,
+ 0.056292,
+ 0.057689,
+ 0.056292,
+ 0.057758,
+ 0.057689,
+ 0.05727,
+ 0.05727,
+ 0.073822,
+ 0.059086,
+ 0.057759,
+ 0.057689
+ ],
+ "warm_rows_per_second": 8875175.510062577
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7521092260422866,
+ "character_perplexity": 5.766753243474986,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00020013482920029312,
+ "mean_prediction_kl_nats": 1.0234737776523216e-05,
+ "prediction_kl_p99_nats": 6.252705906023106e-05,
+ "maximum_prediction_kl_nats": 0.013867672945814268,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.6122770309448242,
+ "maximum_per_target_cross_entropy_increase": 0.8541910648345947
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00020013482920029312,
+ "mean_prediction_kl_nats": 1.0234737776523216e-05,
+ "prediction_kl_p99_nats": 6.252705906023106e-05,
+ "maximum_prediction_kl_nats": 0.013867672945814268,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.6122770309448242,
+ "maximum_per_target_cross_entropy_increase": 0.8541910648345947
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 3010560,
+ "artifact_bytes": 2728478,
+ "unique_rows": 11007,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 344,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 344,
+ "zlib_blocks": 344,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11007
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "6b18e9dd375b79d23d0886c70bf22971c101731ae4466c89ccaf7b515186aecc",
+ "table_sha256": "8b9acbe4b8e85269938eeab09061cb5e1c45520c2c3535cbbb534bb89f43cb19",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2824202699121088,
+ "representatives_sha256": "11e3c66f75f66346adec07ef5cd45185fff371452fb3e41586a7dc1a87b791c7",
+ "structurally_merged_rows": 753,
+ "backend_artifact_bytes": 2729683,
+ "backend_with_key_registry_bytes": 2891408,
+ "backend_with_datrie_key_index_bytes": 2824963,
+ "full_model_serving_artifact_bytes": 3186094,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.910428604511629,
+ "mean_cosine": 0.9998962694002413,
+ "p05_cosine": 0.9996822292021181,
+ "relative_l2_mean": 0.0025830225207301694,
+ "relative_l2_p95": 0.025385540658181598,
+ "relative_l2_max": 0.4227487665187151,
+ "weighted_relative_l2_mean": 0.001174464139748775,
+ "weighted_relative_frobenius": 0.010534642496976436,
+ "weighted_mean_cosine": 0.9999483978634752,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.910428604511629,
+ "mean_cosine": 0.9998962694002413,
+ "p05_cosine": 0.9996822292021181,
+ "relative_l2_mean": 0.0025830225207301694,
+ "relative_l2_p95": 0.025385540658181598,
+ "relative_l2_max": 0.4227487665187151,
+ "weighted_relative_l2_mean": 0.001174464139748775,
+ "weighted_relative_frobenius": 0.010534642496976436,
+ "weighted_mean_cosine": 0.9999483978634752,
+ "max_relative_norm_change": 0.010050490693820427
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 11.280985,
+ "warm_p50_ms": 0.057759,
+ "warm_p95_ms": 0.0625988,
+ "warm_p99_ms": 0.07156652999999999,
+ "samples_ms": [
+ 0.06195,
+ 0.062508,
+ 0.060133,
+ 0.059505,
+ 0.057689,
+ 0.060063,
+ 0.059644,
+ 0.057759,
+ 0.057689,
+ 0.057689,
+ 0.057758,
+ 0.057689,
+ 0.057759,
+ 0.058178,
+ 0.057689,
+ 0.058667,
+ 0.059155,
+ 0.057759,
+ 0.076197,
+ 0.064324,
+ 0.058108,
+ 0.057689,
+ 0.057689,
+ 0.056711,
+ 0.059086,
+ 0.057689,
+ 0.057689,
+ 0.057759,
+ 0.057689,
+ 0.057759,
+ 0.058248,
+ 0.057759,
+ 0.058178,
+ 0.058248,
+ 0.058248,
+ 0.058667,
+ 0.058178,
+ 0.057689,
+ 0.057759,
+ 0.0572
+ ],
+ "warm_rows_per_second": 8864419.397842761
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7521683752870862,
+ "character_perplexity": 5.76709435266236,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.000259297661525257,
+ "mean_prediction_kl_nats": 2.7767721139857577e-05,
+ "prediction_kl_p99_nats": 6.82620767472471e-05,
+ "maximum_prediction_kl_nats": 0.055205727356219876,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.6663031578063965,
+ "maximum_per_target_cross_entropy_increase": 0.8541910648345947
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.000259297661525257,
+ "mean_prediction_kl_nats": 2.7767721139857577e-05,
+ "prediction_kl_p99_nats": 6.82620767472471e-05,
+ "maximum_prediction_kl_nats": 0.055205727356219876,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.6663031578063965,
+ "maximum_per_target_cross_entropy_increase": 0.8541910648345947
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 128,
+ "storage": {
+ "artifact_bytes": 1505280,
+ "unique_rows": 11760,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "8f131fcc83bb844ad4e9d610693f55687d566532851a2724e31525d9369c4017",
+ "table_sha256": "ae77e374b3687298bc905011989dd5cea95a297efdc59cae484cf38c92f8e8d1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.019108804175630212,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1506085,
+ "backend_with_key_registry_bytes": 1667810,
+ "backend_with_datrie_key_index_bytes": 1601365,
+ "full_model_serving_artifact_bytes": 1962496,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999561702593,
+ "mean_cosine": 0.9999999793093656,
+ "p05_cosine": 0.9999999714403758,
+ "relative_l2_mean": 0.0002061393803210051,
+ "relative_l2_p95": 0.00024342303018898745,
+ "relative_l2_max": 0.0002964292448205526,
+ "weighted_relative_l2_mean": 0.00018699371965687282,
+ "weighted_relative_frobenius": 0.00020699072274290548,
+ "weighted_mean_cosine": 0.9999999813742835,
+ "max_relative_norm_change": 0.00017415355567106385
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.425823,
+ "warm_p50_ms": 0.132978,
+ "warm_p95_ms": 0.13661974999999996,
+ "warm_p99_ms": 0.16155541999999998,
+ "samples_ms": [
+ 0.134305,
+ 0.151416,
+ 0.134375,
+ 0.133816,
+ 0.133397,
+ 0.133397,
+ 0.132489,
+ 0.132978,
+ 0.132977,
+ 0.133816,
+ 0.133885,
+ 0.133886,
+ 0.132908,
+ 0.132,
+ 0.132978,
+ 0.132908,
+ 0.132489,
+ 0.133467,
+ 0.132978,
+ 0.133467,
+ 0.132489,
+ 0.132908,
+ 0.132908,
+ 0.131092,
+ 0.132489,
+ 0.132489,
+ 0.132419,
+ 0.132489,
+ 0.13249,
+ 0.132978,
+ 0.168038,
+ 0.135841,
+ 0.132419,
+ 0.133397,
+ 0.133397,
+ 0.132978,
+ 0.131511,
+ 0.133397,
+ 0.133397,
+ 0.132001
+ ],
+ "warm_rows_per_second": 3850260.9454195425
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.751908738569852,
+ "character_perplexity": 5.765597197583279,
+ "next_character_accuracy": 0.5006900782088637,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.7266746803859213e-07,
+ "mean_prediction_kl_nats": 3.967559244862379e-09,
+ "prediction_kl_p99_nats": 2.3306544418521744e-08,
+ "maximum_prediction_kl_nats": 7.241702038288769e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010852813720703125,
+ "maximum_per_target_cross_entropy_increase": 0.0007781982421875
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1491702,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 184,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 184,
+ "zlib_blocks": 184,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "8f131fcc83bb844ad4e9d610693f55687d566532851a2724e31525d9369c4017",
+ "table_sha256": "baf2739a65c9173b3a419ad965d58d8462a683778341881f6f71cad4c7463797",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14929922996088862,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1492909,
+ "backend_with_key_registry_bytes": 1654634,
+ "backend_with_datrie_key_index_bytes": 1588189,
+ "full_model_serving_artifact_bytes": 1949320,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999561702593,
+ "mean_cosine": 0.9999999793093656,
+ "p05_cosine": 0.9999999714403758,
+ "relative_l2_mean": 0.0002061393803210051,
+ "relative_l2_p95": 0.00024342303018898745,
+ "relative_l2_max": 0.0002964292448205526,
+ "weighted_relative_l2_mean": 0.00018699371965687282,
+ "weighted_relative_frobenius": 0.00020699072274290548,
+ "weighted_mean_cosine": 0.9999999813742835,
+ "max_relative_norm_change": 0.00017415355567106385
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.891363,
+ "warm_p50_ms": 0.141429,
+ "warm_p95_ms": 0.1458044,
+ "warm_p99_ms": 0.15016047,
+ "samples_ms": [
+ 0.145759,
+ 0.142406,
+ 0.143384,
+ 0.141498,
+ 0.14513,
+ 0.143873,
+ 0.141987,
+ 0.14101,
+ 0.140032,
+ 0.141917,
+ 0.141918,
+ 0.141499,
+ 0.141918,
+ 0.141428,
+ 0.141429,
+ 0.141918,
+ 0.140521,
+ 0.141009,
+ 0.14101,
+ 0.140032,
+ 0.140521,
+ 0.152394,
+ 0.142895,
+ 0.146667,
+ 0.144781,
+ 0.142337,
+ 0.142407,
+ 0.141429,
+ 0.14101,
+ 0.14101,
+ 0.14052,
+ 0.140451,
+ 0.141499,
+ 0.140521,
+ 0.14094,
+ 0.14101,
+ 0.140521,
+ 0.140521,
+ 0.140521,
+ 0.140521
+ ],
+ "warm_rows_per_second": 3620191.0499261115
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.751908738569852,
+ "character_perplexity": 5.765597197583279,
+ "next_character_accuracy": 0.5006900782088637,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.7266746803859213e-07,
+ "mean_prediction_kl_nats": 3.967559244862379e-09,
+ "prediction_kl_p99_nats": 2.3306544418521744e-08,
+ "maximum_prediction_kl_nats": 7.241702038288769e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010852813720703125,
+ "maximum_per_target_cross_entropy_increase": 0.0007781982421875
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1491702,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 184,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 184,
+ "zlib_blocks": 184,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "8f131fcc83bb844ad4e9d610693f55687d566532851a2724e31525d9369c4017",
+ "table_sha256": "baf2739a65c9173b3a419ad965d58d8462a683778341881f6f71cad4c7463797",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.18716672179289162,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1492908,
+ "backend_with_key_registry_bytes": 1654633,
+ "backend_with_datrie_key_index_bytes": 1588188,
+ "full_model_serving_artifact_bytes": 1949319,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999561702593,
+ "mean_cosine": 0.9999999793093656,
+ "p05_cosine": 0.9999999714403758,
+ "relative_l2_mean": 0.0002061393803210051,
+ "relative_l2_p95": 0.00024342303018898745,
+ "relative_l2_max": 0.0002964292448205526,
+ "weighted_relative_l2_mean": 0.00018699371965687282,
+ "weighted_relative_frobenius": 0.00020699072274290548,
+ "weighted_mean_cosine": 0.9999999813742835,
+ "max_relative_norm_change": 0.00017415355567106385
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.849737,
+ "warm_p50_ms": 0.14170850000000002,
+ "warm_p95_ms": 0.14629335,
+ "warm_p99_ms": 0.16134238,
+ "samples_ms": [
+ 0.144781,
+ 0.142337,
+ 0.141918,
+ 0.141009,
+ 0.141499,
+ 0.140521,
+ 0.147155,
+ 0.146248,
+ 0.142825,
+ 0.144711,
+ 0.142896,
+ 0.14052,
+ 0.141429,
+ 0.140521,
+ 0.140521,
+ 0.140032,
+ 0.141359,
+ 0.170413,
+ 0.14534,
+ 0.141499,
+ 0.1452,
+ 0.142896,
+ 0.143315,
+ 0.142896,
+ 0.14101,
+ 0.143314,
+ 0.142895,
+ 0.144711,
+ 0.144292,
+ 0.143873,
+ 0.141429,
+ 0.142476,
+ 0.141429,
+ 0.14101,
+ 0.140031,
+ 0.14101,
+ 0.141009,
+ 0.140521,
+ 0.14101,
+ 0.141498
+ ],
+ "warm_rows_per_second": 3613050.734430185
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.751908738569852,
+ "character_perplexity": 5.765597197583279,
+ "next_character_accuracy": 0.5006900782088637,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.7266746803859213e-07,
+ "mean_prediction_kl_nats": 3.967559244862379e-09,
+ "prediction_kl_p99_nats": 2.3306544418521744e-08,
+ "maximum_prediction_kl_nats": 7.241702038288769e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010852813720703125,
+ "maximum_per_target_cross_entropy_increase": 0.0007781982421875
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1491702,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 184,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 184,
+ "zlib_blocks": 184,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "8f131fcc83bb844ad4e9d610693f55687d566532851a2724e31525d9369c4017",
+ "table_sha256": "baf2739a65c9173b3a419ad965d58d8462a683778341881f6f71cad4c7463797",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1468518499750644,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 1492908,
+ "backend_with_key_registry_bytes": 1654633,
+ "backend_with_datrie_key_index_bytes": 1588188,
+ "full_model_serving_artifact_bytes": 1949319,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999561702593,
+ "mean_cosine": 0.9999999793093656,
+ "p05_cosine": 0.9999999714403758,
+ "relative_l2_mean": 0.0002061393803210051,
+ "relative_l2_p95": 0.00024342303018898745,
+ "relative_l2_max": 0.0002964292448205526,
+ "weighted_relative_l2_mean": 0.00018699371965687282,
+ "weighted_relative_frobenius": 0.00020699072274290548,
+ "weighted_mean_cosine": 0.9999999813742835,
+ "max_relative_norm_change": 0.00017415355567106385
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.891363,
+ "warm_p50_ms": 0.1416735,
+ "warm_p95_ms": 0.14574574999999998,
+ "warm_p99_ms": 0.17258308,
+ "samples_ms": [
+ 0.145201,
+ 0.142965,
+ 0.141918,
+ 0.141988,
+ 0.14094,
+ 0.141918,
+ 0.142407,
+ 0.141988,
+ 0.142337,
+ 0.141009,
+ 0.141428,
+ 0.14101,
+ 0.141848,
+ 0.141009,
+ 0.141009,
+ 0.140451,
+ 0.141428,
+ 0.142825,
+ 0.183124,
+ 0.156096,
+ 0.143874,
+ 0.142337,
+ 0.14101,
+ 0.141009,
+ 0.141987,
+ 0.142337,
+ 0.142406,
+ 0.141498,
+ 0.142476,
+ 0.139613,
+ 0.141498,
+ 0.141499,
+ 0.142337,
+ 0.14101,
+ 0.141009,
+ 0.140521,
+ 0.141429,
+ 0.141917,
+ 0.14101,
+ 0.14094
+ ],
+ "warm_rows_per_second": 3613943.3274394996
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.751908738569852,
+ "character_perplexity": 5.765597197583279,
+ "next_character_accuracy": 0.5006900782088637,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.7266746803859213e-07,
+ "mean_prediction_kl_nats": 3.967559244862379e-09,
+ "prediction_kl_p99_nats": 2.3306544418521744e-08,
+ "maximum_prediction_kl_nats": 7.241702038288769e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010852813720703125,
+ "maximum_per_target_cross_entropy_increase": 0.0007781982421875
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404516,
+ "unique_rows": 11022,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11022
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "785de7f6356459f0902c06da3b6caad2a0961b50ed6388b3ab2e23be05a0dc31",
+ "table_sha256": "845b2b891603f41d960335876f3379e6c21681d01ced708060391d0dd25d2523",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1694920409936458,
+ "representatives_sha256": "9013105313d6ddeb5e0f2c6c13db3aed609b655ac8c660cda775726425c1bfa6",
+ "structurally_merged_rows": 738,
+ "backend_artifact_bytes": 1405721,
+ "backend_with_key_registry_bytes": 1567446,
+ "backend_with_datrie_key_index_bytes": 1501001,
+ "full_model_serving_artifact_bytes": 1862132,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9936624335326107,
+ "mean_cosine": 0.9999574335860908,
+ "p05_cosine": 0.9996982293524213,
+ "relative_l2_mean": 0.002381315830960158,
+ "relative_l2_p95": 0.02482178038069265,
+ "relative_l2_max": 0.11263148326372621,
+ "weighted_relative_l2_mean": 0.0011114270993146658,
+ "weighted_relative_frobenius": 0.005608434419236457,
+ "weighted_mean_cosine": 0.9999842876418499,
+ "max_relative_norm_change": 0.010028110073227618
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9936621571143124,
+ "mean_cosine": 0.9999574540023862,
+ "p05_cosine": 0.9996979216243842,
+ "relative_l2_mean": 0.002188099155414633,
+ "relative_l2_p95": 0.024799198256564084,
+ "relative_l2_max": 0.11263622270436796,
+ "weighted_relative_l2_mean": 0.0009309170684210872,
+ "weighted_relative_frobenius": 0.005604080214656733,
+ "weighted_mean_cosine": 0.999984306340439,
+ "max_relative_norm_change": 0.009973852896569286
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.445636,
+ "warm_p50_ms": 0.1414285,
+ "warm_p95_ms": 0.14569584999999996,
+ "warm_p99_ms": 0.15571358000000002,
+ "samples_ms": [
+ 0.1452,
+ 0.142895,
+ 0.141499,
+ 0.141429,
+ 0.141917,
+ 0.142406,
+ 0.156095,
+ 0.141987,
+ 0.141498,
+ 0.142896,
+ 0.140521,
+ 0.142406,
+ 0.140521,
+ 0.140521,
+ 0.141917,
+ 0.140521,
+ 0.141429,
+ 0.140031,
+ 0.140521,
+ 0.141009,
+ 0.141499,
+ 0.14101,
+ 0.141498,
+ 0.140521,
+ 0.14101,
+ 0.141498,
+ 0.140521,
+ 0.14101,
+ 0.140521,
+ 0.139543,
+ 0.141428,
+ 0.140521,
+ 0.140521,
+ 0.155117,
+ 0.141987,
+ 0.142476,
+ 0.14101,
+ 0.140521,
+ 0.141498,
+ 0.140521
+ ],
+ "warm_rows_per_second": 3620203.848587802
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519901895170722,
+ "character_perplexity": 5.766066830062059,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 8.108156661531794e-05,
+ "mean_prediction_kl_nats": 2.7003989792803626e-06,
+ "prediction_kl_p99_nats": 4.4022399019958154e-05,
+ "maximum_prediction_kl_nats": 0.00427000501318861,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.24455657601356506,
+ "maximum_per_target_cross_entropy_increase": 0.12400329113006592
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 8.145426443871102e-05,
+ "mean_prediction_kl_nats": 2.6931912207071457e-06,
+ "prediction_kl_p99_nats": 4.41447738182781e-05,
+ "maximum_prediction_kl_nats": 0.004265104875593458,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.24464166164398193,
+ "maximum_per_target_cross_entropy_increase": 0.12392067909240723
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404353,
+ "unique_rows": 11020,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11020
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "78571ac5fc12853614a0d899073f39dd8547b1d0f844cd4be7b36d1301b33539",
+ "table_sha256": "75d931f1d53e3d9797b296d3e9d38beb1623dac5280fbf3c463e20b508167d54",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14509666594676673,
+ "representatives_sha256": "207d7ee8c9849014e5ecc344843ce94ee019d669b1107e94eb340448afd0a919",
+ "structurally_merged_rows": 740,
+ "backend_artifact_bytes": 1405558,
+ "backend_with_key_registry_bytes": 1567283,
+ "backend_with_datrie_key_index_bytes": 1500838,
+ "full_model_serving_artifact_bytes": 1861969,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9836722642304618,
+ "mean_cosine": 0.9999550670186391,
+ "p05_cosine": 0.9996975312475809,
+ "relative_l2_mean": 0.002409581727041581,
+ "relative_l2_p95": 0.024849233211097135,
+ "relative_l2_max": 0.18058696496909937,
+ "weighted_relative_l2_mean": 0.001130755829370565,
+ "weighted_relative_frobenius": 0.005965541474413652,
+ "weighted_mean_cosine": 0.9999828249229119,
+ "max_relative_norm_change": 0.010028110073227618
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.983662787110448,
+ "mean_cosine": 0.9999550867509435,
+ "p05_cosine": 0.9996974624952747,
+ "relative_l2_mean": 0.002216398924696341,
+ "relative_l2_p95": 0.024838888229141796,
+ "relative_l2_max": 0.1806373240056417,
+ "weighted_relative_l2_mean": 0.0009502665207172197,
+ "weighted_relative_frobenius": 0.005961406671631451,
+ "weighted_mean_cosine": 0.9999828438059056,
+ "max_relative_norm_change": 0.009973852896569286
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.424823,
+ "warm_p50_ms": 0.141429,
+ "warm_p95_ms": 0.14564695,
+ "warm_p99_ms": 0.16381553999999998,
+ "samples_ms": [
+ 0.143873,
+ 0.143314,
+ 0.142407,
+ 0.14052,
+ 0.141429,
+ 0.141918,
+ 0.142476,
+ 0.140521,
+ 0.14101,
+ 0.141009,
+ 0.14101,
+ 0.14101,
+ 0.14045,
+ 0.14101,
+ 0.175092,
+ 0.142407,
+ 0.141429,
+ 0.141498,
+ 0.140591,
+ 0.141499,
+ 0.140032,
+ 0.140521,
+ 0.139612,
+ 0.140521,
+ 0.141428,
+ 0.141429,
+ 0.141499,
+ 0.14101,
+ 0.140521,
+ 0.140032,
+ 0.141987,
+ 0.139613,
+ 0.146178,
+ 0.143454,
+ 0.143384,
+ 0.141918,
+ 0.141429,
+ 0.141009,
+ 0.144292,
+ 0.145619
+ ],
+ "warm_rows_per_second": 3620191.0499261115
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519812967441397,
+ "character_perplexity": 5.766015553967019,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 7.218811218665827e-05,
+ "mean_prediction_kl_nats": 2.8673964566398535e-06,
+ "prediction_kl_p99_nats": 4.422877980386387e-05,
+ "maximum_prediction_kl_nats": 0.00427000501318861,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.2781916856765747,
+ "maximum_per_target_cross_entropy_increase": 0.12400329113006592
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 7.256080669574896e-05,
+ "mean_prediction_kl_nats": 2.8610946976357848e-06,
+ "prediction_kl_p99_nats": 4.421483256831372e-05,
+ "maximum_prediction_kl_nats": 0.004265104875593458,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.27823901176452637,
+ "maximum_per_target_cross_entropy_increase": 0.12392067909240723
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1404013,
+ "unique_rows": 11017,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11017
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "4e4e27a7f10c80adedc7b35304bb6d89e97dc17fb0c58379a0d24e41afa70e20",
+ "table_sha256": "444efd7255148d02ece8f2f798bbeec15f92a73f324134f9c1b408efb75bf235",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.15004485612735152,
+ "representatives_sha256": "4b7ad2db943cbe62fafde32d2beb4eb1654c97fc6c4a11076be545993968b9ab",
+ "structurally_merged_rows": 743,
+ "backend_artifact_bytes": 1405218,
+ "backend_with_key_registry_bytes": 1566943,
+ "backend_with_datrie_key_index_bytes": 1500498,
+ "full_model_serving_artifact_bytes": 1861629,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9715812247051491,
+ "mean_cosine": 0.9999464193252622,
+ "p05_cosine": 0.9996918807512631,
+ "relative_l2_mean": 0.0024844494170948132,
+ "relative_l2_p95": 0.025000227633378485,
+ "relative_l2_max": 0.23931966315367953,
+ "weighted_relative_l2_mean": 0.0012089073703037338,
+ "weighted_relative_frobenius": 0.008071870537997206,
+ "weighted_mean_cosine": 0.9999737605864243,
+ "max_relative_norm_change": 0.010028110073227618
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9715763313795763,
+ "mean_cosine": 0.9999464378996576,
+ "p05_cosine": 0.9996921393279198,
+ "relative_l2_mean": 0.0022913286448834774,
+ "relative_l2_p95": 0.024953767743632015,
+ "relative_l2_max": 0.23934554468610783,
+ "weighted_relative_l2_mean": 0.0010284817488151697,
+ "weighted_relative_frobenius": 0.008069216677422612,
+ "weighted_mean_cosine": 0.9999737775473834,
+ "max_relative_norm_change": 0.009973852896569286
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.447102,
+ "warm_p50_ms": 0.1431045,
+ "warm_p95_ms": 0.15956325,
+ "warm_p99_ms": 0.17014537999999998,
+ "samples_ms": [
+ 0.149041,
+ 0.149041,
+ 0.145689,
+ 0.141988,
+ 0.142406,
+ 0.143314,
+ 0.142406,
+ 0.14101,
+ 0.144292,
+ 0.147156,
+ 0.150509,
+ 0.145689,
+ 0.159448,
+ 0.154699,
+ 0.142476,
+ 0.142895,
+ 0.155187,
+ 0.161753,
+ 0.148063,
+ 0.147156,
+ 0.142407,
+ 0.140521,
+ 0.175511,
+ 0.141498,
+ 0.14527,
+ 0.143803,
+ 0.143873,
+ 0.14052,
+ 0.142407,
+ 0.14094,
+ 0.141498,
+ 0.14094,
+ 0.14094,
+ 0.144711,
+ 0.140031,
+ 0.141499,
+ 0.14101,
+ 0.141498,
+ 0.141429,
+ 0.144781
+ ],
+ "warm_rows_per_second": 3577805.0305895344
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519985300844854,
+ "character_perplexity": 5.766114922531722,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 8.942284508012468e-05,
+ "mean_prediction_kl_nats": 4.732999292025376e-06,
+ "prediction_kl_p99_nats": 5.5180821558474446e-05,
+ "maximum_prediction_kl_nats": 0.00427000501318861,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.5952320098876953,
+ "maximum_per_target_cross_entropy_increase": 0.2905094623565674
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 8.979554601204205e-05,
+ "mean_prediction_kl_nats": 4.72634571777004e-06,
+ "prediction_kl_p99_nats": 5.572812791949773e-05,
+ "maximum_prediction_kl_nats": 0.004265104875593458,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.595128059387207,
+ "maximum_per_target_cross_entropy_increase": 0.29048895835876465
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1403486,
+ "unique_rows": 11013,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 173,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 172,
+ "zlib_blocks": 172,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11013
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "32d3efbc900eb2931365249475898406aaba52ee029fc94c02a9afefea84f5dd",
+ "table_sha256": "0992c9d6e5c63fc06ebd1ebbc7f37b595f9d7be3775c4e2512a0d4eed5b092fa",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.14616956911049783,
+ "representatives_sha256": "072a861c31f89876bfc32521508d31c3af9abc29633347f0f2d4dffa8be1d0be",
+ "structurally_merged_rows": 747,
+ "backend_artifact_bytes": 1404691,
+ "backend_with_key_registry_bytes": 1566416,
+ "backend_with_datrie_key_index_bytes": 1499971,
+ "full_model_serving_artifact_bytes": 1861102,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9540935690445206,
+ "mean_cosine": 0.9999332679413712,
+ "p05_cosine": 0.9996868885640952,
+ "relative_l2_mean": 0.0025788275188094374,
+ "relative_l2_p95": 0.025159322591634673,
+ "relative_l2_max": 0.30429112907861555,
+ "weighted_relative_l2_mean": 0.001257991412726988,
+ "weighted_relative_frobenius": 0.009467519517306568,
+ "weighted_mean_cosine": 0.999966966968916,
+ "max_relative_norm_change": 0.010028110073227618
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.954088589280544,
+ "mean_cosine": 0.9999332864986421,
+ "p05_cosine": 0.9996874062379404,
+ "relative_l2_mean": 0.0023857788662846047,
+ "relative_l2_p95": 0.025154377332231567,
+ "relative_l2_max": 0.3043068509506128,
+ "weighted_relative_l2_mean": 0.001077604318228787,
+ "weighted_relative_frobenius": 0.009465423815370745,
+ "weighted_mean_cosine": 0.9999669839465212,
+ "max_relative_norm_change": 0.009973852896569286
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.464632,
+ "warm_p50_ms": 0.140521,
+ "warm_p95_ms": 0.14480895,
+ "warm_p99_ms": 0.15369029,
+ "samples_ms": [
+ 0.14534,
+ 0.142895,
+ 0.14094,
+ 0.14101,
+ 0.144781,
+ 0.144362,
+ 0.142406,
+ 0.142477,
+ 0.140521,
+ 0.141498,
+ 0.141429,
+ 0.139543,
+ 0.140521,
+ 0.140521,
+ 0.140032,
+ 0.140101,
+ 0.140451,
+ 0.159029,
+ 0.141918,
+ 0.141428,
+ 0.140032,
+ 0.140102,
+ 0.140032,
+ 0.14094,
+ 0.14052,
+ 0.140032,
+ 0.14052,
+ 0.139613,
+ 0.140521,
+ 0.143873,
+ 0.140032,
+ 0.139543,
+ 0.140032,
+ 0.140521,
+ 0.14101,
+ 0.141429,
+ 0.141918,
+ 0.140032,
+ 0.139054,
+ 0.139054
+ ],
+ "warm_rows_per_second": 3643583.5213242145
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.752106522501061,
+ "character_perplexity": 5.76673765284093,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00019743075055729176,
+ "mean_prediction_kl_nats": 1.0169879969428095e-05,
+ "prediction_kl_p99_nats": 6.048439962904012e-05,
+ "maximum_prediction_kl_nats": 0.013871782352292441,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.6127772331237793,
+ "maximum_per_target_cross_entropy_increase": 0.85483717918396
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00019780349174025678,
+ "mean_prediction_kl_nats": 1.0161199275241385e-05,
+ "prediction_kl_p99_nats": 6.041215210047706e-05,
+ "maximum_prediction_kl_nats": 0.013866209796825128,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.6127262115478516,
+ "maximum_per_target_cross_entropy_increase": 0.8547048568725586
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 1505280,
+ "artifact_bytes": 1402812,
+ "unique_rows": 11008,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 172,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 172,
+ "zlib_blocks": 172,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11008
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "e15d5dcbbde7631cd554c932b957f8fdd83481b33396efe09409dc8f86e6c974",
+ "table_sha256": "71403bc91623cece5472d7c4efdb6f4ac9c14aad9d857e2c570c554c6e431afb",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.1447559108491987,
+ "representatives_sha256": "4445613e0eb1b8c00a2687569e9f020c68f3bb469b5c5750222604b97661a4b7",
+ "structurally_merged_rows": 752,
+ "backend_artifact_bytes": 1404017,
+ "backend_with_key_registry_bytes": 1565742,
+ "backend_with_datrie_key_index_bytes": 1499297,
+ "full_model_serving_artifact_bytes": 1860428,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9104368676447112,
+ "mean_cosine": 0.9998964605058549,
+ "p05_cosine": 0.9996822992536398,
+ "relative_l2_mean": 0.002769957707369786,
+ "relative_l2_p95": 0.025353362987667258,
+ "relative_l2_max": 0.42271188496719964,
+ "weighted_relative_l2_mean": 0.001349395227162131,
+ "weighted_relative_frobenius": 0.010513068014911243,
+ "weighted_mean_cosine": 0.9999485675167205,
+ "max_relative_norm_change": 0.010028110073227618
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9104365327066158,
+ "mean_cosine": 0.9998964807650583,
+ "p05_cosine": 0.9996831448565078,
+ "relative_l2_mean": 0.0025769816210947107,
+ "relative_l2_p95": 0.0253563989285597,
+ "relative_l2_max": 0.42269607197838727,
+ "weighted_relative_l2_mean": 0.0011690501394823313,
+ "weighted_relative_frobenius": 0.010511287974359379,
+ "weighted_mean_cosine": 0.9999485838283916,
+ "max_relative_norm_change": 0.009973852896569286
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.5006,
+ "warm_p50_ms": 0.14146350000000002,
+ "warm_p95_ms": 0.14599649999999997,
+ "warm_p99_ms": 0.16652199999999998,
+ "samples_ms": [
+ 0.17167,
+ 0.14534,
+ 0.143314,
+ 0.142826,
+ 0.143384,
+ 0.143803,
+ 0.142406,
+ 0.141499,
+ 0.141918,
+ 0.141917,
+ 0.141009,
+ 0.141918,
+ 0.140521,
+ 0.141847,
+ 0.14101,
+ 0.14101,
+ 0.140521,
+ 0.14094,
+ 0.141428,
+ 0.139124,
+ 0.14094,
+ 0.139543,
+ 0.14101,
+ 0.141498,
+ 0.141429,
+ 0.14101,
+ 0.141568,
+ 0.15847,
+ 0.142895,
+ 0.140101,
+ 0.141429,
+ 0.141429,
+ 0.141009,
+ 0.141498,
+ 0.140521,
+ 0.140939,
+ 0.141499,
+ 0.141499,
+ 0.142337,
+ 0.140032
+ ],
+ "warm_rows_per_second": 3619308.160762316
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7521656517845516,
+ "character_perplexity": 5.767078645987661,
+ "next_character_accuracy": 0.5007667535654041,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0002565734565024768,
+ "mean_prediction_kl_nats": 2.7694690963231373e-05,
+ "prediction_kl_p99_nats": 6.599084822732549e-05,
+ "maximum_prediction_kl_nats": 0.05519350090312892,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.6659965515136719,
+ "maximum_per_target_cross_entropy_increase": 0.85483717918396
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00025694621972601253,
+ "mean_prediction_kl_nats": 2.7696791556825498e-05,
+ "prediction_kl_p99_nats": 6.66065197807943e-05,
+ "maximum_prediction_kl_nats": 0.055230586414404835,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.666330099105835,
+ "maximum_per_target_cross_entropy_increase": 0.8547048568725586
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 68,
+ "storage": {
+ "artifact_bytes": 799680,
+ "unique_rows": 11760,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "c24044479e6740e64c7477e0129737cf3fc9833330c87891041560fc8979c289",
+ "table_sha256": "33f22b0adf51608d72dafdd55c404069b6392555fde3ce9cbe2da3193605ee03",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.015448767226189375,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 800527,
+ "backend_with_key_registry_bytes": 962252,
+ "backend_with_datrie_key_index_bytes": 895807,
+ "full_model_serving_artifact_bytes": 1256938,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999476717268412,
+ "mean_cosine": 0.9999868747051318,
+ "p05_cosine": 0.9999750305769112,
+ "relative_l2_mean": 0.005015239779976406,
+ "relative_l2_p95": 0.007111025857858184,
+ "relative_l2_max": 0.010230223191756224,
+ "weighted_relative_l2_mean": 0.005115351005261811,
+ "weighted_relative_frobenius": 0.005551393829243131,
+ "weighted_mean_cosine": 0.9999853386526664,
+ "max_relative_norm_change": 0.0028514043963567587
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.353886,
+ "warm_p50_ms": 0.098895,
+ "warm_p95_ms": 0.10071545,
+ "warm_p99_ms": 0.11117661999999999,
+ "samples_ms": [
+ 0.100781,
+ 0.099873,
+ 0.098337,
+ 0.098895,
+ 0.098826,
+ 0.100292,
+ 0.098406,
+ 0.098895,
+ 0.097917,
+ 0.098406,
+ 0.099314,
+ 0.099314,
+ 0.099454,
+ 0.099315,
+ 0.117823,
+ 0.100712,
+ 0.100292,
+ 0.098476,
+ 0.09701,
+ 0.097499,
+ 0.098476,
+ 0.097009,
+ 0.096451,
+ 0.098896,
+ 0.098966,
+ 0.099315,
+ 0.09701,
+ 0.098407,
+ 0.098895,
+ 0.098406,
+ 0.097498,
+ 0.098895,
+ 0.098895,
+ 0.099315,
+ 0.098477,
+ 0.099315,
+ 0.098476,
+ 0.09701,
+ 0.098896,
+ 0.099315
+ ],
+ "warm_rows_per_second": 5177208.150058143
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.751896889463768,
+ "character_perplexity": 5.765528880815196,
+ "next_character_accuracy": 0.5011501303481061,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.2221698935909417e-05,
+ "mean_prediction_kl_nats": 3.718427634436085e-06,
+ "prediction_kl_p99_nats": 2.5791119079547005e-05,
+ "maximum_prediction_kl_nats": 0.00010015670638372724,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.00046005213924238397,
+ "maximum_absolute_logit_change": 0.03807401657104492,
+ "maximum_per_target_cross_entropy_increase": 0.018783092498779297
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 893276,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 98,
+ "max_block_bytes": 8192,
+ "raw_blocks": 11,
+ "compressed_blocks": 87,
+ "zlib_blocks": 87,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "c24044479e6740e64c7477e0129737cf3fc9833330c87891041560fc8979c289",
+ "table_sha256": "afcbb7c49277c72201ff209b1943a53df7ed001404730910419901465c8bb252",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08559812000021338,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 894522,
+ "backend_with_key_registry_bytes": 1056247,
+ "backend_with_datrie_key_index_bytes": 989802,
+ "full_model_serving_artifact_bytes": 1350933,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999476717268412,
+ "mean_cosine": 0.9999868747051318,
+ "p05_cosine": 0.9999750305769112,
+ "relative_l2_mean": 0.005015239779976406,
+ "relative_l2_p95": 0.007111025857858184,
+ "relative_l2_max": 0.010230223191756224,
+ "weighted_relative_l2_mean": 0.005115351005261811,
+ "weighted_relative_frobenius": 0.005551393829243131,
+ "weighted_mean_cosine": 0.9999853386526664,
+ "max_relative_norm_change": 0.0028514043963567587
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.958247,
+ "warm_p50_ms": 0.11028,
+ "warm_p95_ms": 0.13255535,
+ "warm_p99_ms": 0.15865397999999997,
+ "samples_ms": [
+ 0.11447,
+ 0.174534,
+ 0.132489,
+ 0.112165,
+ 0.111188,
+ 0.111607,
+ 0.110699,
+ 0.110279,
+ 0.10979,
+ 0.111188,
+ 0.110838,
+ 0.108813,
+ 0.108882,
+ 0.11028,
+ 0.109232,
+ 0.109301,
+ 0.109721,
+ 0.110769,
+ 0.111188,
+ 0.107905,
+ 0.11021,
+ 0.11021,
+ 0.109791,
+ 0.10972,
+ 0.110699,
+ 0.110699,
+ 0.109791,
+ 0.110628,
+ 0.109371,
+ 0.11021,
+ 0.109721,
+ 0.109301,
+ 0.109721,
+ 0.110698,
+ 0.11028,
+ 0.11028,
+ 0.133816,
+ 0.112165,
+ 0.109721,
+ 0.110699
+ ],
+ "warm_rows_per_second": 4642727.602466449
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.751896889463768,
+ "character_perplexity": 5.765528880815196,
+ "next_character_accuracy": 0.5011501303481061,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.2221698935909417e-05,
+ "mean_prediction_kl_nats": 3.718427634436085e-06,
+ "prediction_kl_p99_nats": 2.5791119079547005e-05,
+ "maximum_prediction_kl_nats": 0.00010015670638372724,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.00046005213924238397,
+ "maximum_absolute_logit_change": 0.03807401657104492,
+ "maximum_per_target_cross_entropy_increase": 0.018783092498779297
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 893276,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 98,
+ "max_block_bytes": 8192,
+ "raw_blocks": 11,
+ "compressed_blocks": 87,
+ "zlib_blocks": 87,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "c24044479e6740e64c7477e0129737cf3fc9833330c87891041560fc8979c289",
+ "table_sha256": "afcbb7c49277c72201ff209b1943a53df7ed001404730910419901465c8bb252",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.15985735692083836,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 894521,
+ "backend_with_key_registry_bytes": 1056246,
+ "backend_with_datrie_key_index_bytes": 989801,
+ "full_model_serving_artifact_bytes": 1350932,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999476717268412,
+ "mean_cosine": 0.9999868747051318,
+ "p05_cosine": 0.9999750305769112,
+ "relative_l2_mean": 0.005015239779976406,
+ "relative_l2_p95": 0.007111025857858184,
+ "relative_l2_max": 0.010230223191756224,
+ "weighted_relative_l2_mean": 0.005115351005261811,
+ "weighted_relative_frobenius": 0.005551393829243131,
+ "weighted_mean_cosine": 0.9999853386526664,
+ "max_relative_norm_change": 0.0028514043963567587
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.958736,
+ "warm_p50_ms": 0.11024500000000001,
+ "warm_p95_ms": 0.11549685,
+ "warm_p99_ms": 0.13627805999999998,
+ "samples_ms": [
+ 0.115448,
+ 0.111188,
+ 0.111606,
+ 0.112165,
+ 0.11021,
+ 0.110209,
+ 0.111606,
+ 0.111676,
+ 0.110768,
+ 0.111187,
+ 0.148971,
+ 0.111188,
+ 0.110768,
+ 0.110768,
+ 0.111606,
+ 0.110698,
+ 0.116425,
+ 0.11021,
+ 0.11028,
+ 0.110699,
+ 0.109302,
+ 0.109301,
+ 0.109302,
+ 0.10979,
+ 0.108813,
+ 0.11014,
+ 0.111188,
+ 0.108813,
+ 0.109232,
+ 0.109302,
+ 0.108254,
+ 0.109302,
+ 0.109651,
+ 0.108883,
+ 0.109302,
+ 0.111188,
+ 0.109302,
+ 0.110698,
+ 0.109301,
+ 0.11021
+ ],
+ "warm_rows_per_second": 4644201.551090752
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.751896889463768,
+ "character_perplexity": 5.765528880815196,
+ "next_character_accuracy": 0.5011501303481061,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.2221698935909417e-05,
+ "mean_prediction_kl_nats": 3.718427634436085e-06,
+ "prediction_kl_p99_nats": 2.5791119079547005e-05,
+ "maximum_prediction_kl_nats": 0.00010015670638372724,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.00046005213924238397,
+ "maximum_absolute_logit_change": 0.03807401657104492,
+ "maximum_per_target_cross_entropy_increase": 0.018783092498779297
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 893276,
+ "unique_rows": 11760,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 98,
+ "max_block_bytes": 8192,
+ "raw_blocks": 11,
+ "compressed_blocks": 87,
+ "zlib_blocks": 87,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 11760
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "c24044479e6740e64c7477e0129737cf3fc9833330c87891041560fc8979c289",
+ "table_sha256": "afcbb7c49277c72201ff209b1943a53df7ed001404730910419901465c8bb252",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08475534385070205,
+ "representatives_sha256": "1f725dc17215df73ba41553c18f5027884a4b6b9c5f2e739cad9878427946916",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 894521,
+ "backend_with_key_registry_bytes": 1056246,
+ "backend_with_datrie_key_index_bytes": 989801,
+ "full_model_serving_artifact_bytes": 1350932,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999476717268412,
+ "mean_cosine": 0.9999868747051318,
+ "p05_cosine": 0.9999750305769112,
+ "relative_l2_mean": 0.005015239779976406,
+ "relative_l2_p95": 0.007111025857858184,
+ "relative_l2_max": 0.010230223191756224,
+ "weighted_relative_l2_mean": 0.005115351005261811,
+ "weighted_relative_frobenius": 0.005551393829243131,
+ "weighted_mean_cosine": 0.9999853386526664,
+ "max_relative_norm_change": 0.0028514043963567587
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.976685,
+ "warm_p50_ms": 0.109302,
+ "warm_p95_ms": 0.11582514999999997,
+ "warm_p99_ms": 0.13777435,
+ "samples_ms": [
+ 0.115448,
+ 0.112095,
+ 0.10986,
+ 0.110768,
+ 0.147226,
+ 0.112584,
+ 0.111257,
+ 0.109721,
+ 0.108953,
+ 0.109721,
+ 0.111117,
+ 0.110698,
+ 0.110349,
+ 0.109302,
+ 0.108812,
+ 0.110698,
+ 0.111118,
+ 0.108813,
+ 0.109301,
+ 0.108394,
+ 0.108813,
+ 0.108813,
+ 0.108813,
+ 0.109301,
+ 0.109302,
+ 0.10986,
+ 0.109791,
+ 0.108813,
+ 0.108812,
+ 0.108324,
+ 0.109301,
+ 0.108324,
+ 0.109231,
+ 0.107835,
+ 0.109302,
+ 0.10986,
+ 0.109301,
+ 0.108813,
+ 0.107904,
+ 0.122991
+ ],
+ "warm_rows_per_second": 4684269.27229145
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.751896889463768,
+ "character_perplexity": 5.765528880815196,
+ "next_character_accuracy": 0.5011501303481061,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.2221698935909417e-05,
+ "mean_prediction_kl_nats": 3.718427634436085e-06,
+ "prediction_kl_p99_nats": 2.5791119079547005e-05,
+ "maximum_prediction_kl_nats": 0.00010015670638372724,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.00046005213924238397,
+ "maximum_absolute_logit_change": 0.03807401657104492,
+ "maximum_per_target_cross_entropy_increase": 0.018783092498779297
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 843031,
+ "unique_rows": 11027,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 9,
+ "compressed_blocks": 83,
+ "zlib_blocks": 83,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11027
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "6a76caf808febaa2918212471310dcc13a7e77c361da6fd749ffce6f8828f9c1",
+ "table_sha256": "8299fefb1df30f6d5c27208240446621e707cf8d79f5a9facd6586036618133c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08858914603479207,
+ "representatives_sha256": "ddc211604533e762af72c94b24f16feddf8ad1df483ca51a831f088fa4305c99",
+ "structurally_merged_rows": 733,
+ "backend_artifact_bytes": 844274,
+ "backend_with_key_registry_bytes": 1005999,
+ "backend_with_datrie_key_index_bytes": 939554,
+ "full_model_serving_artifact_bytes": 1300685,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9935234690476181,
+ "mean_cosine": 0.9999448499062227,
+ "p05_cosine": 0.9996907266797578,
+ "relative_l2_mean": 0.0069153334884845775,
+ "relative_l2_p95": 0.02504314859298606,
+ "relative_l2_max": 0.11380766731945188,
+ "weighted_relative_l2_mean": 0.00590098430130065,
+ "weighted_relative_frobenius": 0.007915395391086759,
+ "weighted_mean_cosine": 0.9999694724828936,
+ "max_relative_norm_change": 0.01026371520134981
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9935668397229672,
+ "mean_cosine": 0.9999565796102982,
+ "p05_cosine": 0.9996795632367155,
+ "relative_l2_mean": 0.002210186691108473,
+ "relative_l2_p95": 0.025412558685815156,
+ "relative_l2_max": 0.11350018007565987,
+ "weighted_relative_l2_mean": 0.0009643804186688708,
+ "weighted_relative_frobenius": 0.005829303558166024,
+ "weighted_mean_cosine": 0.9999832783732631,
+ "max_relative_norm_change": 0.01000756191502061
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.904329,
+ "warm_p50_ms": 0.109302,
+ "warm_p95_ms": 0.11434419999999995,
+ "warm_p99_ms": 0.13555426999999998,
+ "samples_ms": [
+ 0.113562,
+ 0.111257,
+ 0.129206,
+ 0.110698,
+ 0.109721,
+ 0.11028,
+ 0.111188,
+ 0.110629,
+ 0.110209,
+ 0.109301,
+ 0.109232,
+ 0.109372,
+ 0.109232,
+ 0.109861,
+ 0.108813,
+ 0.108882,
+ 0.109302,
+ 0.109231,
+ 0.109791,
+ 0.108813,
+ 0.109302,
+ 0.109721,
+ 0.108813,
+ 0.107346,
+ 0.108393,
+ 0.109302,
+ 0.109721,
+ 0.108812,
+ 0.109302,
+ 0.109232,
+ 0.109302,
+ 0.109232,
+ 0.109301,
+ 0.108812,
+ 0.109302,
+ 0.10979,
+ 0.109231,
+ 0.139613,
+ 0.112095,
+ 0.109371
+ ],
+ "warm_rows_per_second": 4684269.27229145
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519973031897105,
+ "character_perplexity": 5.766107848119792,
+ "next_character_accuracy": 0.5010734549915657,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 8.819584134555916e-05,
+ "mean_prediction_kl_nats": 6.619619213553747e-06,
+ "prediction_kl_p99_nats": 5.8379613657389136e-05,
+ "maximum_prediction_kl_nats": 0.004350244493226012,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.24921925365924835,
+ "maximum_per_target_cross_entropy_increase": 0.1248863935470581
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00010041876756941332,
+ "mean_prediction_kl_nats": 2.8241785271275015e-06,
+ "prediction_kl_p99_nats": 4.790040985163771e-05,
+ "maximum_prediction_kl_nats": 0.0042292557678851525,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": -7.667535654032331e-05,
+ "maximum_absolute_logit_change": 0.2476276308298111,
+ "maximum_per_target_cross_entropy_increase": 0.12327206134796143
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842884,
+ "unique_rows": 11025,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 9,
+ "compressed_blocks": 83,
+ "zlib_blocks": 83,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11025
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "9cb793383faafe69621c29fbedcf09935e833a5ca1b6577e6e07b5d37f681ffc",
+ "table_sha256": "8a1a378e6c7409e01657aaf8952035ab62ab1d557ce09457130e1fe5f8fa3697",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.09192413999699056,
+ "representatives_sha256": "4e96ec63a944015daeb56d59750a418a1355cb8ec5e00be7bb61faf9c69dbac5",
+ "structurally_merged_rows": 735,
+ "backend_artifact_bytes": 844127,
+ "backend_with_key_registry_bytes": 1005852,
+ "backend_with_datrie_key_index_bytes": 939407,
+ "full_model_serving_artifact_bytes": 1300538,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9835967725835124,
+ "mean_cosine": 0.9999424895823382,
+ "p05_cosine": 0.9996906107736135,
+ "relative_l2_mean": 0.006942767588978875,
+ "relative_l2_p95": 0.025062895425252696,
+ "relative_l2_max": 0.18093346744696945,
+ "weighted_relative_l2_mean": 0.00591965234573629,
+ "weighted_relative_frobenius": 0.008169246504612586,
+ "weighted_mean_cosine": 0.9999680270419147,
+ "max_relative_norm_change": 0.01026371520134981
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9835349675861836,
+ "mean_cosine": 0.9999542193719191,
+ "p05_cosine": 0.9996786278128243,
+ "relative_l2_mean": 0.0022384190422566168,
+ "relative_l2_p95": 0.02552355952616013,
+ "relative_l2_max": 0.18129729136883532,
+ "weighted_relative_l2_mean": 0.0009835451344865425,
+ "weighted_relative_frobenius": 0.006167469880755982,
+ "weighted_mean_cosine": 0.9999818425792947,
+ "max_relative_norm_change": 0.01000756191502061
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.951193,
+ "warm_p50_ms": 0.110209,
+ "warm_p95_ms": 0.11635505,
+ "warm_p99_ms": 0.13429182999999997,
+ "samples_ms": [
+ 0.116356,
+ 0.11447,
+ 0.110628,
+ 0.112165,
+ 0.111187,
+ 0.145759,
+ 0.113562,
+ 0.116355,
+ 0.111117,
+ 0.110209,
+ 0.108255,
+ 0.109791,
+ 0.109721,
+ 0.10979,
+ 0.110209,
+ 0.109721,
+ 0.11028,
+ 0.11028,
+ 0.11028,
+ 0.109791,
+ 0.109791,
+ 0.10972,
+ 0.107835,
+ 0.109372,
+ 0.113562,
+ 0.10986,
+ 0.109791,
+ 0.112585,
+ 0.109302,
+ 0.107835,
+ 0.108883,
+ 0.114959,
+ 0.107486,
+ 0.115377,
+ 0.109232,
+ 0.109791,
+ 0.111118,
+ 0.11028,
+ 0.109721,
+ 0.110698
+ ],
+ "warm_rows_per_second": 4645718.5892259255
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519882733825372,
+ "character_perplexity": 5.7660557815128595,
+ "next_character_accuracy": 0.5010734549915657,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 7.916527855288511e-05,
+ "mean_prediction_kl_nats": 6.8057714999284804e-06,
+ "prediction_kl_p99_nats": 5.931262572062106e-05,
+ "maximum_prediction_kl_nats": 0.004350244493226012,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.28000450134277344,
+ "maximum_per_target_cross_entropy_increase": 0.1248863935470581
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 9.138809440657068e-05,
+ "mean_prediction_kl_nats": 2.987786397601086e-06,
+ "prediction_kl_p99_nats": 4.9591083932686315e-05,
+ "maximum_prediction_kl_nats": 0.0042292557678851525,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": -7.667535654032331e-05,
+ "maximum_absolute_logit_change": 0.27948784828186035,
+ "maximum_per_target_cross_entropy_increase": 0.12327206134796143
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842733,
+ "unique_rows": 11022,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 9,
+ "compressed_blocks": 83,
+ "zlib_blocks": 83,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11022
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "afdcfa804544bb5565cfcf55b990cabf7227427470bf7cc4d63800b7ed0fc753",
+ "table_sha256": "c62a96eadafd51f8dcb0c91248fa0e36f659ea93e1b913ca467a1bb449932ead",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.08975822012871504,
+ "representatives_sha256": "7b6b25be7c4910dfee752a11a070d4486390bdf7141b916423109360b2624506",
+ "structurally_merged_rows": 738,
+ "backend_artifact_bytes": 843976,
+ "backend_with_key_registry_bytes": 1005701,
+ "backend_with_datrie_key_index_bytes": 939256,
+ "full_model_serving_artifact_bytes": 1300387,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9718039636632881,
+ "mean_cosine": 0.9999338860458784,
+ "p05_cosine": 0.9996873580129493,
+ "relative_l2_mean": 0.007016172763185574,
+ "relative_l2_p95": 0.02536507624210821,
+ "relative_l2_max": 0.2384614203989659,
+ "weighted_relative_l2_mean": 0.005996171242510425,
+ "weighted_relative_frobenius": 0.009804938310251428,
+ "weighted_mean_cosine": 0.9999590152236963,
+ "max_relative_norm_change": 0.01026371520134981
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9718008085430542,
+ "mean_cosine": 0.9999455606719624,
+ "p05_cosine": 0.999675080030134,
+ "relative_l2_mean": 0.0023132668469809654,
+ "relative_l2_p95": 0.025674210231420577,
+ "relative_l2_max": 0.23839006453365424,
+ "weighted_relative_l2_mean": 0.0010616306996878465,
+ "weighted_relative_frobenius": 0.00822429441213463,
+ "weighted_mean_cosine": 0.9999727704685598,
+ "max_relative_norm_change": 0.01000756191502061
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.884914,
+ "warm_p50_ms": 0.110768,
+ "warm_p95_ms": 0.114959,
+ "warm_p99_ms": 0.13511034999999996,
+ "samples_ms": [
+ 0.113562,
+ 0.112165,
+ 0.112654,
+ 0.114959,
+ 0.111676,
+ 0.10979,
+ 0.110139,
+ 0.108882,
+ 0.11021,
+ 0.113143,
+ 0.109721,
+ 0.111117,
+ 0.114959,
+ 0.113562,
+ 0.109302,
+ 0.110699,
+ 0.110629,
+ 0.112585,
+ 0.147994,
+ 0.114959,
+ 0.113073,
+ 0.111677,
+ 0.111746,
+ 0.111188,
+ 0.110629,
+ 0.110629,
+ 0.110699,
+ 0.11028,
+ 0.110699,
+ 0.110209,
+ 0.110768,
+ 0.110768,
+ 0.110279,
+ 0.111117,
+ 0.111677,
+ 0.110769,
+ 0.109232,
+ 0.110699,
+ 0.108883,
+ 0.110699
+ ],
+ "warm_rows_per_second": 4622273.580817564
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7520058149132827,
+ "character_perplexity": 5.766156927844759,
+ "next_character_accuracy": 0.5010734549915657,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 9.670835184437542e-05,
+ "mean_prediction_kl_nats": 8.68630843041024e-06,
+ "prediction_kl_p99_nats": 6.698481101176636e-05,
+ "maximum_prediction_kl_nats": 0.004350244493226012,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.6081280708312988,
+ "maximum_per_target_cross_entropy_increase": 0.2908320426940918
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00010893138210684163,
+ "mean_prediction_kl_nats": 4.882093484277991e-06,
+ "prediction_kl_p99_nats": 5.806587418914549e-05,
+ "maximum_prediction_kl_nats": 0.0042292557678851525,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": -7.667535654032331e-05,
+ "maximum_absolute_logit_change": 0.6147077083587646,
+ "maximum_per_target_cross_entropy_increase": 0.2929689884185791
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842471,
+ "unique_rows": 11018,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 10,
+ "compressed_blocks": 82,
+ "zlib_blocks": 82,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11018
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "5e41082603ecc688e3cc05634d6e7a6044a3a438e4063411ccbeffef81325252",
+ "table_sha256": "ae1726ef666d6da487989aa18a00505498337f464d207029fd296a40a6de824d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.09176189918071032,
+ "representatives_sha256": "ee184d53d92e5c9169984d9036ccbafab348e0ccd49bbd1e5093ef78e80edfc3",
+ "structurally_merged_rows": 742,
+ "backend_artifact_bytes": 843715,
+ "backend_with_key_registry_bytes": 1005440,
+ "backend_with_datrie_key_index_bytes": 938995,
+ "full_model_serving_artifact_bytes": 1300126,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9540646943538107,
+ "mean_cosine": 0.999920735356232,
+ "p05_cosine": 0.9996827920998723,
+ "relative_l2_mean": 0.007108773236785336,
+ "relative_l2_p95": 0.025441315375941905,
+ "relative_l2_max": 0.3043084476945251,
+ "weighted_relative_l2_mean": 0.006044384000095293,
+ "weighted_relative_frobenius": 0.010989243866715627,
+ "weighted_mean_cosine": 0.9999521938288747,
+ "max_relative_norm_change": 0.01026371520134981
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9540093139944557,
+ "mean_cosine": 0.9999324396874711,
+ "p05_cosine": 0.9996720616019498,
+ "relative_l2_mean": 0.002407608216982624,
+ "relative_l2_p95": 0.025857957969630616,
+ "relative_l2_max": 0.30465804372306293,
+ "weighted_relative_l2_mean": 0.0011107527975362252,
+ "weighted_relative_frobenius": 0.009598739438612783,
+ "weighted_mean_cosine": 0.9999659814284537,
+ "max_relative_norm_change": 0.01000756191502061
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.846711,
+ "warm_p50_ms": 0.112095,
+ "warm_p95_ms": 0.11526974999999999,
+ "warm_p99_ms": 0.12232507000000001,
+ "samples_ms": [
+ 0.121174,
+ 0.114959,
+ 0.114051,
+ 0.113073,
+ 0.113073,
+ 0.113981,
+ 0.112095,
+ 0.123061,
+ 0.113563,
+ 0.112165,
+ 0.111676,
+ 0.113562,
+ 0.111188,
+ 0.110769,
+ 0.112026,
+ 0.111607,
+ 0.111676,
+ 0.111188,
+ 0.11028,
+ 0.111677,
+ 0.110838,
+ 0.111676,
+ 0.111187,
+ 0.111187,
+ 0.112095,
+ 0.111676,
+ 0.112514,
+ 0.111117,
+ 0.112165,
+ 0.112095,
+ 0.112584,
+ 0.11028,
+ 0.112654,
+ 0.112584,
+ 0.112654,
+ 0.111187,
+ 0.112165,
+ 0.112166,
+ 0.110699,
+ 0.111188
+ ],
+ "warm_rows_per_second": 4567554.306614925
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7521137930062571,
+ "character_perplexity": 5.766779580089415,
+ "next_character_accuracy": 0.5010734549915657,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0002047027176100909,
+ "mean_prediction_kl_nats": 1.4025727395657766e-05,
+ "prediction_kl_p99_nats": 7.21927144228074e-05,
+ "maximum_prediction_kl_nats": 0.01307531830417566,
+ "top1_changed_fraction": 0.0016101824873485662,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.6094646453857422,
+ "maximum_per_target_cross_entropy_increase": 0.8497684001922607
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00021692706776331357,
+ "mean_prediction_kl_nats": 1.0307647465465827e-05,
+ "prediction_kl_p99_nats": 6.713816086402037e-05,
+ "maximum_prediction_kl_nats": 0.013614929907800971,
+ "top1_changed_fraction": 0.0005367274957828554,
+ "accuracy_change": -7.667535654032331e-05,
+ "maximum_absolute_logit_change": 0.6147077083587646,
+ "maximum_per_target_cross_entropy_increase": 0.8540303707122803
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 11760,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 799680,
+ "artifact_bytes": 842173,
+ "unique_rows": 11013,
+ "row_map_bytes": 94080,
+ "n_rows": 11760,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 92,
+ "max_block_bytes": 8192,
+ "raw_blocks": 10,
+ "compressed_blocks": 82,
+ "zlib_blocks": 82,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 11013
+ },
+ "source_float32_sha256": "bde6cf8c4371179f9154641977abf83991df8fdcbb9201ab5fd060bcc3a7ee50",
+ "decoded_float32_sha256": "09b9bc1131925f72342138a773ee0125b11fa4c5913f814ee0264d1dead7d97e",
+ "table_sha256": "c0deda1dc4ffcee3a546e61bf1674f41729ac473d1467336a7d9037e9762bd0b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.09217452094890177,
+ "representatives_sha256": "bd7cc8a0a8c819f2056e2c826f7d6d16ac9fbf79d051fc4e754ffdc8469054a3",
+ "structurally_merged_rows": 747,
+ "backend_artifact_bytes": 843417,
+ "backend_with_key_registry_bytes": 1005142,
+ "backend_with_datrie_key_index_bytes": 938697,
+ "full_model_serving_artifact_bytes": 1299828,
+ "fidelity_vs_float32": {
+ "rows": 11760,
+ "minimum_cosine": 0.9103325952842204,
+ "mean_cosine": 0.9998838947959986,
+ "p05_cosine": 0.999673700062979,
+ "relative_l2_mean": 0.0072978531059799965,
+ "relative_l2_p95": 0.02584965811696233,
+ "relative_l2_max": 0.42304636127986717,
+ "weighted_relative_l2_mean": 0.006134540240910108,
+ "weighted_relative_frobenius": 0.011902670014672963,
+ "weighted_mean_cosine": 0.9999337846253813,
+ "max_relative_norm_change": 0.01026371520134981
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 11760,
+ "minimum_cosine": 0.9106898404260163,
+ "mean_cosine": 0.9998956056307285,
+ "p05_cosine": 0.9996645558119033,
+ "relative_l2_mean": 0.0025988416498503963,
+ "relative_l2_p95": 0.026044948848024656,
+ "relative_l2_max": 0.42225185957192496,
+ "weighted_relative_l2_mean": 0.0012021179656445894,
+ "weighted_relative_frobenius": 0.010629060788327762,
+ "weighted_mean_cosine": 0.9999476288907098,
+ "max_relative_norm_change": 0.01000756191502061
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.888685,
+ "warm_p50_ms": 0.1116415,
+ "warm_p95_ms": 0.12813784999999994,
+ "warm_p99_ms": 0.15508908,
+ "samples_ms": [
+ 0.116845,
+ 0.113493,
+ 0.154279,
+ 0.116426,
+ 0.112095,
+ 0.113073,
+ 0.113004,
+ 0.111118,
+ 0.113073,
+ 0.113003,
+ 0.112515,
+ 0.109791,
+ 0.113562,
+ 0.112584,
+ 0.111117,
+ 0.109791,
+ 0.111188,
+ 0.109232,
+ 0.111188,
+ 0.109302,
+ 0.111187,
+ 0.115867,
+ 0.110629,
+ 0.117264,
+ 0.110769,
+ 0.111118,
+ 0.111607,
+ 0.11021,
+ 0.109302,
+ 0.11014,
+ 0.110769,
+ 0.116356,
+ 0.155607,
+ 0.122083,
+ 0.11028,
+ 0.11028,
+ 0.126762,
+ 0.116844,
+ 0.111676,
+ 0.111257
+ ],
+ "warm_rows_per_second": 4586108.21244788
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7521727945856924,
+ "character_perplexity": 5.767119839230711,
+ "next_character_accuracy": 0.5010734549915657,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00026371811581273876,
+ "mean_prediction_kl_nats": 3.147875233721867e-05,
+ "prediction_kl_p99_nats": 7.920611986419741e-05,
+ "maximum_prediction_kl_nats": 0.05457426993193903,
+ "top1_changed_fraction": 0.0016101824873485662,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.6681852340698242,
+ "maximum_per_target_cross_entropy_increase": 0.8497684001922607
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0002759431872432061,
+ "mean_prediction_kl_nats": 2.771713403775436e-05,
+ "prediction_kl_p99_nats": 7.474973083541969e-05,
+ "maximum_prediction_kl_nats": 0.05451500133355573,
+ "top1_changed_fraction": 0.0005367274957828554,
+ "accuracy_change": -7.667535654032331e-05,
+ "maximum_absolute_logit_change": 0.6665081977844238,
+ "maximum_per_target_cross_entropy_increase": 0.8540303707122803
+ }
+ }
+ ],
+ "baseline_quality": {
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.7519091112373895,
+ "character_perplexity": 5.765599346234589,
+ "next_character_accuracy": 0.5006900782088637,
+ "unknown_targets": 0
+ },
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6054999464765443,
+ "character_perplexity": 4.980348888311733,
+ "next_character_accuracy": 0.531089473280687,
+ "unknown_targets": 0
+ }
+ },
+ "selection": {
+ "fp32": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 2891408,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.605444559273432,
+ "character_perplexity": 4.98007304835536,
+ "next_character_accuracy": 0.5310128038028061,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -5.5385669269262635e-05,
+ "mean_prediction_kl_nats": 2.0394254485204406e-05,
+ "prediction_kl_p99_nats": 6.292058917393448e-05,
+ "maximum_prediction_kl_nats": 0.028937972454724202,
+ "top1_changed_fraction": 0.00038334738940427816,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.6746394634246826,
+ "maximum_per_target_cross_entropy_increase": 0.259132981300354
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00013110323318399718,
+ 2.8910119316644106e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "fp16": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 1565742,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6054448741033316,
+ "character_perplexity": 4.980074616231503,
+ "next_character_accuracy": 0.5310128038028061,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -5.5070856757365275e-05,
+ "mean_prediction_kl_nats": 2.0374639076691894e-05,
+ "prediction_kl_p99_nats": 6.243811826060971e-05,
+ "maximum_prediction_kl_nats": 0.028911209971021787,
+ "top1_changed_fraction": 0.00038334738940427816,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.67431640625,
+ "maximum_per_target_cross_entropy_increase": 0.2590094804763794
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00012999504614585823,
+ 2.8694272122474748e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "int8": {
+ "feasible": true,
+ "name": "int8-raw",
+ "backend_with_key_registry_bytes": 962252,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6055288730005035,
+ "character_perplexity": 4.980492954576834,
+ "next_character_accuracy": 0.5311661427585678,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 2.8926942335284213e-05,
+ "mean_prediction_kl_nats": 3.594473232078083e-06,
+ "prediction_kl_p99_nats": 2.5250991242415894e-05,
+ "maximum_prediction_kl_nats": 9.39906069725258e-05,
+ "top1_changed_fraction": 0.0011500421682128345,
+ "accuracy_change": 7.666947788076417e-05,
+ "maximum_absolute_logit_change": 0.03656506538391113,
+ "maximum_per_target_cross_entropy_increase": 0.019799232482910156
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -2.5258228333548203e-05,
+ 8.626173798269781e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ }
+ },
+ "fixed_quantization_test_controls": {
+ "fp16": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6054995762185011,
+ "character_perplexity": 4.980347044297841,
+ "next_character_accuracy": 0.531089473280687,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -3.702579744376222e-07,
+ "mean_prediction_kl_nats": 3.891884622922279e-09,
+ "prediction_kl_p99_nats": 2.3866377594231902e-08,
+ "maximum_prediction_kl_nats": 1.3001873063553263e-07,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010790824890136719,
+ "maximum_per_target_cross_entropy_increase": 0.0006437301635742188
+ }
+ },
+ "int8": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6055288730005035,
+ "character_perplexity": 4.980492954576834,
+ "next_character_accuracy": 0.5311661427585678,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": 2.8926942335284213e-05,
+ "mean_prediction_kl_nats": 3.594473232078083e-06,
+ "prediction_kl_p99_nats": 2.5250991242415894e-05,
+ "maximum_prediction_kl_nats": 9.39906069725258e-05,
+ "top1_changed_fraction": 0.0011500421682128345,
+ "accuracy_change": 7.666947788076417e-05,
+ "maximum_absolute_logit_change": 0.03656506538391113,
+ "maximum_per_target_cross_entropy_increase": 0.019799232482910156
+ }
+ }
+ },
+ "checkpoint_sha256": "6f98acbd06c462d38cca0009605daa5986430f75993f62e281fc35ba0be9f1eb",
+ "registry_sha256": "7fa792a7a13db17dc47320b6a65018697d338beda3b6328e9775d68716636b54"
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 3,
+ "rows": 36778,
+ "dim": 64,
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "source_float32_bytes": 9415168,
+ "key_registry_bytes": 597521,
+ "unchanged_backbone_artifact_bytes": 699998,
+ "alternative_datrie_key_index_bytes": 294960,
+ "query_rows": 512,
+ "query_ids_sha256": "c0a9c7f836d519eda638b9a28093508b2d62873b170686d55b9ec727411573db",
+ "frequency_trace_sha256": "efa6b79677263433708f5d18746f076a51a377d034e599be829419ac6115bd49",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 9415168,
+ "unique_rows": 36778,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "table_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.050667806062847376,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9415972,
+ "backend_with_key_registry_bytes": 10013493,
+ "backend_with_datrie_key_index_bytes": 9710932,
+ "full_model_serving_artifact_bytes": 10713491,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.292845,
+ "warm_p50_ms": 0.0478065,
+ "warm_p95_ms": 0.048795399999999996,
+ "warm_p99_ms": 0.07334490999999997,
+ "samples_ms": [
+ 0.049658,
+ 0.04826,
+ 0.047353,
+ 0.048261,
+ 0.048191,
+ 0.048331,
+ 0.046864,
+ 0.047282,
+ 0.047771,
+ 0.047212,
+ 0.048749,
+ 0.04819,
+ 0.04826,
+ 0.047353,
+ 0.048261,
+ 0.047841,
+ 0.046864,
+ 0.047772,
+ 0.047282,
+ 0.048749,
+ 0.047771,
+ 0.048679,
+ 0.047282,
+ 0.047352,
+ 0.047771,
+ 0.048749,
+ 0.047771,
+ 0.047353,
+ 0.04875,
+ 0.047772,
+ 0.047772,
+ 0.046794,
+ 0.048679,
+ 0.048191,
+ 0.04826,
+ 0.047283,
+ 0.046374,
+ 0.04826,
+ 0.04833,
+ 0.088489
+ ],
+ "warm_rows_per_second": 10709840.712037066
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594409488138681,
+ "character_perplexity": 6.420146577412286,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 9082790,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1150,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1150,
+ "zlib_blocks": 1150,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "table_sha256": "950b8a20eb569b5f57c76c89f3e6cd1589ac55e0a5e135e564bfce53733c02e8",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.8069775030016899,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9084000,
+ "backend_with_key_registry_bytes": 9681521,
+ "backend_with_datrie_key_index_bytes": 9378960,
+ "full_model_serving_artifact_bytes": 10381519,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 13.541539,
+ "warm_p50_ms": 0.056292,
+ "warm_p95_ms": 0.06284309999999997,
+ "warm_p99_ms": 0.08353361,
+ "samples_ms": [
+ 0.062438,
+ 0.059645,
+ 0.059156,
+ 0.05734,
+ 0.055803,
+ 0.057759,
+ 0.05727,
+ 0.055803,
+ 0.055803,
+ 0.054895,
+ 0.0572,
+ 0.056711,
+ 0.05727,
+ 0.055804,
+ 0.055384,
+ 0.0616,
+ 0.056851,
+ 0.091841,
+ 0.060133,
+ 0.07054,
+ 0.057759,
+ 0.056781,
+ 0.0572,
+ 0.056292,
+ 0.056292,
+ 0.056293,
+ 0.055873,
+ 0.056292,
+ 0.056292,
+ 0.055873,
+ 0.056292,
+ 0.056292,
+ 0.055803,
+ 0.056223,
+ 0.056292,
+ 0.056362,
+ 0.055803,
+ 0.054895,
+ 0.056292,
+ 0.055804
+ ],
+ "warm_rows_per_second": 9095430.967100121
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594409488138681,
+ "character_perplexity": 6.420146577412286,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 9082790,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1150,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1150,
+ "zlib_blocks": 1150,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "table_sha256": "950b8a20eb569b5f57c76c89f3e6cd1589ac55e0a5e135e564bfce53733c02e8",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.923841570969671,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9083999,
+ "backend_with_key_registry_bytes": 9681520,
+ "backend_with_datrie_key_index_bytes": 9378959,
+ "full_model_serving_artifact_bytes": 10381518,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.505699,
+ "warm_p50_ms": 0.055873,
+ "warm_p95_ms": 0.06108735,
+ "warm_p99_ms": 0.06655022999999999,
+ "samples_ms": [
+ 0.061042,
+ 0.057689,
+ 0.057689,
+ 0.05727,
+ 0.05727,
+ 0.057269,
+ 0.061949,
+ 0.069492,
+ 0.056781,
+ 0.055315,
+ 0.05727,
+ 0.055803,
+ 0.055803,
+ 0.055315,
+ 0.055804,
+ 0.055733,
+ 0.055314,
+ 0.055384,
+ 0.055803,
+ 0.055873,
+ 0.055873,
+ 0.054895,
+ 0.054337,
+ 0.055803,
+ 0.055873,
+ 0.056292,
+ 0.056781,
+ 0.056292,
+ 0.054895,
+ 0.056292,
+ 0.055385,
+ 0.056362,
+ 0.055873,
+ 0.055803,
+ 0.056222,
+ 0.056223,
+ 0.056292,
+ 0.055803,
+ 0.055803,
+ 0.055873
+ ],
+ "warm_rows_per_second": 9163638.966942888
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594409488138681,
+ "character_perplexity": 6.420146577412286,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 9082790,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1150,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1150,
+ "zlib_blocks": 1150,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "table_sha256": "950b8a20eb569b5f57c76c89f3e6cd1589ac55e0a5e135e564bfce53733c02e8",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.803027834976092,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9083999,
+ "backend_with_key_registry_bytes": 9681520,
+ "backend_with_datrie_key_index_bytes": 9378959,
+ "full_model_serving_artifact_bytes": 10381518,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.480696,
+ "warm_p50_ms": 0.056257,
+ "warm_p95_ms": 0.058739349999999996,
+ "warm_p99_ms": 0.060133,
+ "samples_ms": [
+ 0.060133,
+ 0.058178,
+ 0.055804,
+ 0.05727,
+ 0.056292,
+ 0.056781,
+ 0.05727,
+ 0.056292,
+ 0.060133,
+ 0.056362,
+ 0.055873,
+ 0.055384,
+ 0.055384,
+ 0.055384,
+ 0.056292,
+ 0.056292,
+ 0.054407,
+ 0.055803,
+ 0.056292,
+ 0.056781,
+ 0.056222,
+ 0.055803,
+ 0.058666,
+ 0.056292,
+ 0.056292,
+ 0.055873,
+ 0.057688,
+ 0.056292,
+ 0.055873,
+ 0.055873,
+ 0.055384,
+ 0.055384,
+ 0.055873,
+ 0.056711,
+ 0.055874,
+ 0.056292,
+ 0.055314,
+ 0.055314,
+ 0.055803,
+ 0.055315
+ ],
+ "warm_rows_per_second": 9101089.642177863
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594409488138681,
+ "character_perplexity": 6.420146577412286,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6841588,
+ "unique_rows": 27388,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 856,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 856,
+ "zlib_blocks": 856,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27388
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "1c7725661f8317f50934a7e4659676736e50fcebf3e970b34f60492a87da20f2",
+ "table_sha256": "827e27d0a480b2dbb2a7508ffea4f56bad0ab6418aabf25ac7e2504fb1027770",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6645706370472908,
+ "representatives_sha256": "14c4fe0f85f8aa238dae85b33062eeebc5d7436acdac3c94b2ed38d87d09f339",
+ "structurally_merged_rows": 9390,
+ "backend_artifact_bytes": 6842794,
+ "backend_with_key_registry_bytes": 7440315,
+ "backend_with_datrie_key_index_bytes": 7137754,
+ "full_model_serving_artifact_bytes": 8140313,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900855318746704,
+ "mean_cosine": 0.9997854104245398,
+ "p05_cosine": 0.9988759100134763,
+ "relative_l2_mean": 0.009961789294431892,
+ "relative_l2_p95": 0.0476501269778969,
+ "relative_l2_max": 0.1407557896887402,
+ "weighted_relative_l2_mean": 0.0024695631320877207,
+ "weighted_relative_frobenius": 0.011433905190859801,
+ "weighted_mean_cosine": 0.9999525446836851,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900855318746704,
+ "mean_cosine": 0.9997854104245398,
+ "p05_cosine": 0.9988759100134763,
+ "relative_l2_mean": 0.009961789294431892,
+ "relative_l2_p95": 0.0476501269778969,
+ "relative_l2_max": 0.1407557896887402,
+ "weighted_relative_l2_mean": 0.0024695631320877207,
+ "weighted_relative_frobenius": 0.011433905190859801,
+ "weighted_mean_cosine": 0.9999525446836851,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.637082,
+ "warm_p50_ms": 0.056781,
+ "warm_p95_ms": 0.0581815,
+ "warm_p99_ms": 0.060548309999999994,
+ "samples_ms": [
+ 0.057689,
+ 0.056292,
+ 0.057759,
+ 0.057689,
+ 0.056781,
+ 0.058248,
+ 0.0572,
+ 0.056851,
+ 0.055873,
+ 0.057689,
+ 0.062019,
+ 0.058178,
+ 0.058178,
+ 0.056851,
+ 0.056781,
+ 0.056851,
+ 0.056292,
+ 0.056781,
+ 0.055873,
+ 0.056781,
+ 0.058177,
+ 0.056781,
+ 0.05727,
+ 0.056711,
+ 0.054406,
+ 0.056781,
+ 0.057269,
+ 0.056292,
+ 0.055733,
+ 0.054895,
+ 0.055803,
+ 0.055734,
+ 0.055315,
+ 0.055873,
+ 0.055873,
+ 0.054895,
+ 0.054895,
+ 0.055385,
+ 0.055873,
+ 0.054476
+ ],
+ "warm_rows_per_second": 9017100.790757472
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594526840220023,
+ "character_perplexity": 6.4202219196107,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.173527699198737e-05,
+ "mean_prediction_kl_nats": 9.793462390566502e-06,
+ "prediction_kl_p99_nats": 0.0001375510538455611,
+ "maximum_prediction_kl_nats": 0.0022487753052020452,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.24757981300354004,
+ "maximum_per_target_cross_entropy_increase": 0.11480426788330078
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 1.173527699198737e-05,
+ "mean_prediction_kl_nats": 9.793462390566502e-06,
+ "prediction_kl_p99_nats": 0.0001375510538455611,
+ "maximum_prediction_kl_nats": 0.0022487753052020452,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.24757981300354004,
+ "maximum_per_target_cross_entropy_increase": 0.11480426788330078
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6836319,
+ "unique_rows": 27366,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 856,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 856,
+ "zlib_blocks": 856,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27366
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "7331d2cf862a77d49588d7ee907cd2df6d09fd45b772ea2b61fa7edd101d475a",
+ "table_sha256": "20a62a52705ab3bf59373bc4f11f72c5620a328bf0773ca3c8d17e3d9a2ef639",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6232985178939998,
+ "representatives_sha256": "8cf46e582cc093f33c06dffa9aca7893e63093ea1e335d018671c59372c2cb08",
+ "structurally_merged_rows": 9412,
+ "backend_artifact_bytes": 6837525,
+ "backend_with_key_registry_bytes": 7435046,
+ "backend_with_datrie_key_index_bytes": 7132485,
+ "full_model_serving_artifact_bytes": 8135044,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9815629527553452,
+ "mean_cosine": 0.9997748831084752,
+ "p05_cosine": 0.9988684734369433,
+ "relative_l2_mean": 0.010080787616436045,
+ "relative_l2_p95": 0.047799700327254446,
+ "relative_l2_max": 0.1924942487658696,
+ "weighted_relative_l2_mean": 0.002527352749144715,
+ "weighted_relative_frobenius": 0.01240934828812918,
+ "weighted_mean_cosine": 0.9999474112652121,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9815629527553452,
+ "mean_cosine": 0.9997748831084752,
+ "p05_cosine": 0.9988684734369433,
+ "relative_l2_mean": 0.010080787616436045,
+ "relative_l2_p95": 0.047799700327254446,
+ "relative_l2_max": 0.1924942487658696,
+ "weighted_relative_l2_mean": 0.002527352749144715,
+ "weighted_relative_frobenius": 0.01240934828812918,
+ "weighted_mean_cosine": 0.9999474112652121,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.505699,
+ "warm_p50_ms": 0.056292,
+ "warm_p95_ms": 0.060979,
+ "warm_p99_ms": 0.08241319999999998,
+ "samples_ms": [
+ 0.060972,
+ 0.058178,
+ 0.058247,
+ 0.056292,
+ 0.056781,
+ 0.05727,
+ 0.056292,
+ 0.055873,
+ 0.055873,
+ 0.055804,
+ 0.056292,
+ 0.057759,
+ 0.056293,
+ 0.055803,
+ 0.056292,
+ 0.058108,
+ 0.056292,
+ 0.056292,
+ 0.055803,
+ 0.056292,
+ 0.055804,
+ 0.055803,
+ 0.096032,
+ 0.061112,
+ 0.057759,
+ 0.056711,
+ 0.056292,
+ 0.056292,
+ 0.056362,
+ 0.055873,
+ 0.0572,
+ 0.056292,
+ 0.056292,
+ 0.056712,
+ 0.056781,
+ 0.055873,
+ 0.056292,
+ 0.054896,
+ 0.056292,
+ 0.056292
+ ],
+ "warm_rows_per_second": 9095430.967100121
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8595709047205486,
+ "character_perplexity": 6.4209809675975675,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0001299643513152377,
+ "mean_prediction_kl_nats": 4.298913405109953e-05,
+ "prediction_kl_p99_nats": 0.00014861154754934762,
+ "maximum_prediction_kl_nats": 0.08667454145389028,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.4670448303222656,
+ "maximum_per_target_cross_entropy_increase": 0.47715890407562256
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0001299643513152377,
+ "mean_prediction_kl_nats": 4.298913405109953e-05,
+ "prediction_kl_p99_nats": 0.00014861154754934762,
+ "maximum_prediction_kl_nats": 0.08667454145389028,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.4670448303222656,
+ "maximum_per_target_cross_entropy_increase": 0.47715890407562256
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6833902,
+ "unique_rows": 27356,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 855,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 855,
+ "zlib_blocks": 855,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27356
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "3e65e968b6f9c8c19e9e2fcf08419f134b7a766ac9a337605696d2e1d5544614",
+ "table_sha256": "e6304456dbdd714753cc346a1bbeca47482e1cc1cad71c033617df5da5224d15",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6291723160538822,
+ "representatives_sha256": "5aea254d1caa9587a4332f8b2ec364b79e79fcf03275a6bcdaff86caa6d063df",
+ "structurally_merged_rows": 9422,
+ "backend_artifact_bytes": 6835108,
+ "backend_with_key_registry_bytes": 7432629,
+ "backend_with_datrie_key_index_bytes": 7130068,
+ "full_model_serving_artifact_bytes": 8132627,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9708261475735659,
+ "mean_cosine": 0.9997643298302546,
+ "p05_cosine": 0.9988642210195045,
+ "relative_l2_mean": 0.010171123678658016,
+ "relative_l2_p95": 0.04787809944360828,
+ "relative_l2_max": 0.24111852823659694,
+ "weighted_relative_l2_mean": 0.002545748913655022,
+ "weighted_relative_frobenius": 0.01270578130435524,
+ "weighted_mean_cosine": 0.9999451854429472,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9708261475735659,
+ "mean_cosine": 0.9997643298302546,
+ "p05_cosine": 0.9988642210195045,
+ "relative_l2_mean": 0.010171123678658016,
+ "relative_l2_p95": 0.04787809944360828,
+ "relative_l2_max": 0.24111852823659694,
+ "weighted_relative_l2_mean": 0.002545748913655022,
+ "weighted_relative_frobenius": 0.01270578130435524,
+ "weighted_mean_cosine": 0.9999451854429472,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.979584,
+ "warm_p50_ms": 0.055873,
+ "warm_p95_ms": 0.059602,
+ "warm_p99_ms": 0.06967683999999999,
+ "samples_ms": [
+ 0.059574,
+ 0.075778,
+ 0.057758,
+ 0.056362,
+ 0.05727,
+ 0.055385,
+ 0.056362,
+ 0.054755,
+ 0.056711,
+ 0.056292,
+ 0.055804,
+ 0.055873,
+ 0.055873,
+ 0.056711,
+ 0.055873,
+ 0.060134,
+ 0.056851,
+ 0.055873,
+ 0.055873,
+ 0.055384,
+ 0.054895,
+ 0.054477,
+ 0.054896,
+ 0.054407,
+ 0.058248,
+ 0.059156,
+ 0.058667,
+ 0.056781,
+ 0.055734,
+ 0.056292,
+ 0.054826,
+ 0.055244,
+ 0.055803,
+ 0.055873,
+ 0.053918,
+ 0.055245,
+ 0.056292,
+ 0.054826,
+ 0.056223,
+ 0.056292
+ ],
+ "warm_rows_per_second": 9163638.966942888
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.859591477530857,
+ "character_perplexity": 6.421113066579826,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0001505400470049534,
+ "mean_prediction_kl_nats": 5.1810581690064186e-05,
+ "prediction_kl_p99_nats": 0.00015809382505059693,
+ "maximum_prediction_kl_nats": 0.08667454145389028,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.5190249681472778,
+ "maximum_per_target_cross_entropy_increase": 0.47715890407562256
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0001505400470049534,
+ "mean_prediction_kl_nats": 5.1810581690064186e-05,
+ "prediction_kl_p99_nats": 0.00015809382505059693,
+ "maximum_prediction_kl_nats": 0.08667454145389028,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.5190249681472778,
+ "maximum_per_target_cross_entropy_increase": 0.47715890407562256
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6831433,
+ "unique_rows": 27346,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 855,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 855,
+ "zlib_blocks": 855,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27346
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "2817d1a62870bf8b0354d9c58a89ba05eaed78cf769b7a1ccb2aca44086c3cb9",
+ "table_sha256": "3a1a8af8cb07bb4b211d74757b973aeb9a29f85fabbebb8c09cb1444107ea2fc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6161359390243888,
+ "representatives_sha256": "f1adb58e0494cb4956a844dc77f063a86c7e9f0926f9c1f4f5315224bdf29a43",
+ "structurally_merged_rows": 9432,
+ "backend_artifact_bytes": 6832639,
+ "backend_with_key_registry_bytes": 7430160,
+ "backend_with_datrie_key_index_bytes": 7127599,
+ "full_model_serving_artifact_bytes": 8130158,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9504960047174151,
+ "mean_cosine": 0.9997458119066871,
+ "p05_cosine": 0.9988581501771227,
+ "relative_l2_mean": 0.010295763089467157,
+ "relative_l2_p95": 0.048060017629512716,
+ "relative_l2_max": 0.3161439583806123,
+ "weighted_relative_l2_mean": 0.0026566571157003198,
+ "weighted_relative_frobenius": 0.01450411579453141,
+ "weighted_mean_cosine": 0.9999284265720063,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9504960047174151,
+ "mean_cosine": 0.9997458119066871,
+ "p05_cosine": 0.9988581501771227,
+ "relative_l2_mean": 0.010295763089467157,
+ "relative_l2_p95": 0.048060017629512716,
+ "relative_l2_max": 0.3161439583806123,
+ "weighted_relative_l2_mean": 0.0026566571157003198,
+ "weighted_relative_frobenius": 0.01450411579453141,
+ "weighted_mean_cosine": 0.9999284265720063,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.883203,
+ "warm_p50_ms": 0.056781,
+ "warm_p95_ms": 0.060175849999999996,
+ "warm_p99_ms": 0.0610137,
+ "samples_ms": [
+ 0.060134,
+ 0.061041,
+ 0.059994,
+ 0.059155,
+ 0.060971,
+ 0.059645,
+ 0.059575,
+ 0.058667,
+ 0.058247,
+ 0.057759,
+ 0.058177,
+ 0.059086,
+ 0.059156,
+ 0.058666,
+ 0.057759,
+ 0.055804,
+ 0.055804,
+ 0.055733,
+ 0.056222,
+ 0.056781,
+ 0.055384,
+ 0.055384,
+ 0.055873,
+ 0.055873,
+ 0.056292,
+ 0.055873,
+ 0.056781,
+ 0.05727,
+ 0.059225,
+ 0.0572,
+ 0.056781,
+ 0.056292,
+ 0.056222,
+ 0.056781,
+ 0.056362,
+ 0.0572,
+ 0.056781,
+ 0.055803,
+ 0.055803,
+ 0.056711
+ ],
+ "warm_rows_per_second": 9017100.790757472
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8595670320928694,
+ "character_perplexity": 6.420956101577093,
+ "next_character_accuracy": 0.4953228032510351,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00012609122783208918,
+ "mean_prediction_kl_nats": 7.923554484544588e-05,
+ "prediction_kl_p99_nats": 0.00022168504694244006,
+ "maximum_prediction_kl_nats": 0.08667454145389028,
+ "top1_changed_fraction": 0.001763533200429382,
+ "accuracy_change": 0.0002300260696212475,
+ "maximum_absolute_logit_change": 1.0080201625823975,
+ "maximum_per_target_cross_entropy_increase": 0.5376654863357544
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00012609122783208918,
+ "mean_prediction_kl_nats": 7.923554484544588e-05,
+ "prediction_kl_p99_nats": 0.00022168504694244006,
+ "maximum_prediction_kl_nats": 0.08667454145389028,
+ "top1_changed_fraction": 0.001763533200429382,
+ "accuracy_change": 0.0002300260696212475,
+ "maximum_absolute_logit_change": 1.0080201625823975,
+ "maximum_per_target_cross_entropy_increase": 0.5376654863357544
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6824522,
+ "unique_rows": 27317,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 854,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 854,
+ "zlib_blocks": 854,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27317
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "6de322ddb4331af83dcbdc2158eb6864b715c23dcb4c13332cc3d5095353b99b",
+ "table_sha256": "e4dfd9fcbe9440fc94361faefca737f02ff988fd9a01d218963ee1b02ff1e1b8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6603471210692078,
+ "representatives_sha256": "c4a6dd1f891ac302499bb67441a2d36b9b204d1c13dbde9ee157bb130b9eed09",
+ "structurally_merged_rows": 9461,
+ "backend_artifact_bytes": 6825728,
+ "backend_with_key_registry_bytes": 7423249,
+ "backend_with_datrie_key_index_bytes": 7120688,
+ "full_model_serving_artifact_bytes": 8123247,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9003755707401346,
+ "mean_cosine": 0.9996617371942216,
+ "p05_cosine": 0.998843291457119,
+ "relative_l2_mean": 0.010714793843897417,
+ "relative_l2_p95": 0.04834942152870287,
+ "relative_l2_max": 0.4471091219856128,
+ "weighted_relative_l2_mean": 0.00280477823538223,
+ "weighted_relative_frobenius": 0.017512481898731582,
+ "weighted_mean_cosine": 0.9998992199370661,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9003755707401346,
+ "mean_cosine": 0.9996617371942216,
+ "p05_cosine": 0.998843291457119,
+ "relative_l2_mean": 0.010714793843897417,
+ "relative_l2_p95": 0.04834942152870287,
+ "relative_l2_max": 0.4471091219856128,
+ "weighted_relative_l2_mean": 0.00280477823538223,
+ "weighted_relative_frobenius": 0.017512481898731582,
+ "weighted_mean_cosine": 0.9998992199370661,
+ "max_relative_norm_change": 0.010100686111549882
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.336334,
+ "warm_p50_ms": 0.0569905,
+ "warm_p95_ms": 0.05968945,
+ "warm_p99_ms": 0.062043839999999996,
+ "samples_ms": [
+ 0.062997,
+ 0.057758,
+ 0.058737,
+ 0.060553,
+ 0.058667,
+ 0.057759,
+ 0.058178,
+ 0.056292,
+ 0.057759,
+ 0.056292,
+ 0.057689,
+ 0.0572,
+ 0.058108,
+ 0.05727,
+ 0.05727,
+ 0.056292,
+ 0.057759,
+ 0.0572,
+ 0.056292,
+ 0.056781,
+ 0.056292,
+ 0.054826,
+ 0.055803,
+ 0.056222,
+ 0.0572,
+ 0.056781,
+ 0.056711,
+ 0.058247,
+ 0.056781,
+ 0.057689,
+ 0.056781,
+ 0.054965,
+ 0.056292,
+ 0.055873,
+ 0.05727,
+ 0.054825,
+ 0.056781,
+ 0.055454,
+ 0.056781,
+ 0.059644
+ ],
+ "warm_rows_per_second": 8983953.465928532
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8597678277036642,
+ "character_perplexity": 6.4222455308309305,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0003269323205221838,
+ "mean_prediction_kl_nats": 0.00018569737129286721,
+ "prediction_kl_p99_nats": 0.0003782698632015452,
+ "maximum_prediction_kl_nats": 0.364815827427902,
+ "top1_changed_fraction": 0.0020702346265910137,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 1.1867132186889648,
+ "maximum_per_target_cross_entropy_increase": 0.9962930083274841
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0003269323205221838,
+ "mean_prediction_kl_nats": 0.00018569737129286721,
+ "prediction_kl_p99_nats": 0.0003782698632015452,
+ "maximum_prediction_kl_nats": 0.364815827427902,
+ "top1_changed_fraction": 0.0020702346265910137,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 1.1867132186889648,
+ "maximum_per_target_cross_entropy_increase": 0.9962930083274841
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 128,
+ "storage": {
+ "artifact_bytes": 4707584,
+ "unique_rows": 36778,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "bac7d3501f88195092d63ba68dffadd151fe5e79cbc8b496139da8f07b095ef3",
+ "table_sha256": "633749180e09d591673ccc3cf1af107d9c7a628d9e13375f718a588d76d8dfc8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06076218606904149,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4708389,
+ "backend_with_key_registry_bytes": 5305910,
+ "backend_with_datrie_key_index_bytes": 5003349,
+ "full_model_serving_artifact_bytes": 6005908,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999541484265,
+ "mean_cosine": 0.9999999792144507,
+ "p05_cosine": 0.9999999715937654,
+ "relative_l2_mean": 0.00020648124862410823,
+ "relative_l2_p95": 0.00024227354853237684,
+ "relative_l2_max": 0.0003216770229553896,
+ "weighted_relative_l2_mean": 0.00011706746574026251,
+ "weighted_relative_frobenius": 0.00020782228325645768,
+ "weighted_mean_cosine": 0.9999999883375167,
+ "max_relative_norm_change": 0.00018137923830007608
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.352559,
+ "warm_p50_ms": 0.1376575,
+ "warm_p95_ms": 0.14418384999999995,
+ "warm_p99_ms": 0.16755114,
+ "samples_ms": [
+ 0.139543,
+ 0.137727,
+ 0.137728,
+ 0.137169,
+ 0.138705,
+ 0.137168,
+ 0.137238,
+ 0.136679,
+ 0.137587,
+ 0.173625,
+ 0.140521,
+ 0.138635,
+ 0.137238,
+ 0.137657,
+ 0.136749,
+ 0.136749,
+ 0.138216,
+ 0.138635,
+ 0.137727,
+ 0.138216,
+ 0.138635,
+ 0.137657,
+ 0.137727,
+ 0.137727,
+ 0.143454,
+ 0.136609,
+ 0.137168,
+ 0.138635,
+ 0.137658,
+ 0.13668,
+ 0.137168,
+ 0.136749,
+ 0.137238,
+ 0.136749,
+ 0.137727,
+ 0.13675,
+ 0.136191,
+ 0.137658,
+ 0.158051,
+ 0.137308
+ ],
+ "warm_rows_per_second": 3719375.9875052213
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594393890174032,
+ "character_perplexity": 6.42013656329816,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.5597952482588541e-06,
+ "mean_prediction_kl_nats": 4.213260959177702e-09,
+ "prediction_kl_p99_nats": 2.854386703663781e-08,
+ "maximum_prediction_kl_nats": 1.3921878729996988e-07,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0011134147644042969,
+ "maximum_per_target_cross_entropy_increase": 0.0008730888366699219
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 4649895,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 575,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 575,
+ "zlib_blocks": 575,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "bac7d3501f88195092d63ba68dffadd151fe5e79cbc8b496139da8f07b095ef3",
+ "table_sha256": "b6d46b5f2b902dd35a8345e7b1b572239da9cd3d5f289262b8ad5a719e8c5a7d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4254511680919677,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4651103,
+ "backend_with_key_registry_bytes": 5248624,
+ "backend_with_datrie_key_index_bytes": 4946063,
+ "full_model_serving_artifact_bytes": 5948622,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999541484265,
+ "mean_cosine": 0.9999999792144507,
+ "p05_cosine": 0.9999999715937654,
+ "relative_l2_mean": 0.00020648124862410823,
+ "relative_l2_p95": 0.00024227354853237684,
+ "relative_l2_max": 0.0003216770229553896,
+ "weighted_relative_l2_mean": 0.00011706746574026251,
+ "weighted_relative_frobenius": 0.00020782228325645768,
+ "weighted_mean_cosine": 0.9999999883375167,
+ "max_relative_norm_change": 0.00018137923830007608
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.82307,
+ "warm_p50_ms": 0.146737,
+ "warm_p95_ms": 0.15446759999999998,
+ "warm_p99_ms": 0.16009572,
+ "samples_ms": [
+ 0.148553,
+ 0.154279,
+ 0.147575,
+ 0.147086,
+ 0.146666,
+ 0.147156,
+ 0.148553,
+ 0.161403,
+ 0.152324,
+ 0.148064,
+ 0.146247,
+ 0.146737,
+ 0.146178,
+ 0.146737,
+ 0.146736,
+ 0.147156,
+ 0.147086,
+ 0.146178,
+ 0.146178,
+ 0.14527,
+ 0.149041,
+ 0.147155,
+ 0.148063,
+ 0.148134,
+ 0.146667,
+ 0.147086,
+ 0.146667,
+ 0.146597,
+ 0.145619,
+ 0.146667,
+ 0.146178,
+ 0.146178,
+ 0.146178,
+ 0.146597,
+ 0.158051,
+ 0.148064,
+ 0.146666,
+ 0.147156,
+ 0.146667,
+ 0.145689
+ ],
+ "warm_rows_per_second": 3489235.843720398
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594393890174032,
+ "character_perplexity": 6.42013656329816,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.5597952482588541e-06,
+ "mean_prediction_kl_nats": 4.213260959177702e-09,
+ "prediction_kl_p99_nats": 2.854386703663781e-08,
+ "maximum_prediction_kl_nats": 1.3921878729996988e-07,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0011134147644042969,
+ "maximum_per_target_cross_entropy_increase": 0.0008730888366699219
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 4649895,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 575,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 575,
+ "zlib_blocks": 575,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "bac7d3501f88195092d63ba68dffadd151fe5e79cbc8b496139da8f07b095ef3",
+ "table_sha256": "b6d46b5f2b902dd35a8345e7b1b572239da9cd3d5f289262b8ad5a719e8c5a7d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6825302711222321,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4651102,
+ "backend_with_key_registry_bytes": 5248623,
+ "backend_with_datrie_key_index_bytes": 4946062,
+ "full_model_serving_artifact_bytes": 5948621,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999541484265,
+ "mean_cosine": 0.9999999792144507,
+ "p05_cosine": 0.9999999715937654,
+ "relative_l2_mean": 0.00020648124862410823,
+ "relative_l2_p95": 0.00024227354853237684,
+ "relative_l2_max": 0.0003216770229553896,
+ "weighted_relative_l2_mean": 0.00011706746574026251,
+ "weighted_relative_frobenius": 0.00020782228325645768,
+ "weighted_mean_cosine": 0.9999999883375167,
+ "max_relative_norm_change": 0.00018137923830007608
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.766778,
+ "warm_p50_ms": 0.1471205,
+ "warm_p95_ms": 0.15076294999999998,
+ "warm_p99_ms": 0.1582056,
+ "samples_ms": [
+ 0.150508,
+ 0.147156,
+ 0.149041,
+ 0.147156,
+ 0.146667,
+ 0.147644,
+ 0.148064,
+ 0.145271,
+ 0.148134,
+ 0.147156,
+ 0.147155,
+ 0.145689,
+ 0.14527,
+ 0.146667,
+ 0.145689,
+ 0.146667,
+ 0.146178,
+ 0.147155,
+ 0.147156,
+ 0.147086,
+ 0.147086,
+ 0.1452,
+ 0.146527,
+ 0.159867,
+ 0.147644,
+ 0.147645,
+ 0.155607,
+ 0.14995,
+ 0.148064,
+ 0.146178,
+ 0.145619,
+ 0.146597,
+ 0.147156,
+ 0.145131,
+ 0.146597,
+ 0.146177,
+ 0.14946,
+ 0.146737,
+ 0.146667,
+ 0.147156
+ ],
+ "warm_rows_per_second": 3480140.429104034
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594393890174032,
+ "character_perplexity": 6.42013656329816,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.5597952482588541e-06,
+ "mean_prediction_kl_nats": 4.213260959177702e-09,
+ "prediction_kl_p99_nats": 2.854386703663781e-08,
+ "maximum_prediction_kl_nats": 1.3921878729996988e-07,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0011134147644042969,
+ "maximum_per_target_cross_entropy_increase": 0.0008730888366699219
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 4649895,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 575,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 575,
+ "zlib_blocks": 575,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "bac7d3501f88195092d63ba68dffadd151fe5e79cbc8b496139da8f07b095ef3",
+ "table_sha256": "b6d46b5f2b902dd35a8345e7b1b572239da9cd3d5f289262b8ad5a719e8c5a7d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4385594120249152,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4651102,
+ "backend_with_key_registry_bytes": 5248623,
+ "backend_with_datrie_key_index_bytes": 4946062,
+ "full_model_serving_artifact_bytes": 5948621,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999541484265,
+ "mean_cosine": 0.9999999792144507,
+ "p05_cosine": 0.9999999715937654,
+ "relative_l2_mean": 0.00020648124862410823,
+ "relative_l2_p95": 0.00024227354853237684,
+ "relative_l2_max": 0.0003216770229553896,
+ "weighted_relative_l2_mean": 0.00011706746574026251,
+ "weighted_relative_frobenius": 0.00020782228325645768,
+ "weighted_mean_cosine": 0.9999999883375167,
+ "max_relative_norm_change": 0.00018137923830007608
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.819718,
+ "warm_p50_ms": 0.14670149999999998,
+ "warm_p95_ms": 0.15149619999999997,
+ "warm_p99_ms": 0.16776439999999998,
+ "samples_ms": [
+ 0.149949,
+ 0.148133,
+ 0.148553,
+ 0.147575,
+ 0.162311,
+ 0.148623,
+ 0.147575,
+ 0.148064,
+ 0.147086,
+ 0.147155,
+ 0.146667,
+ 0.146178,
+ 0.147645,
+ 0.146597,
+ 0.1452,
+ 0.146667,
+ 0.146108,
+ 0.147156,
+ 0.145689,
+ 0.146597,
+ 0.145759,
+ 0.146667,
+ 0.147156,
+ 0.146248,
+ 0.14534,
+ 0.146247,
+ 0.147226,
+ 0.146736,
+ 0.146248,
+ 0.145689,
+ 0.171251,
+ 0.150927,
+ 0.146597,
+ 0.148063,
+ 0.146737,
+ 0.147574,
+ 0.146667,
+ 0.146178,
+ 0.146667,
+ 0.146667
+ ],
+ "warm_rows_per_second": 3490080.1968623362
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594393890174032,
+ "character_perplexity": 6.42013656329816,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.5597952482588541e-06,
+ "mean_prediction_kl_nats": 4.213260959177702e-09,
+ "prediction_kl_p99_nats": 2.854386703663781e-08,
+ "maximum_prediction_kl_nats": 1.3921878729996988e-07,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0011134147644042969,
+ "maximum_per_target_cross_entropy_increase": 0.0008730888366699219
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3542370,
+ "unique_rows": 27385,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27385
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "0694fbbb70e23e1f2330f4e4a4bd67e90524d754e1f7814d7d30e1fdccb6ab54",
+ "table_sha256": "9dc33891ac3391dad255582baca52071e1a01ae67e1858a6afc3099dd7c541ef",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.35916201607324183,
+ "representatives_sha256": "db04327588f9a3aa5cc31d2f0c093936e74a790de83f31538f13a3dfc326e268",
+ "structurally_merged_rows": 9393,
+ "backend_artifact_bytes": 3543576,
+ "backend_with_key_registry_bytes": 4141097,
+ "backend_with_datrie_key_index_bytes": 3838536,
+ "full_model_serving_artifact_bytes": 4841095,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900865899339991,
+ "mean_cosine": 0.9997853877044258,
+ "p05_cosine": 0.9988764796096378,
+ "relative_l2_mean": 0.010117944363448678,
+ "relative_l2_p95": 0.04765732889351271,
+ "relative_l2_max": 0.1407469103080355,
+ "weighted_relative_l2_mean": 0.0025720274252035076,
+ "weighted_relative_frobenius": 0.011436593582149768,
+ "weighted_mean_cosine": 0.999952534177041,
+ "max_relative_norm_change": 0.010205182003499553
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900863895097766,
+ "mean_cosine": 0.9997853991215832,
+ "p05_cosine": 0.9988759034487058,
+ "relative_l2_mean": 0.009964370503308997,
+ "relative_l2_p95": 0.04766672712855255,
+ "relative_l2_max": 0.14074909543199726,
+ "weighted_relative_l2_mean": 0.002469979984095149,
+ "weighted_relative_frobenius": 0.011436043946536932,
+ "weighted_mean_cosine": 0.9999525401329082,
+ "max_relative_norm_change": 0.010100744870441596
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.549571,
+ "warm_p50_ms": 0.146667,
+ "warm_p95_ms": 0.15252289999999996,
+ "warm_p99_ms": 0.17079081,
+ "samples_ms": [
+ 0.150997,
+ 0.151835,
+ 0.149461,
+ 0.148134,
+ 0.147225,
+ 0.146247,
+ 0.147156,
+ 0.165593,
+ 0.148971,
+ 0.147645,
+ 0.146666,
+ 0.146178,
+ 0.145689,
+ 0.145759,
+ 0.146667,
+ 0.146248,
+ 0.144781,
+ 0.147155,
+ 0.146667,
+ 0.146178,
+ 0.1452,
+ 0.147155,
+ 0.146667,
+ 0.146666,
+ 0.146667,
+ 0.146667,
+ 0.147994,
+ 0.146248,
+ 0.146666,
+ 0.147156,
+ 0.146667,
+ 0.146667,
+ 0.146248,
+ 0.174114,
+ 0.150857,
+ 0.146666,
+ 0.149041,
+ 0.148553,
+ 0.148623,
+ 0.150927
+ ],
+ "warm_rows_per_second": 3490901.157042825
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594520654432896,
+ "character_perplexity": 6.420217948199318,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.1116691211380284e-05,
+ "mean_prediction_kl_nats": 9.825111773194992e-06,
+ "prediction_kl_p99_nats": 0.00013710213579112544,
+ "maximum_prediction_kl_nats": 0.0022618663058666774,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.24773406982421875,
+ "maximum_per_target_cross_entropy_increase": 0.11475634574890137
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 1.2676506232393324e-05,
+ "mean_prediction_kl_nats": 9.825667869041563e-06,
+ "prediction_kl_p99_nats": 0.00013732140618753075,
+ "maximum_prediction_kl_nats": 0.0022635538646430566,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.24774503707885742,
+ "maximum_per_target_cross_entropy_increase": 0.11481094360351562
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3539851,
+ "unique_rows": 27363,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27363
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "b4b13434b713229f49673aed4779bc9357dc31162eaf3f85fcb08abf2dfe92f7",
+ "table_sha256": "fbbea0d2c93d4326946256178b4c0eb6be8bcbb59597482b19bc9d1e47d6415f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3622098241467029,
+ "representatives_sha256": "e915520c3051aec64f3383aa90753a5e554c370c05985bb432d2ccd9933afa9f",
+ "structurally_merged_rows": 9415,
+ "backend_artifact_bytes": 3541057,
+ "backend_with_key_registry_bytes": 4138578,
+ "backend_with_datrie_key_index_bytes": 3836017,
+ "full_model_serving_artifact_bytes": 4838576,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9815501946545637,
+ "mean_cosine": 0.9997748591525092,
+ "p05_cosine": 0.9988684684285702,
+ "relative_l2_mean": 0.010236830526590428,
+ "relative_l2_p95": 0.0477853131292943,
+ "relative_l2_max": 0.19247651088946513,
+ "weighted_relative_l2_mean": 0.0026297548699123043,
+ "weighted_relative_frobenius": 0.012411801737386293,
+ "weighted_mean_cosine": 0.9999474011934931,
+ "max_relative_norm_change": 0.010205182003499553
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9815500673213043,
+ "mean_cosine": 0.9997748709067246,
+ "p05_cosine": 0.9988685995725132,
+ "relative_l2_mean": 0.010083375224941304,
+ "relative_l2_p95": 0.04778673273894737,
+ "relative_l2_max": 0.1924785660982116,
+ "weighted_relative_l2_mean": 0.002527764750321193,
+ "weighted_relative_frobenius": 0.012411215838177433,
+ "weighted_mean_cosine": 0.9999474077045158,
+ "max_relative_norm_change": 0.010100744870441596
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.189468,
+ "warm_p50_ms": 0.146702,
+ "warm_p95_ms": 0.1501133,
+ "warm_p99_ms": 0.15735839999999998,
+ "samples_ms": [
+ 0.151905,
+ 0.148483,
+ 0.148622,
+ 0.147645,
+ 0.147575,
+ 0.147225,
+ 0.147086,
+ 0.147156,
+ 0.145759,
+ 0.150019,
+ 0.148622,
+ 0.147575,
+ 0.147156,
+ 0.146178,
+ 0.145689,
+ 0.146667,
+ 0.148064,
+ 0.147644,
+ 0.160845,
+ 0.148134,
+ 0.146248,
+ 0.144781,
+ 0.14527,
+ 0.146178,
+ 0.1452,
+ 0.146666,
+ 0.147086,
+ 0.147155,
+ 0.146667,
+ 0.147156,
+ 0.14534,
+ 0.145829,
+ 0.145759,
+ 0.145759,
+ 0.146667,
+ 0.146248,
+ 0.146737,
+ 0.146667,
+ 0.146667,
+ 0.146178
+ ],
+ "warm_rows_per_second": 3490068.301727311
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8595702675941401,
+ "character_perplexity": 6.420976876622327,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00012932714230602158,
+ "mean_prediction_kl_nats": 4.3033201686805616e-05,
+ "prediction_kl_p99_nats": 0.0001490815844712177,
+ "maximum_prediction_kl_nats": 0.08671272076410663,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.46732473373413086,
+ "maximum_per_target_cross_entropy_increase": 0.47728192806243896
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00013088714171142216,
+ "mean_prediction_kl_nats": 4.3025091672299196e-05,
+ "prediction_kl_p99_nats": 0.00014879088700169135,
+ "maximum_prediction_kl_nats": 0.08669032413353486,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.46715736389160156,
+ "maximum_per_target_cross_entropy_increase": 0.477230966091156
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3538665,
+ "unique_rows": 27353,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27353
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "73fbf3e81b31d9ebc64e3faa28792d2de5001aff0a38b0a24c7abc32e86b1726",
+ "table_sha256": "ca8169d6a027dca645927287981415c86193c1765f9d390c238e4979702823ab",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.35938104009255767,
+ "representatives_sha256": "2db618c1abe835b770e2a465ebca6f5ff9cefbd274bb379a2f47e97ba8554502",
+ "structurally_merged_rows": 9425,
+ "backend_artifact_bytes": 3539871,
+ "backend_with_key_registry_bytes": 4137392,
+ "backend_with_datrie_key_index_bytes": 3834831,
+ "full_model_serving_artifact_bytes": 4837390,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9708309883932715,
+ "mean_cosine": 0.9997643074180224,
+ "p05_cosine": 0.9988638168652845,
+ "relative_l2_mean": 0.01032710216882639,
+ "relative_l2_p95": 0.04787850560692622,
+ "relative_l2_max": 0.24109966425724155,
+ "weighted_relative_l2_mean": 0.002648136997974743,
+ "weighted_relative_frobenius": 0.012708131555205573,
+ "weighted_mean_cosine": 0.9999451757415506,
+ "max_relative_norm_change": 0.010205182003499553
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9708361119773989,
+ "mean_cosine": 0.9997643183447282,
+ "p05_cosine": 0.9988642426036147,
+ "relative_l2_mean": 0.01017370620197273,
+ "relative_l2_p95": 0.04788957832527853,
+ "relative_l2_max": 0.24108222725641676,
+ "weighted_relative_l2_mean": 0.0025461581452442426,
+ "weighted_relative_frobenius": 0.01270749710197277,
+ "weighted_mean_cosine": 0.9999451826424861,
+ "max_relative_norm_change": 0.010100744870441596
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.905144,
+ "warm_p50_ms": 0.1459685,
+ "warm_p95_ms": 0.1539962,
+ "warm_p99_ms": 0.15718058000000001,
+ "samples_ms": [
+ 0.14953,
+ 0.156584,
+ 0.15386,
+ 0.148064,
+ 0.147086,
+ 0.147575,
+ 0.1452,
+ 0.146666,
+ 0.145689,
+ 0.146178,
+ 0.152324,
+ 0.14946,
+ 0.146178,
+ 0.144712,
+ 0.145689,
+ 0.145689,
+ 0.145689,
+ 0.145759,
+ 0.146178,
+ 0.146178,
+ 0.145759,
+ 0.14527,
+ 0.145759,
+ 0.14534,
+ 0.157562,
+ 0.146667,
+ 0.146177,
+ 0.145759,
+ 0.146178,
+ 0.145689,
+ 0.14527,
+ 0.144851,
+ 0.145829,
+ 0.144362,
+ 0.146108,
+ 0.146178,
+ 0.145689,
+ 0.145689,
+ 0.146178,
+ 0.144223
+ ],
+ "warm_rows_per_second": 3507606.0930954283
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8595908286842588,
+ "character_perplexity": 6.421108900263808,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0001498911029398651,
+ "mean_prediction_kl_nats": 5.18532733136159e-05,
+ "prediction_kl_p99_nats": 0.0001583106405913671,
+ "maximum_prediction_kl_nats": 0.08671272076410663,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.5191551446914673,
+ "maximum_per_target_cross_entropy_increase": 0.47728192806243896
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00015145113442088378,
+ "mean_prediction_kl_nats": 5.1843275615881814e-05,
+ "prediction_kl_p99_nats": 0.00015814003066883615,
+ "maximum_prediction_kl_nats": 0.08669032413353486,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.5189602375030518,
+ "maximum_per_target_cross_entropy_increase": 0.477230966091156
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3537485,
+ "unique_rows": 27343,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27343
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "b5f977750c9cc4bfff19896be63ba2779a853aa1ea486ed4d1bbe54b751dbf3b",
+ "table_sha256": "bc7e6d5865c7e395ea55d498a9fca421c3e204562d97c57b21e069148f98f9d9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3649441129527986,
+ "representatives_sha256": "ecc53a425cfa97d6138f491f1e91e1a92379f83822f506c469ec39ef2d34416b",
+ "structurally_merged_rows": 9435,
+ "backend_artifact_bytes": 3538691,
+ "backend_with_key_registry_bytes": 4136212,
+ "backend_with_datrie_key_index_bytes": 3833651,
+ "full_model_serving_artifact_bytes": 4836210,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9504953761164726,
+ "mean_cosine": 0.9997457904811464,
+ "p05_cosine": 0.9988581478455986,
+ "relative_l2_mean": 0.010451682632623872,
+ "relative_l2_p95": 0.048061964801845554,
+ "relative_l2_max": 0.31614041843033924,
+ "weighted_relative_l2_mean": 0.002758998145107052,
+ "weighted_relative_frobenius": 0.014506072163270795,
+ "weighted_mean_cosine": 0.9999284180438613,
+ "max_relative_norm_change": 0.010205182003499553
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.950498967859504,
+ "mean_cosine": 0.9997458022570964,
+ "p05_cosine": 0.9988580421762674,
+ "relative_l2_mean": 0.010298339746944081,
+ "relative_l2_p95": 0.048069961827879114,
+ "relative_l2_max": 0.31613755897546153,
+ "weighted_relative_l2_mean": 0.0026570630640665033,
+ "weighted_relative_frobenius": 0.014505419604932797,
+ "weighted_mean_cosine": 0.9999284259934664,
+ "max_relative_norm_change": 0.010100744870441596
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.909405,
+ "warm_p50_ms": 0.146178,
+ "warm_p95_ms": 0.15272564999999996,
+ "warm_p99_ms": 0.17185356,
+ "samples_ms": [
+ 0.150019,
+ 0.147085,
+ 0.147156,
+ 0.146597,
+ 0.146178,
+ 0.146178,
+ 0.146108,
+ 0.147155,
+ 0.145689,
+ 0.168318,
+ 0.147574,
+ 0.146667,
+ 0.146108,
+ 0.1452,
+ 0.1452,
+ 0.145689,
+ 0.146178,
+ 0.1452,
+ 0.1452,
+ 0.144712,
+ 0.1452,
+ 0.146667,
+ 0.144851,
+ 0.14527,
+ 0.144781,
+ 0.147156,
+ 0.151905,
+ 0.149042,
+ 0.149041,
+ 0.146178,
+ 0.146108,
+ 0.145689,
+ 0.14527,
+ 0.1452,
+ 0.144781,
+ 0.146248,
+ 0.174114,
+ 0.14946,
+ 0.146248,
+ 0.14527
+ ],
+ "warm_rows_per_second": 3502579.0474626822
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8595663824726743,
+ "character_perplexity": 6.420951930395692,
+ "next_character_accuracy": 0.4953228032510351,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00012544152593662485,
+ "mean_prediction_kl_nats": 7.928011616664785e-05,
+ "prediction_kl_p99_nats": 0.00022322135069379742,
+ "maximum_prediction_kl_nats": 0.08671272076410663,
+ "top1_changed_fraction": 0.001763533200429382,
+ "accuracy_change": 0.0002300260696212475,
+ "maximum_absolute_logit_change": 1.0077931880950928,
+ "maximum_per_target_cross_entropy_increase": 0.5376447439193726
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00012700151928124998,
+ "mean_prediction_kl_nats": 7.927086675416382e-05,
+ "prediction_kl_p99_nats": 0.0002215309973247075,
+ "maximum_prediction_kl_nats": 0.08669032413353486,
+ "top1_changed_fraction": 0.001763533200429382,
+ "accuracy_change": 0.0002300260696212475,
+ "maximum_absolute_logit_change": 1.0080904960632324,
+ "maximum_per_target_cross_entropy_increase": 0.5377774238586426
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3533853,
+ "unique_rows": 27314,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 427,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 427,
+ "zlib_blocks": 427,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27314
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "e3e7f26162f4067923c05e65b37de65e3da42c896ff4778b65872707de2501a6",
+ "table_sha256": "698b7300fd5ff947d50097f6da6afd4e67705b130321a25074e3b4c4125a08cd",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3561051299329847,
+ "representatives_sha256": "58e57637a3c052278b4ce8afb52efc7e656cbfedd09f62f0197042726756758c",
+ "structurally_merged_rows": 9464,
+ "backend_artifact_bytes": 3535059,
+ "backend_with_key_registry_bytes": 4132580,
+ "backend_with_datrie_key_index_bytes": 3830019,
+ "full_model_serving_artifact_bytes": 4832578,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9003836142050164,
+ "mean_cosine": 0.9996617185665663,
+ "p05_cosine": 0.9988430210537634,
+ "relative_l2_mean": 0.010870543231925107,
+ "relative_l2_p95": 0.048348905670925466,
+ "relative_l2_max": 0.44709213369033746,
+ "weighted_relative_l2_mean": 0.00290703968431716,
+ "weighted_relative_frobenius": 0.017513905292490955,
+ "weighted_mean_cosine": 0.9998992124310611,
+ "max_relative_norm_change": 0.010205182003499553
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9003784435428869,
+ "mean_cosine": 0.9996617274141952,
+ "p05_cosine": 0.9988431173619823,
+ "relative_l2_mean": 0.01071737221999517,
+ "relative_l2_p95": 0.048325204758531096,
+ "relative_l2_max": 0.4471100803532612,
+ "weighted_relative_l2_mean": 0.0028051833016739594,
+ "weighted_relative_frobenius": 0.017513626537221234,
+ "weighted_mean_cosine": 0.999899218814722,
+ "max_relative_norm_change": 0.010100744870441596
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.170541,
+ "warm_p50_ms": 0.1471905,
+ "warm_p95_ms": 0.150858,
+ "warm_p99_ms": 0.15554401999999998,
+ "samples_ms": [
+ 0.150858,
+ 0.150438,
+ 0.149041,
+ 0.148971,
+ 0.14946,
+ 0.148063,
+ 0.148063,
+ 0.146667,
+ 0.145758,
+ 0.14534,
+ 0.147086,
+ 0.147575,
+ 0.148064,
+ 0.147645,
+ 0.146597,
+ 0.146248,
+ 0.145759,
+ 0.147156,
+ 0.148064,
+ 0.14995,
+ 0.150858,
+ 0.149461,
+ 0.146248,
+ 0.146248,
+ 0.145759,
+ 0.147226,
+ 0.147575,
+ 0.147645,
+ 0.146597,
+ 0.147156,
+ 0.148134,
+ 0.146247,
+ 0.147156,
+ 0.146666,
+ 0.14534,
+ 0.146247,
+ 0.146178,
+ 0.146666,
+ 0.15854,
+ 0.147225
+ ],
+ "warm_rows_per_second": 3478485.3642048906
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8597671584603643,
+ "character_perplexity": 6.422241232787577,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00032626285864902467,
+ "mean_prediction_kl_nats": 0.0001857141024099959,
+ "prediction_kl_p99_nats": 0.0003793403623519644,
+ "maximum_prediction_kl_nats": 0.36471007120037646,
+ "top1_changed_fraction": 0.0020702346265910137,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 1.1866064071655273,
+ "maximum_per_target_cross_entropy_increase": 0.9961137175559998
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00032782316523429894,
+ "mean_prediction_kl_nats": 0.00018572698214540625,
+ "prediction_kl_p99_nats": 0.00037744275644388547,
+ "maximum_prediction_kl_nats": 0.3647813022894302,
+ "top1_changed_fraction": 0.0020702346265910137,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 1.1867146492004395,
+ "maximum_per_target_cross_entropy_increase": 0.9962016940116882
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 68,
+ "storage": {
+ "artifact_bytes": 2500904,
+ "unique_rows": 36778,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "7a21ec1590d5c8c461d39c7936eefcd8c2a50245cde875af4712b68897f74de2",
+ "table_sha256": "23ed365ab52b2ccf14770d2c13290520d0466caddd4128ccac1db828a6843ff3",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06059561506845057,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2501752,
+ "backend_with_key_registry_bytes": 3099273,
+ "backend_with_datrie_key_index_bytes": 2796712,
+ "full_model_serving_artifact_bytes": 3799271,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999235421852686,
+ "mean_cosine": 0.9999882155834846,
+ "p05_cosine": 0.9999769308878261,
+ "relative_l2_mean": 0.004745605336158613,
+ "relative_l2_p95": 0.006838408959516379,
+ "relative_l2_max": 0.012370012480415811,
+ "weighted_relative_l2_mean": 0.003092517752458587,
+ "weighted_relative_frobenius": 0.0053606944760871186,
+ "weighted_mean_cosine": 0.9999913748526958,
+ "max_relative_norm_change": 0.0033418314673439907
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.323505,
+ "warm_p50_ms": 0.09697449999999999,
+ "warm_p95_ms": 0.10129095,
+ "warm_p99_ms": 0.11442762999999999,
+ "samples_ms": [
+ 0.099314,
+ 0.097498,
+ 0.097917,
+ 0.097499,
+ 0.122572,
+ 0.097918,
+ 0.098896,
+ 0.097498,
+ 0.097009,
+ 0.096521,
+ 0.09694,
+ 0.096451,
+ 0.09694,
+ 0.098895,
+ 0.09652,
+ 0.09694,
+ 0.096521,
+ 0.096032,
+ 0.096521,
+ 0.096939,
+ 0.096521,
+ 0.097429,
+ 0.096521,
+ 0.096521,
+ 0.096032,
+ 0.097917,
+ 0.101689,
+ 0.096591,
+ 0.097009,
+ 0.097429,
+ 0.096521,
+ 0.098477,
+ 0.097918,
+ 0.097498,
+ 0.095543,
+ 0.096451,
+ 0.10127,
+ 0.096521,
+ 0.096032,
+ 0.097499
+ ],
+ "warm_rows_per_second": 5279738.487953019
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594522439986367,
+ "character_perplexity": 6.420219094563666,
+ "next_character_accuracy": 0.49547615396411593,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.1295248559438053e-05,
+ "mean_prediction_kl_nats": 3.648268143178588e-06,
+ "prediction_kl_p99_nats": 2.577223089372882e-05,
+ "maximum_prediction_kl_nats": 0.00010449443355391112,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.031384944915771484,
+ "maximum_per_target_cross_entropy_increase": 0.028478145599365234
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2802301,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 307,
+ "max_block_bytes": 8192,
+ "raw_blocks": 131,
+ "compressed_blocks": 176,
+ "zlib_blocks": 176,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "7a21ec1590d5c8c461d39c7936eefcd8c2a50245cde875af4712b68897f74de2",
+ "table_sha256": "fe1f98da27c853a902917702ab5d04fb502ae7236f0ac69e95da5273a0e4a755",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24412318295799196,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2803554,
+ "backend_with_key_registry_bytes": 3401075,
+ "backend_with_datrie_key_index_bytes": 3098514,
+ "full_model_serving_artifact_bytes": 4101073,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999235421852686,
+ "mean_cosine": 0.9999882155834846,
+ "p05_cosine": 0.9999769308878261,
+ "relative_l2_mean": 0.004745605336158613,
+ "relative_l2_p95": 0.006838408959516379,
+ "relative_l2_max": 0.012370012480415811,
+ "weighted_relative_l2_mean": 0.003092517752458587,
+ "weighted_relative_frobenius": 0.0053606944760871186,
+ "weighted_mean_cosine": 0.9999913748526958,
+ "max_relative_norm_change": 0.0033418314673439907
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.727223,
+ "warm_p50_ms": 0.11447,
+ "warm_p95_ms": 0.11701515,
+ "warm_p99_ms": 0.14012006,
+ "samples_ms": [
+ 0.120267,
+ 0.116844,
+ 0.116356,
+ 0.116426,
+ 0.115937,
+ 0.115866,
+ 0.113562,
+ 0.114051,
+ 0.113981,
+ 0.11454,
+ 0.114121,
+ 0.11447,
+ 0.114959,
+ 0.11447,
+ 0.113911,
+ 0.114959,
+ 0.113632,
+ 0.114959,
+ 0.115448,
+ 0.11447,
+ 0.114051,
+ 0.11454,
+ 0.11447,
+ 0.114959,
+ 0.114051,
+ 0.152813,
+ 0.113003,
+ 0.114051,
+ 0.113074,
+ 0.114051,
+ 0.113632,
+ 0.114958,
+ 0.11447,
+ 0.113073,
+ 0.115378,
+ 0.114051,
+ 0.113981,
+ 0.113562,
+ 0.115028,
+ 0.113562
+ ],
+ "warm_rows_per_second": 4472787.629946711
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594522439986367,
+ "character_perplexity": 6.420219094563666,
+ "next_character_accuracy": 0.49547615396411593,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.1295248559438053e-05,
+ "mean_prediction_kl_nats": 3.648268143178588e-06,
+ "prediction_kl_p99_nats": 2.577223089372882e-05,
+ "maximum_prediction_kl_nats": 0.00010449443355391112,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.031384944915771484,
+ "maximum_per_target_cross_entropy_increase": 0.028478145599365234
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2802301,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 307,
+ "max_block_bytes": 8192,
+ "raw_blocks": 131,
+ "compressed_blocks": 176,
+ "zlib_blocks": 176,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "7a21ec1590d5c8c461d39c7936eefcd8c2a50245cde875af4712b68897f74de2",
+ "table_sha256": "fe1f98da27c853a902917702ab5d04fb502ae7236f0ac69e95da5273a0e4a755",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3854801938869059,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2803553,
+ "backend_with_key_registry_bytes": 3401074,
+ "backend_with_datrie_key_index_bytes": 3098513,
+ "full_model_serving_artifact_bytes": 4101072,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999235421852686,
+ "mean_cosine": 0.9999882155834846,
+ "p05_cosine": 0.9999769308878261,
+ "relative_l2_mean": 0.004745605336158613,
+ "relative_l2_p95": 0.006838408959516379,
+ "relative_l2_max": 0.012370012480415811,
+ "weighted_relative_l2_mean": 0.003092517752458587,
+ "weighted_relative_frobenius": 0.0053606944760871186,
+ "weighted_mean_cosine": 0.9999913748526958,
+ "max_relative_norm_change": 0.0033418314673439907
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.744265,
+ "warm_p50_ms": 0.114539,
+ "warm_p95_ms": 0.11700479999999999,
+ "warm_p99_ms": 0.12482267999999999,
+ "samples_ms": [
+ 0.11873,
+ 0.115448,
+ 0.115448,
+ 0.115448,
+ 0.113493,
+ 0.115029,
+ 0.115867,
+ 0.116914,
+ 0.115448,
+ 0.116426,
+ 0.11454,
+ 0.115029,
+ 0.128718,
+ 0.115936,
+ 0.114051,
+ 0.114539,
+ 0.113422,
+ 0.112584,
+ 0.113074,
+ 0.112584,
+ 0.113073,
+ 0.113981,
+ 0.114051,
+ 0.11454,
+ 0.11447,
+ 0.113981,
+ 0.114051,
+ 0.113981,
+ 0.113492,
+ 0.114051,
+ 0.114959,
+ 0.11454,
+ 0.11454,
+ 0.1144,
+ 0.114959,
+ 0.11454,
+ 0.11447,
+ 0.114539,
+ 0.113981,
+ 0.113911
+ ],
+ "warm_rows_per_second": 4470093.15604292
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594522439986367,
+ "character_perplexity": 6.420219094563666,
+ "next_character_accuracy": 0.49547615396411593,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.1295248559438053e-05,
+ "mean_prediction_kl_nats": 3.648268143178588e-06,
+ "prediction_kl_p99_nats": 2.577223089372882e-05,
+ "maximum_prediction_kl_nats": 0.00010449443355391112,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.031384944915771484,
+ "maximum_per_target_cross_entropy_increase": 0.028478145599365234
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2802301,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 307,
+ "max_block_bytes": 8192,
+ "raw_blocks": 131,
+ "compressed_blocks": 176,
+ "zlib_blocks": 176,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "7a21ec1590d5c8c461d39c7936eefcd8c2a50245cde875af4712b68897f74de2",
+ "table_sha256": "fe1f98da27c853a902917702ab5d04fb502ae7236f0ac69e95da5273a0e4a755",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.251635808032006,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2803553,
+ "backend_with_key_registry_bytes": 3401074,
+ "backend_with_datrie_key_index_bytes": 3098513,
+ "full_model_serving_artifact_bytes": 4101072,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999235421852686,
+ "mean_cosine": 0.9999882155834846,
+ "p05_cosine": 0.9999769308878261,
+ "relative_l2_mean": 0.004745605336158613,
+ "relative_l2_p95": 0.006838408959516379,
+ "relative_l2_max": 0.012370012480415811,
+ "weighted_relative_l2_mean": 0.003092517752458587,
+ "weighted_relative_frobenius": 0.0053606944760871186,
+ "weighted_mean_cosine": 0.9999913748526958,
+ "max_relative_norm_change": 0.0033418314673439907
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.734767,
+ "warm_p50_ms": 0.11450450000000001,
+ "warm_p95_ms": 0.1220409,
+ "warm_p99_ms": 0.13058091,
+ "samples_ms": [
+ 0.122013,
+ 0.117753,
+ 0.114959,
+ 0.115447,
+ 0.116426,
+ 0.116356,
+ 0.116845,
+ 0.135702,
+ 0.122571,
+ 0.114889,
+ 0.121105,
+ 0.115028,
+ 0.113562,
+ 0.115937,
+ 0.114959,
+ 0.114959,
+ 0.112165,
+ 0.113632,
+ 0.113562,
+ 0.115029,
+ 0.114051,
+ 0.113911,
+ 0.112585,
+ 0.113073,
+ 0.112514,
+ 0.112584,
+ 0.112585,
+ 0.11447,
+ 0.114051,
+ 0.114539,
+ 0.112654,
+ 0.114051,
+ 0.113562,
+ 0.113981,
+ 0.114051,
+ 0.113981,
+ 0.114958,
+ 0.114959,
+ 0.114889,
+ 0.11447
+ ],
+ "warm_rows_per_second": 4471439.987074743
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594522439986367,
+ "character_perplexity": 6.420219094563666,
+ "next_character_accuracy": 0.49547615396411593,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.1295248559438053e-05,
+ "mean_prediction_kl_nats": 3.648268143178588e-06,
+ "prediction_kl_p99_nats": 2.577223089372882e-05,
+ "maximum_prediction_kl_nats": 0.00010449443355391112,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.031384944915771484,
+ "maximum_per_target_cross_entropy_increase": 0.028478145599365234
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2161923,
+ "unique_rows": 27408,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 229,
+ "max_block_bytes": 8192,
+ "raw_blocks": 60,
+ "compressed_blocks": 169,
+ "zlib_blocks": 169,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27408
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "f01aaa4197a58748b13a903bda6add998e61bb3afcaf191d51ed3a2f9980f3ed",
+ "table_sha256": "e3d8d172bc0e5ffb7b700333a1324985f365aaee68130d8ca92920cc84bc3427",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2589109721593559,
+ "representatives_sha256": "c404d7a33bd005c51ff59902036cf11911785c3a3c86e4f2a72f3e382b6de83f",
+ "structurally_merged_rows": 9370,
+ "backend_artifact_bytes": 2163173,
+ "backend_with_key_registry_bytes": 2760694,
+ "backend_with_datrie_key_index_bytes": 2458133,
+ "full_model_serving_artifact_bytes": 3460692,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9902830008809947,
+ "mean_cosine": 0.9997745580967632,
+ "p05_cosine": 0.9988623939400101,
+ "relative_l2_mean": 0.01360689390174835,
+ "relative_l2_p95": 0.04794353444322512,
+ "relative_l2_max": 0.13922942015502465,
+ "weighted_relative_l2_mean": 0.005234278781483011,
+ "weighted_relative_frobenius": 0.01266636913425628,
+ "weighted_mean_cosine": 0.9999437981882358,
+ "max_relative_norm_change": 0.011360291955908557
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9903389209589103,
+ "mean_cosine": 0.9997809738838896,
+ "p05_cosine": 0.9988492390424457,
+ "relative_l2_mean": 0.01007816341469381,
+ "relative_l2_p95": 0.04822341063628851,
+ "relative_l2_max": 0.1388218665797682,
+ "weighted_relative_l2_mean": 0.0025294326379207833,
+ "weighted_relative_frobenius": 0.011752570284712499,
+ "weighted_mean_cosine": 0.999950583796412,
+ "max_relative_norm_change": 0.01010074239305728
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.911255,
+ "warm_p50_ms": 0.111607,
+ "warm_p95_ms": 0.11677864999999998,
+ "warm_p99_ms": 0.12543893,
+ "samples_ms": [
+ 0.114051,
+ 0.113003,
+ 0.126692,
+ 0.113982,
+ 0.112584,
+ 0.116356,
+ 0.111607,
+ 0.111118,
+ 0.111607,
+ 0.112095,
+ 0.111676,
+ 0.111676,
+ 0.111746,
+ 0.113492,
+ 0.114051,
+ 0.110699,
+ 0.111607,
+ 0.111677,
+ 0.111257,
+ 0.10979,
+ 0.111536,
+ 0.111187,
+ 0.111257,
+ 0.111187,
+ 0.111607,
+ 0.111606,
+ 0.111606,
+ 0.109232,
+ 0.111118,
+ 0.110279,
+ 0.10979,
+ 0.116426,
+ 0.108813,
+ 0.112515,
+ 0.110629,
+ 0.112096,
+ 0.109302,
+ 0.123479,
+ 0.112584,
+ 0.112165
+ ],
+ "warm_rows_per_second": 4587525.872033116
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594851865060764,
+ "character_perplexity": 6.42043059616263,
+ "next_character_accuracy": 0.49555282932065636,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 4.423867070959386e-05,
+ "mean_prediction_kl_nats": 1.40456145856038e-05,
+ "prediction_kl_p99_nats": 0.00016217386332785115,
+ "maximum_prediction_kl_nats": 0.002340309688407767,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": 0.000460052139242495,
+ "maximum_absolute_logit_change": 0.25087714195251465,
+ "maximum_per_target_cross_entropy_increase": 0.11739516258239746
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 3.294305005021718e-05,
+ "mean_prediction_kl_nats": 1.0321497770350946e-05,
+ "prediction_kl_p99_nats": 0.0001315307878000659,
+ "maximum_prediction_kl_nats": 0.0023158615785463044,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.2475517988204956,
+ "maximum_per_target_cross_entropy_increase": 0.11541223526000977
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2160397,
+ "unique_rows": 27387,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 229,
+ "max_block_bytes": 8192,
+ "raw_blocks": 62,
+ "compressed_blocks": 167,
+ "zlib_blocks": 167,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27387
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "ad282e29aaa0515112a327c85e8c943fe32bd5ff0a65f16b6b135aa9a4d708c7",
+ "table_sha256": "71db6e7b0acf5bfcc0df7266267c23968ab42c9c466f667f98cb5e4119911644",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22282429388724267,
+ "representatives_sha256": "d2de16a859b62db0947eb52fa6f461884bc24f9e8349b3b13475bf3d37fdab95",
+ "structurally_merged_rows": 9391,
+ "backend_artifact_bytes": 2161647,
+ "backend_with_key_registry_bytes": 2759168,
+ "backend_with_datrie_key_index_bytes": 2456607,
+ "full_model_serving_artifact_bytes": 3459166,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9816421009306339,
+ "mean_cosine": 0.9997652403899993,
+ "p05_cosine": 0.9988548508913819,
+ "relative_l2_mean": 0.013712651661030487,
+ "relative_l2_p95": 0.048130194023205695,
+ "relative_l2_max": 0.19229178536981864,
+ "weighted_relative_l2_mean": 0.005289604356728374,
+ "weighted_relative_frobenius": 0.013628023787982177,
+ "weighted_mean_cosine": 0.999938779029103,
+ "max_relative_norm_change": 0.011360291955908557
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.981848720195287,
+ "mean_cosine": 0.9997716323604746,
+ "p05_cosine": 0.9988420028161887,
+ "relative_l2_mean": 0.010186784702438669,
+ "relative_l2_p95": 0.04838511617363744,
+ "relative_l2_max": 0.19109288303028826,
+ "weighted_relative_l2_mean": 0.0025862251985459338,
+ "weighted_relative_frobenius": 0.012785517298615963,
+ "weighted_mean_cosine": 0.9999455515365314,
+ "max_relative_norm_change": 0.01010074239305728
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.999744,
+ "warm_p50_ms": 0.11220050000000001,
+ "warm_p95_ms": 0.11740729999999999,
+ "warm_p99_ms": 0.12165480000000001,
+ "samples_ms": [
+ 0.115448,
+ 0.111677,
+ 0.113562,
+ 0.112235,
+ 0.113981,
+ 0.113982,
+ 0.113562,
+ 0.112165,
+ 0.112235,
+ 0.1188,
+ 0.113493,
+ 0.113562,
+ 0.114051,
+ 0.112166,
+ 0.112096,
+ 0.112096,
+ 0.112166,
+ 0.112096,
+ 0.117334,
+ 0.111188,
+ 0.113492,
+ 0.113003,
+ 0.112095,
+ 0.112584,
+ 0.12348,
+ 0.113562,
+ 0.112096,
+ 0.117334,
+ 0.111677,
+ 0.111677,
+ 0.112096,
+ 0.112095,
+ 0.113003,
+ 0.112095,
+ 0.112095,
+ 0.116844,
+ 0.112165,
+ 0.112095,
+ 0.112095,
+ 0.112165
+ ],
+ "warm_rows_per_second": 4563259.52201639
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.85954489198257,
+ "character_perplexity": 6.42081394247449,
+ "next_character_accuracy": 0.49539947860757555,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00010394857098016971,
+ "mean_prediction_kl_nats": 4.382602376892404e-05,
+ "prediction_kl_p99_nats": 0.00016823848963472371,
+ "maximum_prediction_kl_nats": 0.08511632811208247,
+ "top1_changed_fraction": 0.0021469099831314216,
+ "accuracy_change": 0.00030670142616168183,
+ "maximum_absolute_logit_change": 0.4676833152770996,
+ "maximum_per_target_cross_entropy_increase": 0.4725503921508789
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 9.265227589024589e-05,
+ "mean_prediction_kl_nats": 4.034427924357419e-05,
+ "prediction_kl_p99_nats": 0.0001563639767156741,
+ "maximum_prediction_kl_nats": 0.08632221023776415,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": -7.667535654037883e-05,
+ "maximum_absolute_logit_change": 0.46900463104248047,
+ "maximum_per_target_cross_entropy_increase": 0.4750744104385376
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2159741,
+ "unique_rows": 27378,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 229,
+ "max_block_bytes": 8192,
+ "raw_blocks": 61,
+ "compressed_blocks": 168,
+ "zlib_blocks": 168,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27378
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "69458614fd13cf52ebf575d8817b1c4d821a8c1c7e9d947652c124918186a5db",
+ "table_sha256": "97d860efd673013410d4637fb123871f0ecb46f3e99ea34c5c949d00936cf699",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22295161592774093,
+ "representatives_sha256": "e842e6993c70cf7eb2ccf091e48159221e64ec8e727d0a14df416cab65eaca09",
+ "structurally_merged_rows": 9400,
+ "backend_artifact_bytes": 2160991,
+ "backend_with_key_registry_bytes": 2758512,
+ "backend_with_datrie_key_index_bytes": 2455951,
+ "full_model_serving_artifact_bytes": 3458510,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9707975827186466,
+ "mean_cosine": 0.9997560166171646,
+ "p05_cosine": 0.9988474667289791,
+ "relative_l2_mean": 0.013791136960217677,
+ "relative_l2_p95": 0.048213008711153964,
+ "relative_l2_max": 0.24134291688908707,
+ "weighted_relative_l2_mean": 0.005307653165989682,
+ "weighted_relative_frobenius": 0.013898882213643262,
+ "weighted_mean_cosine": 0.9999365518952043,
+ "max_relative_norm_change": 0.011360291955908557
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9707887824572112,
+ "mean_cosine": 0.9997624007244594,
+ "p05_cosine": 0.9988381018375099,
+ "relative_l2_mean": 0.010266526746003363,
+ "relative_l2_p95": 0.04848659250948559,
+ "relative_l2_max": 0.2413508426857053,
+ "weighted_relative_l2_mean": 0.0026046390875125703,
+ "weighted_relative_frobenius": 0.013074109055989915,
+ "weighted_mean_cosine": 0.999943322863491,
+ "max_relative_norm_change": 0.01010074239305728
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.145434,
+ "warm_p50_ms": 0.11185149999999999,
+ "warm_p95_ms": 0.11409974999999999,
+ "warm_p99_ms": 0.13914864999999998,
+ "samples_ms": [
+ 0.116356,
+ 0.113143,
+ 0.113004,
+ 0.113981,
+ 0.112584,
+ 0.113003,
+ 0.112095,
+ 0.111606,
+ 0.113073,
+ 0.112165,
+ 0.111187,
+ 0.111677,
+ 0.112096,
+ 0.111188,
+ 0.112026,
+ 0.111118,
+ 0.109791,
+ 0.113562,
+ 0.113143,
+ 0.111188,
+ 0.153721,
+ 0.113562,
+ 0.111187,
+ 0.111676,
+ 0.111536,
+ 0.112095,
+ 0.110279,
+ 0.111187,
+ 0.110698,
+ 0.111118,
+ 0.111677,
+ 0.111676,
+ 0.112584,
+ 0.112165,
+ 0.110768,
+ 0.111606,
+ 0.111188,
+ 0.112515,
+ 0.112584,
+ 0.111606
+ ],
+ "warm_rows_per_second": 4577497.843122355
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8595651472084707,
+ "character_perplexity": 6.420943998828518,
+ "next_character_accuracy": 0.49539947860757555,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0001242061075425597,
+ "mean_prediction_kl_nats": 5.2638764337911946e-05,
+ "prediction_kl_p99_nats": 0.00018133778386500096,
+ "maximum_prediction_kl_nats": 0.08511632811208247,
+ "top1_changed_fraction": 0.0022235853396718294,
+ "accuracy_change": 0.00030670142616168183,
+ "maximum_absolute_logit_change": 0.5142704248428345,
+ "maximum_per_target_cross_entropy_increase": 0.4725503921508789
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00011290958364130969,
+ "mean_prediction_kl_nats": 4.9197622857850525e-05,
+ "prediction_kl_p99_nats": 0.00016645288080595016,
+ "maximum_prediction_kl_nats": 0.08632221023776415,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": -7.667535654037883e-05,
+ "maximum_absolute_logit_change": 0.5174161195755005,
+ "maximum_per_target_cross_entropy_increase": 0.4750744104385376
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2159196,
+ "unique_rows": 27369,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 229,
+ "max_block_bytes": 8192,
+ "raw_blocks": 58,
+ "compressed_blocks": 171,
+ "zlib_blocks": 171,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27369
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "046c7c0f394517bf27a820dfec412434ce72494adaf29e4930cd4f18a4015163",
+ "table_sha256": "be41f5beb6281bad05b67299662ee40ddec9ae1155bc3567ac86f20373e4a284",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22292598406784236,
+ "representatives_sha256": "d4ebb46412064fb5a48afc1145b3825651cd0041a3b99efba712dca65fe57243",
+ "structurally_merged_rows": 9409,
+ "backend_artifact_bytes": 2160446,
+ "backend_with_key_registry_bytes": 2757967,
+ "backend_with_datrie_key_index_bytes": 2455406,
+ "full_model_serving_artifact_bytes": 3457965,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9503393835533169,
+ "mean_cosine": 0.9997387406374143,
+ "p05_cosine": 0.9988440679432582,
+ "relative_l2_mean": 0.013906177449907827,
+ "relative_l2_p95": 0.048287317929070005,
+ "relative_l2_max": 0.3165663548628563,
+ "weighted_relative_l2_mean": 0.005416989166314837,
+ "weighted_relative_frobenius": 0.01555448737128707,
+ "weighted_mean_cosine": 0.9999198514354147,
+ "max_relative_norm_change": 0.011360291955908557
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9503933292350666,
+ "mean_cosine": 0.9997451200762226,
+ "p05_cosine": 0.9988322194559154,
+ "relative_l2_mean": 0.010382896028430847,
+ "relative_l2_p95": 0.048593494091304056,
+ "relative_l2_max": 0.3162155535296128,
+ "weighted_relative_l2_mean": 0.0027153392047157185,
+ "weighted_relative_frobenius": 0.014827773285010164,
+ "weighted_mean_cosine": 0.999926567519293,
+ "max_relative_norm_change": 0.01010074239305728
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.229313,
+ "warm_p50_ms": 0.11143149999999999,
+ "warm_p95_ms": 0.11547239999999999,
+ "warm_p99_ms": 0.12714108999999998,
+ "samples_ms": [
+ 0.115448,
+ 0.113492,
+ 0.113073,
+ 0.113003,
+ 0.111607,
+ 0.113073,
+ 0.111187,
+ 0.113073,
+ 0.134305,
+ 0.114051,
+ 0.111187,
+ 0.112235,
+ 0.110209,
+ 0.111117,
+ 0.111187,
+ 0.111257,
+ 0.111257,
+ 0.111676,
+ 0.111606,
+ 0.112095,
+ 0.112095,
+ 0.10986,
+ 0.111606,
+ 0.112095,
+ 0.111117,
+ 0.110698,
+ 0.111257,
+ 0.11021,
+ 0.11447,
+ 0.108743,
+ 0.110769,
+ 0.11028,
+ 0.11014,
+ 0.111187,
+ 0.113073,
+ 0.111677,
+ 0.109791,
+ 0.115936,
+ 0.110279,
+ 0.111118
+ ],
+ "warm_rows_per_second": 4594751.035389455
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8595402974727455,
+ "character_perplexity": 6.420784442049521,
+ "next_character_accuracy": 0.49562950467719674,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 9.935359411882117e-05,
+ "mean_prediction_kl_nats": 8.015656015131716e-05,
+ "prediction_kl_p99_nats": 0.00021484519682922647,
+ "maximum_prediction_kl_nats": 0.08511632811208247,
+ "top1_changed_fraction": 0.002453611409293053,
+ "accuracy_change": 0.0005367274957828738,
+ "maximum_absolute_logit_change": 0.9997484683990479,
+ "maximum_per_target_cross_entropy_increase": 0.5335344076156616
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 8.805735092971689e-05,
+ "mean_prediction_kl_nats": 7.68182050921748e-05,
+ "prediction_kl_p99_nats": 0.0002090726426707212,
+ "maximum_prediction_kl_nats": 0.08632221023776415,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 1.0042335987091064,
+ "maximum_per_target_cross_entropy_increase": 0.5382595062255859
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2157353,
+ "unique_rows": 27342,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 56,
+ "compressed_blocks": 172,
+ "zlib_blocks": 172,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27342
+ },
+ "source_float32_sha256": "0e8b9cb031160e8df3fa6ebc3f76ab5e718940d0564b573772b46a06d6812b63",
+ "decoded_float32_sha256": "383b1e1b5a7b13f93bfa7966f558e8d70b522e53a597c2151a8e7c2a638c1e8d",
+ "table_sha256": "9d711388510a55735272b6ab8f3c6caf339089ec1272fc9069d360ac7a08974c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2181943708565086,
+ "representatives_sha256": "b8af1ff6069e469f3da9b4564d33d1ddef34b63c7a7cb9e594537cae38d574ee",
+ "structurally_merged_rows": 9436,
+ "backend_artifact_bytes": 2158603,
+ "backend_with_key_registry_bytes": 2756124,
+ "backend_with_datrie_key_index_bytes": 2453563,
+ "full_model_serving_artifact_bytes": 3456122,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9003178663665001,
+ "mean_cosine": 0.9996587291999485,
+ "p05_cosine": 0.9988343072054923,
+ "relative_l2_mean": 0.014300957383312534,
+ "relative_l2_p95": 0.04852503555078502,
+ "relative_l2_max": 0.4473976520047374,
+ "weighted_relative_l2_mean": 0.00555003852404553,
+ "weighted_relative_frobenius": 0.018010361448504417,
+ "weighted_mean_cosine": 0.9998933863856327,
+ "max_relative_norm_change": 0.011360291955908557
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9001092385885267,
+ "mean_cosine": 0.9996650778254199,
+ "p05_cosine": 0.9988186190338227,
+ "relative_l2_mean": 0.010781167898732758,
+ "relative_l2_p95": 0.04883380186376116,
+ "relative_l2_max": 0.44752145877237726,
+ "weighted_relative_l2_mean": 0.0028503829810415608,
+ "weighted_relative_frobenius": 0.017390962375218838,
+ "weighted_mean_cosine": 0.9999000452268496,
+ "max_relative_norm_change": 0.01010074239305728
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.429827,
+ "warm_p50_ms": 0.1088475,
+ "warm_p95_ms": 0.1150601,
+ "warm_p99_ms": 0.12666128999999998,
+ "samples_ms": [
+ 0.118311,
+ 0.110768,
+ 0.110209,
+ 0.109302,
+ 0.109232,
+ 0.110768,
+ 0.110209,
+ 0.10979,
+ 0.107905,
+ 0.109232,
+ 0.106508,
+ 0.108394,
+ 0.109232,
+ 0.108813,
+ 0.108813,
+ 0.132,
+ 0.108743,
+ 0.108882,
+ 0.109721,
+ 0.109791,
+ 0.108813,
+ 0.108394,
+ 0.108813,
+ 0.108394,
+ 0.108813,
+ 0.109371,
+ 0.108324,
+ 0.114889,
+ 0.109302,
+ 0.108812,
+ 0.108813,
+ 0.108813,
+ 0.106997,
+ 0.107835,
+ 0.108394,
+ 0.109791,
+ 0.108393,
+ 0.109372,
+ 0.109721,
+ 0.107346
+ ],
+ "warm_rows_per_second": 4703828.751234525
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.859759182445135,
+ "character_perplexity": 6.422190009097978,
+ "next_character_accuracy": 0.49555282932065636,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.000318284272960733,
+ "mean_prediction_kl_nats": 0.00018617670185500242,
+ "prediction_kl_p99_nats": 0.00033826775677929126,
+ "maximum_prediction_kl_nats": 0.36318883394217805,
+ "top1_changed_fraction": 0.002760312835454685,
+ "accuracy_change": 0.000460052139242495,
+ "maximum_absolute_logit_change": 1.1912569999694824,
+ "maximum_per_target_cross_entropy_increase": 0.9961937069892883
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00030698555692312536,
+ "mean_prediction_kl_nats": 0.00018294799506168375,
+ "prediction_kl_p99_nats": 0.0003370419292004499,
+ "maximum_prediction_kl_nats": 0.36399144749440143,
+ "top1_changed_fraction": 0.001993559270050606,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.1943440437316895,
+ "maximum_per_target_cross_entropy_increase": 0.995843231678009
+ }
+ }
+ ],
+ "baseline_quality": {
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8594409488138681,
+ "character_perplexity": 6.420146577412286,
+ "next_character_accuracy": 0.49509277718141387,
+ "unknown_targets": 0
+ },
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.732981677202601,
+ "character_perplexity": 5.657497616365887,
+ "next_character_accuracy": 0.5241125507935291,
+ "unknown_targets": 0
+ }
+ },
+ "selection": {
+ "fp32": {
+ "feasible": true,
+ "name": "fp32-joint-0.95",
+ "backend_with_key_registry_bytes": 7430160,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.733085906821203,
+ "character_perplexity": 5.658087325916751,
+ "next_character_accuracy": 0.5241892202714099,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00010423505069745625,
+ "mean_prediction_kl_nats": 0.00013875969726630407,
+ "prediction_kl_p99_nats": 0.00038196038107028794,
+ "maximum_prediction_kl_nats": 0.09296217712975313,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": 7.666947788076417e-05,
+ "maximum_absolute_logit_change": 1.504372239112854,
+ "maximum_per_target_cross_entropy_increase": 0.5252523422241211
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00022564232065017386,
+ 0.0003975916668243442
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "fp16": {
+ "feasible": true,
+ "name": "fp16-joint-0.95",
+ "backend_with_key_registry_bytes": 4136212,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.733086864041349,
+ "character_perplexity": 5.658092741954519,
+ "next_character_accuracy": 0.5241892202714099,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00010519237107762973,
+ "mean_prediction_kl_nats": 0.0001388013511152666,
+ "prediction_kl_p99_nats": 0.0003827823135934586,
+ "maximum_prediction_kl_nats": 0.09302745751300914,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": 7.666947788076417e-05,
+ "maximum_absolute_logit_change": 1.5042415857315063,
+ "maximum_per_target_cross_entropy_increase": 0.5253055095672607
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00022603695046941323,
+ 0.00039856463735664214
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "int8": {
+ "feasible": true,
+ "name": "int8-joint-0.95",
+ "backend_with_key_registry_bytes": 2757967,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7330446700904785,
+ "character_perplexity": 5.657854009703907,
+ "next_character_accuracy": 0.5242658897492908,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 6.299487197111428e-05,
+ "mean_prediction_kl_nats": 0.00013896683904643248,
+ "prediction_kl_p99_nats": 0.00035191132917633707,
+ "maximum_prediction_kl_nats": 0.09188349093556392,
+ "top1_changed_fraction": 0.0017633979912596795,
+ "accuracy_change": 0.00015333895576163936,
+ "maximum_absolute_logit_change": 1.4925585985183716,
+ "maximum_per_target_cross_entropy_increase": 0.5185873508453369
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0002516761914661864,
+ 0.00035003994847103626
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ }
+ },
+ "fixed_quantization_test_controls": {
+ "fp16": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7329812414730335,
+ "character_perplexity": 5.657495151227435,
+ "next_character_accuracy": 0.5241125507935291,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -4.3572947252552845e-07,
+ "mean_prediction_kl_nats": 3.895838651310438e-09,
+ "prediction_kl_p99_nats": 2.7689389975956324e-08,
+ "maximum_prediction_kl_nats": 1.3003728296797735e-07,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0011608600616455078,
+ "maximum_per_target_cross_entropy_increase": 0.0008046627044677734
+ }
+ },
+ "int8": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.732985057497423,
+ "character_perplexity": 5.657516740408108,
+ "next_character_accuracy": 0.5241125507935291,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": 3.380300535330562e-06,
+ "mean_prediction_kl_nats": 3.53532370279057e-06,
+ "prediction_kl_p99_nats": 2.8211052007782173e-05,
+ "maximum_prediction_kl_nats": 0.0001297718267378109,
+ "top1_changed_fraction": 0.0009200337345702676,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.029512405395507812,
+ "maximum_per_target_cross_entropy_increase": 0.020233631134033203
+ }
+ }
+ },
+ "checkpoint_sha256": "0e8078af3c09f2b3fc9f9326eeadf80476c38c986ac134e9d92181c35b724578",
+ "registry_sha256": "e22a7692100a89679e3088226c3399678ae46a06d73d93bef7d898aa00f6fb7c"
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 7,
+ "rows": 36778,
+ "dim": 64,
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "source_float32_bytes": 9415168,
+ "key_registry_bytes": 597521,
+ "unchanged_backbone_artifact_bytes": 699998,
+ "alternative_datrie_key_index_bytes": 294960,
+ "query_rows": 512,
+ "query_ids_sha256": "c0a9c7f836d519eda638b9a28093508b2d62873b170686d55b9ec727411573db",
+ "frequency_trace_sha256": "efa6b79677263433708f5d18746f076a51a377d034e599be829419ac6115bd49",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 9415168,
+ "unique_rows": 36778,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "table_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.052108424017205834,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9415972,
+ "backend_with_key_registry_bytes": 10013493,
+ "backend_with_datrie_key_index_bytes": 9710932,
+ "full_model_serving_artifact_bytes": 10713491,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.288095,
+ "warm_p50_ms": 0.0482605,
+ "warm_p95_ms": 0.0487744,
+ "warm_p99_ms": 0.049238609999999995,
+ "samples_ms": [
+ 0.049238,
+ 0.049239,
+ 0.04875,
+ 0.04875,
+ 0.048261,
+ 0.048331,
+ 0.048261,
+ 0.048749,
+ 0.048749,
+ 0.046864,
+ 0.048261,
+ 0.04868,
+ 0.046375,
+ 0.046374,
+ 0.046863,
+ 0.047353,
+ 0.047842,
+ 0.04875,
+ 0.048261,
+ 0.048191,
+ 0.048749,
+ 0.047841,
+ 0.048749,
+ 0.046794,
+ 0.048261,
+ 0.04875,
+ 0.048261,
+ 0.047772,
+ 0.047842,
+ 0.046375,
+ 0.046863,
+ 0.047771,
+ 0.046375,
+ 0.047772,
+ 0.047772,
+ 0.046374,
+ 0.04833,
+ 0.048749,
+ 0.04826,
+ 0.047771
+ ],
+ "warm_rows_per_second": 10609090.249790201
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525171642296288,
+ "character_perplexity": 6.375848398003521,
+ "next_character_accuracy": 0.5017635332004294,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 9082639,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1150,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1150,
+ "zlib_blocks": 1150,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "table_sha256": "30b4d7b29aac1d211e9639854476d82b2936e9a52bada58e85b31d536ede5398",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.7903805440291762,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9083849,
+ "backend_with_key_registry_bytes": 9681370,
+ "backend_with_datrie_key_index_bytes": 9378809,
+ "full_model_serving_artifact_bytes": 10381368,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.489635,
+ "warm_p50_ms": 0.056327,
+ "warm_p95_ms": 0.05920835,
+ "warm_p99_ms": 0.06821229999999999,
+ "samples_ms": [
+ 0.060203,
+ 0.073333,
+ 0.059156,
+ 0.056222,
+ 0.056292,
+ 0.057758,
+ 0.05727,
+ 0.056781,
+ 0.056292,
+ 0.05727,
+ 0.056362,
+ 0.056362,
+ 0.056292,
+ 0.056292,
+ 0.056781,
+ 0.056222,
+ 0.056292,
+ 0.05685,
+ 0.056781,
+ 0.05727,
+ 0.055734,
+ 0.056222,
+ 0.056362,
+ 0.056361,
+ 0.055873,
+ 0.055873,
+ 0.056292,
+ 0.055803,
+ 0.056781,
+ 0.0572,
+ 0.05727,
+ 0.055873,
+ 0.054476,
+ 0.056362,
+ 0.054965,
+ 0.056293,
+ 0.056781,
+ 0.056292,
+ 0.055734,
+ 0.056223
+ ],
+ "warm_rows_per_second": 9089779.324302731
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525171642296288,
+ "character_perplexity": 6.375848398003521,
+ "next_character_accuracy": 0.5017635332004294,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 9082639,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1150,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1150,
+ "zlib_blocks": 1150,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "table_sha256": "30b4d7b29aac1d211e9639854476d82b2936e9a52bada58e85b31d536ede5398",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.9153684892226011,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9083848,
+ "backend_with_key_registry_bytes": 9681369,
+ "backend_with_datrie_key_index_bytes": 9378808,
+ "full_model_serving_artifact_bytes": 10381367,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.460791,
+ "warm_p50_ms": 0.055559,
+ "warm_p95_ms": 0.0596929,
+ "warm_p99_ms": 0.07587382999999999,
+ "samples_ms": [
+ 0.059644,
+ 0.060622,
+ 0.057688,
+ 0.055873,
+ 0.057689,
+ 0.056781,
+ 0.055314,
+ 0.056222,
+ 0.053917,
+ 0.056362,
+ 0.05727,
+ 0.0572,
+ 0.055384,
+ 0.055873,
+ 0.054476,
+ 0.054895,
+ 0.055874,
+ 0.055873,
+ 0.054895,
+ 0.054965,
+ 0.055873,
+ 0.085625,
+ 0.057689,
+ 0.055385,
+ 0.055873,
+ 0.055384,
+ 0.054895,
+ 0.055314,
+ 0.054895,
+ 0.055384,
+ 0.054895,
+ 0.055314,
+ 0.054895,
+ 0.055803,
+ 0.056223,
+ 0.055315,
+ 0.055314,
+ 0.055733,
+ 0.054825,
+ 0.054406
+ ],
+ "warm_rows_per_second": 9215428.643424109
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525171642296288,
+ "character_perplexity": 6.375848398003521,
+ "next_character_accuracy": 0.5017635332004294,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 9082639,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1150,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1150,
+ "zlib_blocks": 1150,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "table_sha256": "30b4d7b29aac1d211e9639854476d82b2936e9a52bada58e85b31d536ede5398",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.8107393679674715,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9083848,
+ "backend_with_key_registry_bytes": 9681369,
+ "backend_with_datrie_key_index_bytes": 9378808,
+ "full_model_serving_artifact_bytes": 10381367,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.510029,
+ "warm_p50_ms": 0.056292,
+ "warm_p95_ms": 0.05867045,
+ "warm_p99_ms": 0.05988646,
+ "samples_ms": [
+ 0.060622,
+ 0.058667,
+ 0.057689,
+ 0.056292,
+ 0.056362,
+ 0.055873,
+ 0.057689,
+ 0.055804,
+ 0.056781,
+ 0.056292,
+ 0.055734,
+ 0.0572,
+ 0.055454,
+ 0.055384,
+ 0.055873,
+ 0.055873,
+ 0.055873,
+ 0.055384,
+ 0.055873,
+ 0.058178,
+ 0.057689,
+ 0.055384,
+ 0.057758,
+ 0.0572,
+ 0.058736,
+ 0.057759,
+ 0.05727,
+ 0.055873,
+ 0.056292,
+ 0.055803,
+ 0.055384,
+ 0.056293,
+ 0.058667,
+ 0.058178,
+ 0.055803,
+ 0.056292,
+ 0.055384,
+ 0.056292,
+ 0.058178,
+ 0.056361
+ ],
+ "warm_rows_per_second": 9095430.967100121
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525171642296288,
+ "character_perplexity": 6.375848398003521,
+ "next_character_accuracy": 0.5017635332004294,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6834157,
+ "unique_rows": 27356,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 855,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 855,
+ "zlib_blocks": 855,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27356
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "f28974969a3a043c3a80922400a7d15e6c9bbcc97a8358eae5e42f49be64e9c6",
+ "table_sha256": "25bac41ea095284e22567e966e065bee4654e87ba02a7b8e518a043bb76b742f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6719967280514538,
+ "representatives_sha256": "7a281a43d6ba2ebdaab9799fb9ded5634de78634979b5a471bd864fdfb7ff118",
+ "structurally_merged_rows": 9422,
+ "backend_artifact_bytes": 6835363,
+ "backend_with_key_registry_bytes": 7432884,
+ "backend_with_datrie_key_index_bytes": 7130323,
+ "full_model_serving_artifact_bytes": 8132882,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9901735396747365,
+ "mean_cosine": 0.9997820335582835,
+ "p05_cosine": 0.9988600506997894,
+ "relative_l2_mean": 0.010062147313454804,
+ "relative_l2_p95": 0.04801019397134542,
+ "relative_l2_max": 0.1398521329039012,
+ "weighted_relative_l2_mean": 0.002481341189480067,
+ "weighted_relative_frobenius": 0.011233518943283134,
+ "weighted_mean_cosine": 0.9999522274010958,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9901735396747365,
+ "mean_cosine": 0.9997820335582835,
+ "p05_cosine": 0.9988600506997894,
+ "relative_l2_mean": 0.010062147313454804,
+ "relative_l2_p95": 0.04801019397134542,
+ "relative_l2_max": 0.1398521329039012,
+ "weighted_relative_l2_mean": 0.002481341189480067,
+ "weighted_relative_frobenius": 0.011233518943283134,
+ "weighted_mean_cosine": 0.9999522274010958,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.079946,
+ "warm_p50_ms": 0.0558385,
+ "warm_p95_ms": 0.0583632,
+ "warm_p99_ms": 0.06055261,
+ "samples_ms": [
+ 0.060552,
+ 0.058177,
+ 0.0572,
+ 0.060553,
+ 0.058178,
+ 0.058109,
+ 0.05727,
+ 0.058248,
+ 0.056781,
+ 0.05727,
+ 0.055314,
+ 0.055314,
+ 0.055804,
+ 0.055384,
+ 0.0572,
+ 0.056362,
+ 0.056292,
+ 0.057759,
+ 0.056293,
+ 0.057759,
+ 0.055454,
+ 0.055873,
+ 0.054965,
+ 0.055803,
+ 0.055384,
+ 0.054965,
+ 0.055314,
+ 0.055873,
+ 0.055874,
+ 0.055384,
+ 0.055384,
+ 0.055803,
+ 0.053428,
+ 0.055873,
+ 0.054476,
+ 0.053987,
+ 0.054895,
+ 0.054895,
+ 0.054895,
+ 0.054825
+ ],
+ "warm_rows_per_second": 9169300.751273762
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.852589936519802,
+ "character_perplexity": 6.376312399976279,
+ "next_character_accuracy": 0.5013801564177274,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 7.277493814049522e-05,
+ "mean_prediction_kl_nats": 5.884713828666284e-06,
+ "prediction_kl_p99_nats": 0.00010471392741432696,
+ "maximum_prediction_kl_nats": 0.0006206674105765104,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": -0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.23345136642456055,
+ "maximum_per_target_cross_entropy_increase": 0.0630180835723877
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 7.277493814049522e-05,
+ "mean_prediction_kl_nats": 5.884713828666284e-06,
+ "prediction_kl_p99_nats": 0.00010471392741432696,
+ "maximum_prediction_kl_nats": 0.0006206674105765104,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": -0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.23345136642456055,
+ "maximum_per_target_cross_entropy_increase": 0.0630180835723877
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6829367,
+ "unique_rows": 27337,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 855,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 855,
+ "zlib_blocks": 855,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27337
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "97f4cadb60613c7e71e4479292720b7b36028f674ba9b1e0e79e6e354ef56bf9",
+ "table_sha256": "27ad7aa908dcfdd5e6173437e75c7bfed5ac1ea58c61ef59b44d436e820434b5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6307932629715651,
+ "representatives_sha256": "5eff92e33cd485bb98a3eac5abcec46c200631815f44387ae3f54f1c15f6dd8b",
+ "structurally_merged_rows": 9441,
+ "backend_artifact_bytes": 6830573,
+ "backend_with_key_registry_bytes": 7428094,
+ "backend_with_datrie_key_index_bytes": 7125533,
+ "full_model_serving_artifact_bytes": 8128092,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9806045429468069,
+ "mean_cosine": 0.9997744007293583,
+ "p05_cosine": 0.9988530409975274,
+ "relative_l2_mean": 0.010152244071902218,
+ "relative_l2_p95": 0.04814969321960311,
+ "relative_l2_max": 0.19798703115127078,
+ "weighted_relative_l2_mean": 0.002568375099506692,
+ "weighted_relative_frobenius": 0.01331560186565728,
+ "weighted_mean_cosine": 0.9999448020457253,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9806045429468069,
+ "mean_cosine": 0.9997744007293583,
+ "p05_cosine": 0.9988530409975274,
+ "relative_l2_mean": 0.010152244071902218,
+ "relative_l2_p95": 0.04814969321960311,
+ "relative_l2_max": 0.19798703115127078,
+ "weighted_relative_l2_mean": 0.002568375099506692,
+ "weighted_relative_frobenius": 0.01331560186565728,
+ "weighted_mean_cosine": 0.9999448020457253,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.774321,
+ "warm_p50_ms": 0.055733,
+ "warm_p95_ms": 0.05873585,
+ "warm_p99_ms": 0.06905257,
+ "samples_ms": [
+ 0.060063,
+ 0.058666,
+ 0.05727,
+ 0.05685,
+ 0.055384,
+ 0.05727,
+ 0.054825,
+ 0.055315,
+ 0.055315,
+ 0.056292,
+ 0.055733,
+ 0.055733,
+ 0.055803,
+ 0.055734,
+ 0.054826,
+ 0.053918,
+ 0.056223,
+ 0.054337,
+ 0.058248,
+ 0.058108,
+ 0.055873,
+ 0.056851,
+ 0.053987,
+ 0.054406,
+ 0.054965,
+ 0.055873,
+ 0.0748,
+ 0.0572,
+ 0.055733,
+ 0.054825,
+ 0.054336,
+ 0.055384,
+ 0.053498,
+ 0.054895,
+ 0.053917,
+ 0.055873,
+ 0.055454,
+ 0.054476,
+ 0.055873,
+ 0.056362
+ ],
+ "warm_rows_per_second": 9186657.814939085
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525007685305952,
+ "character_perplexity": 6.375743862369073,
+ "next_character_accuracy": 0.5016101824873486,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.6395564624773986e-05,
+ "mean_prediction_kl_nats": 2.1939708812914536e-05,
+ "prediction_kl_p99_nats": 0.00011821235869118451,
+ "maximum_prediction_kl_nats": 0.03485956466685582,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.3628110885620117,
+ "maximum_per_target_cross_entropy_increase": 0.2883567810058594
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -1.6395564624773986e-05,
+ "mean_prediction_kl_nats": 2.1939708812914536e-05,
+ "prediction_kl_p99_nats": 0.00011821235869118451,
+ "maximum_prediction_kl_nats": 0.03485956466685582,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.3628110885620117,
+ "maximum_per_target_cross_entropy_increase": 0.2883567810058594
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6827086,
+ "unique_rows": 27327,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 854,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 854,
+ "zlib_blocks": 854,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27327
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "24b9eb6ce21379fa95579811cd63fe48c92a4b0071cfc389b3b94d36c58d684f",
+ "table_sha256": "31efef37403c38fb200012740cb063ef06eb0f2a419b0e17967e33fc5a4c0d1d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6226584929972887,
+ "representatives_sha256": "1ea48882d40270fe86550574bf480b401cc53b5667517caf97945698b9e913aa",
+ "structurally_merged_rows": 9451,
+ "backend_artifact_bytes": 6828292,
+ "backend_with_key_registry_bytes": 7425813,
+ "backend_with_datrie_key_index_bytes": 7123252,
+ "full_model_serving_artifact_bytes": 8125811,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9709786172758579,
+ "mean_cosine": 0.9997646300753811,
+ "p05_cosine": 0.9988468769184199,
+ "relative_l2_mean": 0.010237178448012207,
+ "relative_l2_p95": 0.048278424585558345,
+ "relative_l2_max": 0.2415125233327801,
+ "weighted_relative_l2_mean": 0.002605521992310969,
+ "weighted_relative_frobenius": 0.014693300249397735,
+ "weighted_mean_cosine": 0.9999405173643419,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9709786172758579,
+ "mean_cosine": 0.9997646300753811,
+ "p05_cosine": 0.9988468769184199,
+ "relative_l2_mean": 0.010237178448012207,
+ "relative_l2_p95": 0.048278424585558345,
+ "relative_l2_max": 0.2415125233327801,
+ "weighted_relative_l2_mean": 0.002605521992310969,
+ "weighted_relative_frobenius": 0.014693300249397735,
+ "weighted_mean_cosine": 0.9999405173643419,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.766288,
+ "warm_p50_ms": 0.055873,
+ "warm_p95_ms": 0.0596694,
+ "warm_p99_ms": 0.07368109999999999,
+ "samples_ms": [
+ 0.059085,
+ 0.060133,
+ 0.059645,
+ 0.056781,
+ 0.059085,
+ 0.082343,
+ 0.057689,
+ 0.056851,
+ 0.057689,
+ 0.058178,
+ 0.05727,
+ 0.055315,
+ 0.055803,
+ 0.055873,
+ 0.055454,
+ 0.055873,
+ 0.056362,
+ 0.055873,
+ 0.055803,
+ 0.057689,
+ 0.056223,
+ 0.055803,
+ 0.055803,
+ 0.055733,
+ 0.055314,
+ 0.055315,
+ 0.055734,
+ 0.055315,
+ 0.056362,
+ 0.056292,
+ 0.055873,
+ 0.055873,
+ 0.055873,
+ 0.055384,
+ 0.055803,
+ 0.056292,
+ 0.055873,
+ 0.055384,
+ 0.056711,
+ 0.055733
+ ],
+ "warm_rows_per_second": 9163638.966942888
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.852489701405089,
+ "character_perplexity": 6.375673301602006,
+ "next_character_accuracy": 0.5016101824873486,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.74624474398571e-05,
+ "mean_prediction_kl_nats": 2.696258723223034e-05,
+ "prediction_kl_p99_nats": 0.00012362042407023673,
+ "maximum_prediction_kl_nats": 0.049842036475113476,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.743290901184082,
+ "maximum_per_target_cross_entropy_increase": 0.2883567810058594
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -2.74624474398571e-05,
+ "mean_prediction_kl_nats": 2.696258723223034e-05,
+ "prediction_kl_p99_nats": 0.00012362042407023673,
+ "maximum_prediction_kl_nats": 0.049842036475113476,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.743290901184082,
+ "maximum_per_target_cross_entropy_increase": 0.2883567810058594
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6824580,
+ "unique_rows": 27316,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 854,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 854,
+ "zlib_blocks": 854,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27316
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "01a343e3ceb0c65c4480ff127064c7876c3e6b42561e1a57db961044e89c4275",
+ "table_sha256": "81cfd761163b880ee94e57808ea7582077121d49861ec2c415cf2e051feb5eeb",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6392606198787689,
+ "representatives_sha256": "0d6ba1d3a2d0cdae0fe754e6813613693a7a16717c6d3309e723887c0e943da5",
+ "structurally_merged_rows": 9462,
+ "backend_artifact_bytes": 6825786,
+ "backend_with_key_registry_bytes": 7423307,
+ "backend_with_datrie_key_index_bytes": 7120746,
+ "full_model_serving_artifact_bytes": 8123305,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9513234441624956,
+ "mean_cosine": 0.9997470418346616,
+ "p05_cosine": 0.9988408482469365,
+ "relative_l2_mean": 0.010360222319673776,
+ "relative_l2_p95": 0.048376897574973085,
+ "relative_l2_max": 0.31129419888155907,
+ "weighted_relative_l2_mean": 0.0027232213547155218,
+ "weighted_relative_frobenius": 0.01834001658780078,
+ "weighted_mean_cosine": 0.9999244913850499,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9513234441624956,
+ "mean_cosine": 0.9997470418346616,
+ "p05_cosine": 0.9988408482469365,
+ "relative_l2_mean": 0.010360222319673776,
+ "relative_l2_p95": 0.048376897574973085,
+ "relative_l2_max": 0.31129419888155907,
+ "weighted_relative_l2_mean": 0.0027232213547155218,
+ "weighted_relative_frobenius": 0.01834001658780078,
+ "weighted_mean_cosine": 0.9999244913850499,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.87126,
+ "warm_p50_ms": 0.055804,
+ "warm_p95_ms": 0.05917595,
+ "warm_p99_ms": 0.0676687,
+ "samples_ms": [
+ 0.059574,
+ 0.058178,
+ 0.056781,
+ 0.054476,
+ 0.056292,
+ 0.056781,
+ 0.057759,
+ 0.059155,
+ 0.055873,
+ 0.054476,
+ 0.055384,
+ 0.056781,
+ 0.055874,
+ 0.054896,
+ 0.055384,
+ 0.055873,
+ 0.055873,
+ 0.055385,
+ 0.056292,
+ 0.055244,
+ 0.055803,
+ 0.055314,
+ 0.055314,
+ 0.054895,
+ 0.053987,
+ 0.056781,
+ 0.055804,
+ 0.056293,
+ 0.055803,
+ 0.05727,
+ 0.055314,
+ 0.072844,
+ 0.055803,
+ 0.054826,
+ 0.054826,
+ 0.055804,
+ 0.055314,
+ 0.056292,
+ 0.054825,
+ 0.056292
+ ],
+ "warm_rows_per_second": 9174969.536233962
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8521260846628,
+ "character_perplexity": 6.373355421481727,
+ "next_character_accuracy": 0.5019168839135102,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0003910031051826457,
+ "mean_prediction_kl_nats": 0.00012668706532695237,
+ "prediction_kl_p99_nats": 0.0001307250665160562,
+ "maximum_prediction_kl_nats": 0.12414267612917526,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.8881440162658691,
+ "maximum_per_target_cross_entropy_increase": 0.2883567810058594
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.0003910031051826457,
+ "mean_prediction_kl_nats": 0.00012668706532695237,
+ "prediction_kl_p99_nats": 0.0001307250665160562,
+ "maximum_prediction_kl_nats": 0.12414267612917526,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.8881440162658691,
+ "maximum_per_target_cross_entropy_increase": 0.2883567810058594
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6818694,
+ "unique_rows": 27292,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 853,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 853,
+ "zlib_blocks": 853,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27292
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "1b9166492d0bc6751440d4327dfd61cd7ac83033da680e93ac8222b01362268f",
+ "table_sha256": "55bf023fb3fc3f81cc2585f60d3a700947480aa2b8ba42109becac4929c08f16",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6190395930316299,
+ "representatives_sha256": "ed134a131973f62254def80cb9064931fc73fe7eef505181819eafb931307193",
+ "structurally_merged_rows": 9486,
+ "backend_artifact_bytes": 6819900,
+ "backend_with_key_registry_bytes": 7417421,
+ "backend_with_datrie_key_index_bytes": 7114860,
+ "full_model_serving_artifact_bytes": 8117419,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9000058268534885,
+ "mean_cosine": 0.9996821959897936,
+ "p05_cosine": 0.9988298505418146,
+ "relative_l2_mean": 0.010683446807069317,
+ "relative_l2_p95": 0.04867132648328578,
+ "relative_l2_max": 0.4457487621960916,
+ "weighted_relative_l2_mean": 0.003096513084103833,
+ "weighted_relative_frobenius": 0.02401041382452133,
+ "weighted_mean_cosine": 0.9998493620848846,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9000058268534885,
+ "mean_cosine": 0.9996821959897936,
+ "p05_cosine": 0.9988298505418146,
+ "relative_l2_mean": 0.010683446807069317,
+ "relative_l2_p95": 0.04867132648328578,
+ "relative_l2_max": 0.4457487621960916,
+ "weighted_relative_l2_mean": 0.003096513084103833,
+ "weighted_relative_frobenius": 0.02401041382452133,
+ "weighted_mean_cosine": 0.9998493620848846,
+ "max_relative_norm_change": 0.010064555420672167
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.088816,
+ "warm_p50_ms": 0.056292,
+ "warm_p95_ms": 0.060132849999999995,
+ "warm_p99_ms": 0.07649893999999999,
+ "samples_ms": [
+ 0.06146,
+ 0.058736,
+ 0.058178,
+ 0.058178,
+ 0.055384,
+ 0.055733,
+ 0.056223,
+ 0.056851,
+ 0.055873,
+ 0.055873,
+ 0.060063,
+ 0.057689,
+ 0.056362,
+ 0.058248,
+ 0.057689,
+ 0.058247,
+ 0.056292,
+ 0.056292,
+ 0.055803,
+ 0.056223,
+ 0.058667,
+ 0.058667,
+ 0.0572,
+ 0.056222,
+ 0.055873,
+ 0.055384,
+ 0.055384,
+ 0.055384,
+ 0.059575,
+ 0.056292,
+ 0.055385,
+ 0.056292,
+ 0.055803,
+ 0.086114,
+ 0.058667,
+ 0.056292,
+ 0.055873,
+ 0.054476,
+ 0.053918,
+ 0.054825
+ ],
+ "warm_rows_per_second": 9095430.967100121
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8521563373934884,
+ "character_perplexity": 6.373548235803439,
+ "next_character_accuracy": 0.5019168839135102,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0003607617459665677,
+ "mean_prediction_kl_nats": 0.00022579112662206196,
+ "prediction_kl_p99_nats": 0.0003546683857314661,
+ "maximum_prediction_kl_nats": 0.12414267612917526,
+ "top1_changed_fraction": 0.001993559270050606,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 1.0743937492370605,
+ "maximum_per_target_cross_entropy_increase": 0.6843762397766113
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.0003607617459665677,
+ "mean_prediction_kl_nats": 0.00022579112662206196,
+ "prediction_kl_p99_nats": 0.0003546683857314661,
+ "maximum_prediction_kl_nats": 0.12414267612917526,
+ "top1_changed_fraction": 0.001993559270050606,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 1.0743937492370605,
+ "maximum_per_target_cross_entropy_increase": 0.6843762397766113
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 128,
+ "storage": {
+ "artifact_bytes": 4707584,
+ "unique_rows": 36778,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "5e1e9eb12dad3c914b817161ca8ba461f69a9a80c798a2f75f2a1c88f00a27a2",
+ "table_sha256": "b9d730c7176cf58fede8bab1219fb42cb9ca785bdab964772b80efc4d4c0dcab",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.05573004693724215,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4708389,
+ "backend_with_key_registry_bytes": 5305910,
+ "backend_with_datrie_key_index_bytes": 5003349,
+ "full_model_serving_artifact_bytes": 6005908,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999588217124,
+ "mean_cosine": 0.999999979190866,
+ "p05_cosine": 0.9999999715194372,
+ "relative_l2_mean": 0.0002065236265028307,
+ "relative_l2_p95": 0.00024247128939954417,
+ "relative_l2_max": 0.00030652577504051233,
+ "weighted_relative_l2_mean": 0.00011753808371651295,
+ "weighted_relative_frobenius": 0.00020704368102547602,
+ "weighted_mean_cosine": 0.9999999882675509,
+ "max_relative_norm_change": 0.00018710984770758387
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.348718,
+ "warm_p50_ms": 0.138565,
+ "warm_p95_ms": 0.14061415,
+ "warm_p99_ms": 0.14722647,
+ "samples_ms": [
+ 0.139962,
+ 0.138216,
+ 0.139543,
+ 0.137657,
+ 0.138146,
+ 0.138635,
+ 0.137658,
+ 0.139543,
+ 0.143733,
+ 0.140032,
+ 0.137658,
+ 0.139054,
+ 0.137098,
+ 0.138565,
+ 0.14045,
+ 0.137727,
+ 0.137727,
+ 0.14946,
+ 0.139543,
+ 0.138636,
+ 0.139054,
+ 0.137098,
+ 0.138565,
+ 0.137169,
+ 0.13668,
+ 0.138566,
+ 0.138565,
+ 0.137587,
+ 0.138565,
+ 0.138635,
+ 0.137239,
+ 0.138635,
+ 0.137727,
+ 0.138635,
+ 0.136819,
+ 0.137727,
+ 0.138146,
+ 0.138635,
+ 0.138146,
+ 0.137238
+ ],
+ "warm_rows_per_second": 3695016.7791289287
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525145874085385,
+ "character_perplexity": 6.375831968604068,
+ "next_character_accuracy": 0.5017635332004294,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.5768177702404637e-06,
+ "mean_prediction_kl_nats": 4.324500096867378e-09,
+ "prediction_kl_p99_nats": 3.221988317093933e-08,
+ "maximum_prediction_kl_nats": 9.34618077573458e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010292530059814453,
+ "maximum_per_target_cross_entropy_increase": 0.0007352828979492188
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 4650449,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 575,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 575,
+ "zlib_blocks": 575,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "5e1e9eb12dad3c914b817161ca8ba461f69a9a80c798a2f75f2a1c88f00a27a2",
+ "table_sha256": "be57013cde218e1ae57134f57ab3cfa76271302604acccd66f3fcb396f2839ef",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.43717795005068183,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4651657,
+ "backend_with_key_registry_bytes": 5249178,
+ "backend_with_datrie_key_index_bytes": 4946617,
+ "full_model_serving_artifact_bytes": 5949176,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999588217124,
+ "mean_cosine": 0.999999979190866,
+ "p05_cosine": 0.9999999715194372,
+ "relative_l2_mean": 0.0002065236265028307,
+ "relative_l2_p95": 0.00024247128939954417,
+ "relative_l2_max": 0.00030652577504051233,
+ "weighted_relative_l2_mean": 0.00011753808371651295,
+ "weighted_relative_frobenius": 0.00020704368102547602,
+ "weighted_mean_cosine": 0.9999999882675509,
+ "max_relative_norm_change": 0.00018710984770758387
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.80505,
+ "warm_p50_ms": 0.146178,
+ "warm_p95_ms": 0.15052895,
+ "warm_p99_ms": 0.16651981999999999,
+ "samples_ms": [
+ 0.149601,
+ 0.148064,
+ 0.148552,
+ 0.147016,
+ 0.147086,
+ 0.148134,
+ 0.146666,
+ 0.146178,
+ 0.148063,
+ 0.149041,
+ 0.147645,
+ 0.148553,
+ 0.147645,
+ 0.146178,
+ 0.1452,
+ 0.144362,
+ 0.14527,
+ 0.145339,
+ 0.150927,
+ 0.145689,
+ 0.176489,
+ 0.14995,
+ 0.146108,
+ 0.1452,
+ 0.14562,
+ 0.150508,
+ 0.146737,
+ 0.145759,
+ 0.14534,
+ 0.145759,
+ 0.145759,
+ 0.146248,
+ 0.145689,
+ 0.146178,
+ 0.145759,
+ 0.145759,
+ 0.144781,
+ 0.145619,
+ 0.150438,
+ 0.145759
+ ],
+ "warm_rows_per_second": 3502579.0474626822
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525145874085385,
+ "character_perplexity": 6.375831968604068,
+ "next_character_accuracy": 0.5017635332004294,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.5768177702404637e-06,
+ "mean_prediction_kl_nats": 4.324500096867378e-09,
+ "prediction_kl_p99_nats": 3.221988317093933e-08,
+ "maximum_prediction_kl_nats": 9.34618077573458e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010292530059814453,
+ "maximum_per_target_cross_entropy_increase": 0.0007352828979492188
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 4650449,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 575,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 575,
+ "zlib_blocks": 575,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "5e1e9eb12dad3c914b817161ca8ba461f69a9a80c798a2f75f2a1c88f00a27a2",
+ "table_sha256": "be57013cde218e1ae57134f57ab3cfa76271302604acccd66f3fcb396f2839ef",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5636359148193151,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4651656,
+ "backend_with_key_registry_bytes": 5249177,
+ "backend_with_datrie_key_index_bytes": 4946616,
+ "full_model_serving_artifact_bytes": 5949175,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999588217124,
+ "mean_cosine": 0.999999979190866,
+ "p05_cosine": 0.9999999715194372,
+ "relative_l2_mean": 0.0002065236265028307,
+ "relative_l2_p95": 0.00024247128939954417,
+ "relative_l2_max": 0.00030652577504051233,
+ "weighted_relative_l2_mean": 0.00011753808371651295,
+ "weighted_relative_frobenius": 0.00020704368102547602,
+ "weighted_mean_cosine": 0.9999999882675509,
+ "max_relative_norm_change": 0.00018710984770758387
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.784168,
+ "warm_p50_ms": 0.14733000000000002,
+ "warm_p95_ms": 0.17061604999999994,
+ "warm_p99_ms": 0.19245598,
+ "samples_ms": [
+ 0.14946,
+ 0.148622,
+ 0.146597,
+ 0.146737,
+ 0.148483,
+ 0.147575,
+ 0.186406,
+ 0.1452,
+ 0.14527,
+ 0.147155,
+ 0.148552,
+ 0.1452,
+ 0.14562,
+ 0.145689,
+ 0.145689,
+ 0.147575,
+ 0.1452,
+ 0.147644,
+ 0.150438,
+ 0.149041,
+ 0.147645,
+ 0.14527,
+ 0.14527,
+ 0.145689,
+ 0.146667,
+ 0.145689,
+ 0.147155,
+ 0.146667,
+ 0.145689,
+ 0.149041,
+ 0.148622,
+ 0.14527,
+ 0.157073,
+ 0.147505,
+ 0.146666,
+ 0.196324,
+ 0.151346,
+ 0.169785,
+ 0.168947,
+ 0.168877
+ ],
+ "warm_rows_per_second": 3475191.746419602
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525145874085385,
+ "character_perplexity": 6.375831968604068,
+ "next_character_accuracy": 0.5017635332004294,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.5768177702404637e-06,
+ "mean_prediction_kl_nats": 4.324500096867378e-09,
+ "prediction_kl_p99_nats": 3.221988317093933e-08,
+ "maximum_prediction_kl_nats": 9.34618077573458e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010292530059814453,
+ "maximum_per_target_cross_entropy_increase": 0.0007352828979492188
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 4650449,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 575,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 575,
+ "zlib_blocks": 575,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "5e1e9eb12dad3c914b817161ca8ba461f69a9a80c798a2f75f2a1c88f00a27a2",
+ "table_sha256": "be57013cde218e1ae57134f57ab3cfa76271302604acccd66f3fcb396f2839ef",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.45110794715583324,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4651656,
+ "backend_with_key_registry_bytes": 5249177,
+ "backend_with_datrie_key_index_bytes": 4946616,
+ "full_model_serving_artifact_bytes": 5949175,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999588217124,
+ "mean_cosine": 0.999999979190866,
+ "p05_cosine": 0.9999999715194372,
+ "relative_l2_mean": 0.0002065236265028307,
+ "relative_l2_p95": 0.00024247128939954417,
+ "relative_l2_max": 0.00030652577504051233,
+ "weighted_relative_l2_mean": 0.00011753808371651295,
+ "weighted_relative_frobenius": 0.00020704368102547602,
+ "weighted_mean_cosine": 0.9999999882675509,
+ "max_relative_norm_change": 0.00018710984770758387
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.764333,
+ "warm_p50_ms": 0.147121,
+ "warm_p95_ms": 0.1518454,
+ "warm_p99_ms": 0.17373103,
+ "samples_ms": [
+ 0.150857,
+ 0.149041,
+ 0.148553,
+ 0.147645,
+ 0.147574,
+ 0.148064,
+ 0.147086,
+ 0.146736,
+ 0.181657,
+ 0.148553,
+ 0.148064,
+ 0.145619,
+ 0.145689,
+ 0.146109,
+ 0.147575,
+ 0.147156,
+ 0.147156,
+ 0.146248,
+ 0.147086,
+ 0.150927,
+ 0.145689,
+ 0.146597,
+ 0.147156,
+ 0.146248,
+ 0.147156,
+ 0.145759,
+ 0.147156,
+ 0.146737,
+ 0.14534,
+ 0.147575,
+ 0.146248,
+ 0.14527,
+ 0.146178,
+ 0.146666,
+ 0.161334,
+ 0.148064,
+ 0.14527,
+ 0.147086,
+ 0.146597,
+ 0.151346
+ ],
+ "warm_rows_per_second": 3480128.601627232
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525145874085385,
+ "character_perplexity": 6.375831968604068,
+ "next_character_accuracy": 0.5017635332004294,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.5768177702404637e-06,
+ "mean_prediction_kl_nats": 4.324500096867378e-09,
+ "prediction_kl_p99_nats": 3.221988317093933e-08,
+ "maximum_prediction_kl_nats": 9.34618077573458e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0010292530059814453,
+ "maximum_per_target_cross_entropy_increase": 0.0007352828979492188
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3538979,
+ "unique_rows": 27356,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27356
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "d12039e13a084e6288da830c7f215584bd264df7430402ac4a44a5536e21a4fd",
+ "table_sha256": "f17fe763c9a1f5d876554c3d10a78248e046947a9549e4c446c2856ebbf013f8",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3687351008411497,
+ "representatives_sha256": "4c62dcd1ad5b215aa55f4e8e28f4ac90fcf41ceb90a25e17d213683799437f9b",
+ "structurally_merged_rows": 9422,
+ "backend_artifact_bytes": 3540185,
+ "backend_with_key_registry_bytes": 4137706,
+ "backend_with_datrie_key_index_bytes": 3835145,
+ "full_model_serving_artifact_bytes": 4837704,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9901762297460064,
+ "mean_cosine": 0.9997820555543439,
+ "p05_cosine": 0.9988603489157089,
+ "relative_l2_mean": 0.010214953275116508,
+ "relative_l2_p95": 0.04802668373932069,
+ "relative_l2_max": 0.139832944885907,
+ "weighted_relative_l2_mean": 0.0025829471377984657,
+ "weighted_relative_frobenius": 0.011234116695951897,
+ "weighted_mean_cosine": 0.9999522363805332,
+ "max_relative_norm_change": 0.01010576944316186
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9901807346928359,
+ "mean_cosine": 0.9997820653830614,
+ "p05_cosine": 0.9988596000008526,
+ "relative_l2_mean": 0.010061514977180067,
+ "relative_l2_p95": 0.04801683914824954,
+ "relative_l2_max": 0.13980063605374016,
+ "weighted_relative_l2_mean": 0.002480531716172749,
+ "weighted_relative_frobenius": 0.011232898757745064,
+ "weighted_mean_cosine": 0.9999522432203907,
+ "max_relative_norm_change": 0.010098569885149627
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.905633,
+ "warm_p50_ms": 0.146422,
+ "warm_p95_ms": 0.1685689,
+ "warm_p99_ms": 0.17256018,
+ "samples_ms": [
+ 0.14946,
+ 0.147644,
+ 0.168387,
+ 0.168527,
+ 0.169365,
+ 0.145759,
+ 0.144712,
+ 0.145689,
+ 0.146108,
+ 0.14995,
+ 0.145759,
+ 0.145759,
+ 0.146737,
+ 0.144851,
+ 0.146667,
+ 0.146737,
+ 0.151905,
+ 0.145689,
+ 0.146736,
+ 0.146667,
+ 0.146666,
+ 0.145759,
+ 0.167479,
+ 0.167968,
+ 0.174603,
+ 0.146178,
+ 0.1452,
+ 0.145619,
+ 0.146178,
+ 0.167899,
+ 0.168039,
+ 0.146177,
+ 0.14527,
+ 0.146178,
+ 0.146178,
+ 0.145759,
+ 0.144712,
+ 0.145689,
+ 0.167898,
+ 0.167968
+ ],
+ "warm_rows_per_second": 3496742.2928248486
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525869792546004,
+ "character_perplexity": 6.376293543557386,
+ "next_character_accuracy": 0.5013801564177274,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.981746209715848e-05,
+ "mean_prediction_kl_nats": 5.893158849198016e-06,
+ "prediction_kl_p99_nats": 0.00010359299454403867,
+ "maximum_prediction_kl_nats": 0.0006213262813282408,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": -0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.2334275245666504,
+ "maximum_per_target_cross_entropy_increase": 0.0630805492401123
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 7.239446641474648e-05,
+ "mean_prediction_kl_nats": 5.884694380958763e-06,
+ "prediction_kl_p99_nats": 0.00010475046088103621,
+ "maximum_prediction_kl_nats": 0.0006223747706168132,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": -0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.23338603973388672,
+ "maximum_per_target_cross_entropy_increase": 0.06292223930358887
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3536848,
+ "unique_rows": 27338,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27338
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "937d1df77474fb2cf4272c65e4fcc0860b5097c45fdd8f80a70eba56cbfe9619",
+ "table_sha256": "dc24bbc961478837897ffaecc48bdcba295e4b1f3e1c1f223644c86431169cb0",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.36003342689946294,
+ "representatives_sha256": "df8348472b426cae6fd7a92bad1815eaf4cb6bbd4a1432defac943607997210a",
+ "structurally_merged_rows": 9440,
+ "backend_artifact_bytes": 3538054,
+ "backend_with_key_registry_bytes": 4135575,
+ "backend_with_datrie_key_index_bytes": 3833014,
+ "full_model_serving_artifact_bytes": 4835573,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9806124557008215,
+ "mean_cosine": 0.999774823139785,
+ "p05_cosine": 0.9988537376886683,
+ "relative_l2_mean": 0.010300252832578329,
+ "relative_l2_p95": 0.04812333611612598,
+ "relative_l2_max": 0.1979532702102594,
+ "weighted_relative_l2_mean": 0.0026642370846242046,
+ "weighted_relative_frobenius": 0.01323357293932989,
+ "weighted_mean_cosine": 0.9999452922090167,
+ "max_relative_norm_change": 0.01010576944316186
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9806137170452336,
+ "mean_cosine": 0.9997748339692338,
+ "p05_cosine": 0.9988534109705031,
+ "relative_l2_mean": 0.010146909778610541,
+ "relative_l2_p95": 0.04811074740344134,
+ "relative_l2_max": 0.19794634814101575,
+ "weighted_relative_l2_mean": 0.002561917971575951,
+ "weighted_relative_frobenius": 0.013232561349049259,
+ "weighted_mean_cosine": 0.9999452995059956,
+ "max_relative_norm_change": 0.010098569885149627
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.058028,
+ "warm_p50_ms": 0.1460035,
+ "warm_p95_ms": 0.16206725,
+ "warm_p99_ms": 0.16884807999999998,
+ "samples_ms": [
+ 0.147086,
+ 0.147156,
+ 0.146178,
+ 0.150439,
+ 0.1452,
+ 0.145759,
+ 0.146736,
+ 0.146667,
+ 0.14527,
+ 0.146667,
+ 0.150997,
+ 0.14527,
+ 0.161753,
+ 0.146667,
+ 0.143803,
+ 0.1452,
+ 0.146248,
+ 0.143873,
+ 0.144851,
+ 0.146248,
+ 0.14527,
+ 0.146736,
+ 0.144781,
+ 0.144781,
+ 0.143803,
+ 0.144781,
+ 0.14527,
+ 0.145829,
+ 0.144292,
+ 0.145759,
+ 0.151346,
+ 0.145689,
+ 0.169366,
+ 0.168038,
+ 0.146248,
+ 0.14534,
+ 0.151905,
+ 0.144781,
+ 0.158959,
+ 0.146667
+ ],
+ "warm_rows_per_second": 3506765.2487782827
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.852498542202316,
+ "character_perplexity": 6.375729667886012,
+ "next_character_accuracy": 0.5016101824873486,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.8621853923990522e-05,
+ "mean_prediction_kl_nats": 2.1932618984387973e-05,
+ "prediction_kl_p99_nats": 0.00011857932071118246,
+ "maximum_prediction_kl_nats": 0.03487250158408524,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.3628807067871094,
+ "maximum_per_target_cross_entropy_increase": 0.28847646713256836
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -1.604507749899088e-05,
+ "mean_prediction_kl_nats": 2.1927497150990994e-05,
+ "prediction_kl_p99_nats": 0.00011877888677840257,
+ "maximum_prediction_kl_nats": 0.034874330753904434,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.3629586696624756,
+ "maximum_per_target_cross_entropy_increase": 0.28856325149536133
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3535500,
+ "unique_rows": 27328,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 427,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 427,
+ "zlib_blocks": 427,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27328
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "d6037ca8e04418feae7eaa8e60f1af3a9834ffafb67ef4912ab3476ec9410a91",
+ "table_sha256": "7e1e0bcd89720660613d6f86f4284b263e2ab07238e0d2b010eb1abeee2fc603",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3660884629935026,
+ "representatives_sha256": "f87cf6f37333a9c633b7ebdec21578fef04fd104c1f2a6a8878fbf5c2368da98",
+ "structurally_merged_rows": 9450,
+ "backend_artifact_bytes": 3536706,
+ "backend_with_key_registry_bytes": 4134227,
+ "backend_with_datrie_key_index_bytes": 3831666,
+ "full_model_serving_artifact_bytes": 4834225,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9709827022114997,
+ "mean_cosine": 0.9997650521906866,
+ "p05_cosine": 0.9988471832102334,
+ "relative_l2_mean": 0.010385134048582139,
+ "relative_l2_p95": 0.048265962543471375,
+ "relative_l2_max": 0.24149555041857607,
+ "weighted_relative_l2_mean": 0.002701355191716331,
+ "weighted_relative_frobenius": 0.014619358124001142,
+ "weighted_mean_cosine": 0.9999410069544857,
+ "max_relative_norm_change": 0.01010576944316186
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.970984984297336,
+ "mean_cosine": 0.9997650624190308,
+ "p05_cosine": 0.9988474333802698,
+ "relative_l2_mean": 0.010231848281874937,
+ "relative_l2_p95": 0.04824688144478295,
+ "relative_l2_max": 0.24149217024845834,
+ "weighted_relative_l2_mean": 0.0025990682858739013,
+ "weighted_relative_frobenius": 0.014618411167244827,
+ "weighted_mean_cosine": 0.9999410141982245,
+ "max_relative_norm_change": 0.010098569885149627
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.073602,
+ "warm_p50_ms": 0.146178,
+ "warm_p95_ms": 0.15004345000000002,
+ "warm_p99_ms": 0.15826231999999998,
+ "samples_ms": [
+ 0.149949,
+ 0.147156,
+ 0.146667,
+ 0.14527,
+ 0.14527,
+ 0.14527,
+ 0.144781,
+ 0.148134,
+ 0.147155,
+ 0.146667,
+ 0.145689,
+ 0.146178,
+ 0.150508,
+ 0.16322,
+ 0.148553,
+ 0.147155,
+ 0.146178,
+ 0.146667,
+ 0.146178,
+ 0.146178,
+ 0.144781,
+ 0.144781,
+ 0.147645,
+ 0.146178,
+ 0.146178,
+ 0.146597,
+ 0.144781,
+ 0.145689,
+ 0.145689,
+ 0.14527,
+ 0.145759,
+ 0.14527,
+ 0.14534,
+ 0.144781,
+ 0.145689,
+ 0.146178,
+ 0.14569,
+ 0.144781,
+ 0.150019,
+ 0.146248
+ ],
+ "warm_rows_per_second": 3502579.0474626822
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8524874754607605,
+ "character_perplexity": 6.375659109723975,
+ "next_character_accuracy": 0.5016101824873486,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.9688328161007978e-05,
+ "mean_prediction_kl_nats": 2.6959347392600337e-05,
+ "prediction_kl_p99_nats": 0.0001232882379827201,
+ "maximum_prediction_kl_nats": 0.04987204538153633,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.7435135841369629,
+ "maximum_per_target_cross_entropy_increase": 0.28847646713256836
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -2.7111580252369287e-05,
+ "mean_prediction_kl_nats": 2.6952233424099144e-05,
+ "prediction_kl_p99_nats": 0.00012372796579465564,
+ "maximum_prediction_kl_nats": 0.049850215883357044,
+ "top1_changed_fraction": 0.0013801564177273425,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.7435731887817383,
+ "maximum_per_target_cross_entropy_increase": 0.28856325149536133
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3534205,
+ "unique_rows": 27317,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 427,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 427,
+ "zlib_blocks": 427,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27317
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "7d4c96dab65308fcac02e0971eb24acda40c83bdf0f213aabb1f7f634d5ba5b4",
+ "table_sha256": "25d43cb46dc00061786907d19646096d7d17cbac7a69e70d20ccd23a33fb66e9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3691822250839323,
+ "representatives_sha256": "5b4650fa10885d37c0ca166e016822c9a5e5da800134c1f27d7ecca2ffcd8cdc",
+ "structurally_merged_rows": 9461,
+ "backend_artifact_bytes": 3535411,
+ "backend_with_key_registry_bytes": 4132932,
+ "backend_with_datrie_key_index_bytes": 3830371,
+ "full_model_serving_artifact_bytes": 4832930,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9513163194091709,
+ "mean_cosine": 0.999747462156296,
+ "p05_cosine": 0.9988418588909616,
+ "relative_l2_mean": 0.01050812330519457,
+ "relative_l2_p95": 0.04836003612999388,
+ "relative_l2_max": 0.311314808110222,
+ "weighted_relative_l2_mean": 0.002819016285307103,
+ "weighted_relative_frobenius": 0.01828155946369207,
+ "weighted_mean_cosine": 0.9999249770262345,
+ "max_relative_norm_change": 0.01010576944316186
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9513112397926677,
+ "mean_cosine": 0.999747472245378,
+ "p05_cosine": 0.9988421820784226,
+ "relative_l2_mean": 0.010354899471465413,
+ "relative_l2_p95": 0.048361269465342645,
+ "relative_l2_max": 0.31133352707774775,
+ "weighted_relative_l2_mean": 0.002716787097585304,
+ "weighted_relative_frobenius": 0.018280694607926192,
+ "weighted_mean_cosine": 0.999924984237895,
+ "max_relative_norm_change": 0.010098569885149627
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.426999,
+ "warm_p50_ms": 0.146143,
+ "warm_p95_ms": 0.1505288,
+ "warm_p99_ms": 0.17334230999999997,
+ "samples_ms": [
+ 0.150438,
+ 0.148133,
+ 0.148553,
+ 0.147156,
+ 0.14527,
+ 0.147156,
+ 0.149531,
+ 0.148064,
+ 0.148063,
+ 0.147575,
+ 0.147086,
+ 0.146736,
+ 0.145689,
+ 0.14527,
+ 0.145689,
+ 0.146597,
+ 0.146178,
+ 0.186825,
+ 0.152254,
+ 0.147086,
+ 0.147086,
+ 0.145201,
+ 0.146178,
+ 0.145689,
+ 0.14527,
+ 0.145689,
+ 0.14527,
+ 0.146248,
+ 0.145759,
+ 0.14527,
+ 0.14534,
+ 0.144781,
+ 0.1452,
+ 0.145689,
+ 0.146108,
+ 0.1452,
+ 0.145689,
+ 0.144781,
+ 0.146667,
+ 0.14527
+ ],
+ "warm_rows_per_second": 3503417.8852220085
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.852123829611428,
+ "character_perplexity": 6.373341049254044,
+ "next_character_accuracy": 0.5019168839135102,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.00039325727228118256,
+ "mean_prediction_kl_nats": 0.00012677548255543656,
+ "prediction_kl_p99_nats": 0.00013120910889351578,
+ "maximum_prediction_kl_nats": 0.12425860140083476,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.8884759545326233,
+ "maximum_per_target_cross_entropy_increase": 0.28847646713256836
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.00039068146122587394,
+ "mean_prediction_kl_nats": 0.00012671646408585208,
+ "prediction_kl_p99_nats": 0.00013086450566998734,
+ "maximum_prediction_kl_nats": 0.12420780785095903,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.8883553147315979,
+ "maximum_per_target_cross_entropy_increase": 0.28856325149536133
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3531723,
+ "unique_rows": 27293,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 427,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 427,
+ "zlib_blocks": 427,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27293
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "1e479118a5075d1ef1b4218623a12cf99944f1bb3c6b655c7b32b9c58f19194e",
+ "table_sha256": "7c7a26b4b5b735fa1b933861cb29cc40cccc070092717cffc74dc36b0112039e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.36045743408612907,
+ "representatives_sha256": "e4ad195c6a7a47cd0145ade2e137180ae92aedf3b273b0e3f77e2acfd45d7c24",
+ "structurally_merged_rows": 9485,
+ "backend_artifact_bytes": 3532929,
+ "backend_with_key_registry_bytes": 4130450,
+ "backend_with_datrie_key_index_bytes": 3827889,
+ "full_model_serving_artifact_bytes": 4830448,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9000114158323158,
+ "mean_cosine": 0.9996845787112755,
+ "p05_cosine": 0.9988313432754199,
+ "relative_l2_mean": 0.010822072466038597,
+ "relative_l2_p95": 0.04865110429368451,
+ "relative_l2_max": 0.4457365469788921,
+ "weighted_relative_l2_mean": 0.0031921173076031014,
+ "weighted_relative_frobenius": 0.02396636485356996,
+ "weighted_mean_cosine": 0.9998498425719602,
+ "max_relative_norm_change": 0.01010576944316186
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9000173825782672,
+ "mean_cosine": 0.9996845898874618,
+ "p05_cosine": 0.998831055200147,
+ "relative_l2_mean": 0.010668983896081393,
+ "relative_l2_p95": 0.048643796092008695,
+ "relative_l2_max": 0.44572907924552096,
+ "weighted_relative_l2_mean": 0.003090102065101061,
+ "weighted_relative_frobenius": 0.023965896259199232,
+ "weighted_mean_cosine": 0.9998498482765686,
+ "max_relative_norm_change": 0.010098569885149627
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.418967,
+ "warm_p50_ms": 0.1464575,
+ "warm_p95_ms": 0.15809009999999996,
+ "warm_p99_ms": 0.19397468999999998,
+ "samples_ms": [
+ 0.150927,
+ 0.148971,
+ 0.148552,
+ 0.146667,
+ 0.147155,
+ 0.146178,
+ 0.157493,
+ 0.209664,
+ 0.152953,
+ 0.147644,
+ 0.146667,
+ 0.146667,
+ 0.146667,
+ 0.146178,
+ 0.145689,
+ 0.146667,
+ 0.144711,
+ 0.147155,
+ 0.146667,
+ 0.146178,
+ 0.145759,
+ 0.144781,
+ 0.146248,
+ 0.145759,
+ 0.146247,
+ 0.14527,
+ 0.146737,
+ 0.146248,
+ 0.145759,
+ 0.14527,
+ 0.14527,
+ 0.144781,
+ 0.146248,
+ 0.148552,
+ 0.169435,
+ 0.150438,
+ 0.146667,
+ 0.145689,
+ 0.146248,
+ 0.145201
+ ],
+ "warm_rows_per_second": 3495894.713483434
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8521540448173928,
+ "character_perplexity": 6.373533623975859,
+ "next_character_accuracy": 0.5019168839135102,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0003630534923612783,
+ "mean_prediction_kl_nats": 0.00022589165931384048,
+ "prediction_kl_p99_nats": 0.0003547592823600955,
+ "maximum_prediction_kl_nats": 0.12425860140083476,
+ "top1_changed_fraction": 0.001993559270050606,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 1.0748991966247559,
+ "maximum_per_target_cross_entropy_increase": 0.6844158172607422
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.0003604776034761322,
+ "mean_prediction_kl_nats": 0.00022582647437277177,
+ "prediction_kl_p99_nats": 0.0003549476837928677,
+ "maximum_prediction_kl_nats": 0.12420780785095903,
+ "top1_changed_fraction": 0.001993559270050606,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 1.074732780456543,
+ "maximum_per_target_cross_entropy_increase": 0.6844334602355957
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 68,
+ "storage": {
+ "artifact_bytes": 2500904,
+ "unique_rows": 36778,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "825e9083d4fb773e7477d19b942e81d71aa7941df0d31375fb16b12950b6837f",
+ "table_sha256": "7971ce6013cd10940ea0859b9f441072b9e4740e2e29e5921eacca0a1b7e229b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.05032146302983165,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2501752,
+ "backend_with_key_registry_bytes": 3099273,
+ "backend_with_datrie_key_index_bytes": 2796712,
+ "full_model_serving_artifact_bytes": 3799271,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999423853983437,
+ "mean_cosine": 0.9999882608898237,
+ "p05_cosine": 0.9999771263451054,
+ "relative_l2_mean": 0.0047384062543476555,
+ "relative_l2_p95": 0.006803712913917385,
+ "relative_l2_max": 0.010788400016504664,
+ "weighted_relative_l2_mean": 0.0030901968715466564,
+ "weighted_relative_frobenius": 0.005367400629039276,
+ "weighted_mean_cosine": 0.9999913956326278,
+ "max_relative_norm_change": 0.0030120548461313864
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.321759,
+ "warm_p50_ms": 0.096032,
+ "warm_p95_ms": 0.09845489999999998,
+ "warm_p99_ms": 0.11029019,
+ "samples_ms": [
+ 0.098406,
+ 0.097428,
+ 0.096521,
+ 0.095473,
+ 0.095962,
+ 0.09701,
+ 0.097428,
+ 0.096451,
+ 0.095054,
+ 0.096032,
+ 0.094635,
+ 0.095124,
+ 0.096102,
+ 0.094635,
+ 0.096032,
+ 0.099384,
+ 0.097498,
+ 0.095054,
+ 0.096102,
+ 0.095193,
+ 0.095124,
+ 0.095613,
+ 0.094565,
+ 0.095543,
+ 0.095473,
+ 0.096032,
+ 0.098337,
+ 0.096032,
+ 0.117263,
+ 0.095612,
+ 0.094705,
+ 0.097009,
+ 0.096032,
+ 0.096102,
+ 0.096101,
+ 0.096521,
+ 0.096521,
+ 0.095124,
+ 0.096032,
+ 0.095054
+ ],
+ "warm_rows_per_second": 5331556.147950683
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525349765882355,
+ "character_perplexity": 6.37596196791308,
+ "next_character_accuracy": 0.5016101824873486,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.7812517247741032e-05,
+ "mean_prediction_kl_nats": 3.922783085973939e-06,
+ "prediction_kl_p99_nats": 3.02223781367696e-05,
+ "maximum_prediction_kl_nats": 7.174838273647961e-05,
+ "top1_changed_fraction": 0.0012268057046465265,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.03138601779937744,
+ "maximum_per_target_cross_entropy_increase": 0.03577136993408203
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2802430,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 307,
+ "max_block_bytes": 8192,
+ "raw_blocks": 130,
+ "compressed_blocks": 177,
+ "zlib_blocks": 177,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "825e9083d4fb773e7477d19b942e81d71aa7941df0d31375fb16b12950b6837f",
+ "table_sha256": "c455ed4ebdc5a63b131ba6c6f986de04703fa0b5e64a378a46a5675bbbb32ab2",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24732666602358222,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2803683,
+ "backend_with_key_registry_bytes": 3401204,
+ "backend_with_datrie_key_index_bytes": 3098643,
+ "full_model_serving_artifact_bytes": 4101202,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999423853983437,
+ "mean_cosine": 0.9999882608898237,
+ "p05_cosine": 0.9999771263451054,
+ "relative_l2_mean": 0.0047384062543476555,
+ "relative_l2_p95": 0.006803712913917385,
+ "relative_l2_max": 0.010788400016504664,
+ "weighted_relative_l2_mean": 0.0030901968715466564,
+ "weighted_relative_frobenius": 0.005367400629039276,
+ "weighted_mean_cosine": 0.9999913956326278,
+ "max_relative_norm_change": 0.0030120548461313864
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.801046,
+ "warm_p50_ms": 0.114051,
+ "warm_p95_ms": 0.11917695,
+ "warm_p99_ms": 0.12750440999999998,
+ "samples_ms": [
+ 0.119149,
+ 0.116914,
+ 0.114121,
+ 0.113562,
+ 0.115448,
+ 0.116775,
+ 0.113562,
+ 0.115029,
+ 0.115028,
+ 0.113492,
+ 0.114539,
+ 0.114051,
+ 0.114051,
+ 0.113981,
+ 0.11447,
+ 0.113493,
+ 0.11454,
+ 0.113003,
+ 0.113981,
+ 0.113562,
+ 0.112025,
+ 0.114051,
+ 0.117823,
+ 0.113981,
+ 0.113981,
+ 0.112654,
+ 0.113562,
+ 0.113562,
+ 0.113562,
+ 0.112584,
+ 0.119708,
+ 0.114051,
+ 0.132489,
+ 0.11447,
+ 0.113981,
+ 0.113632,
+ 0.11447,
+ 0.114051,
+ 0.115378,
+ 0.112584
+ ],
+ "warm_rows_per_second": 4489219.73503082
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525349765882355,
+ "character_perplexity": 6.37596196791308,
+ "next_character_accuracy": 0.5016101824873486,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.7812517247741032e-05,
+ "mean_prediction_kl_nats": 3.922783085973939e-06,
+ "prediction_kl_p99_nats": 3.02223781367696e-05,
+ "maximum_prediction_kl_nats": 7.174838273647961e-05,
+ "top1_changed_fraction": 0.0012268057046465265,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.03138601779937744,
+ "maximum_per_target_cross_entropy_increase": 0.03577136993408203
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2802430,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 307,
+ "max_block_bytes": 8192,
+ "raw_blocks": 130,
+ "compressed_blocks": 177,
+ "zlib_blocks": 177,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "825e9083d4fb773e7477d19b942e81d71aa7941df0d31375fb16b12950b6837f",
+ "table_sha256": "c455ed4ebdc5a63b131ba6c6f986de04703fa0b5e64a378a46a5675bbbb32ab2",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.38247065991163254,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2803682,
+ "backend_with_key_registry_bytes": 3401203,
+ "backend_with_datrie_key_index_bytes": 3098642,
+ "full_model_serving_artifact_bytes": 4101201,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999423853983437,
+ "mean_cosine": 0.9999882608898237,
+ "p05_cosine": 0.9999771263451054,
+ "relative_l2_mean": 0.0047384062543476555,
+ "relative_l2_p95": 0.006803712913917385,
+ "relative_l2_max": 0.010788400016504664,
+ "weighted_relative_l2_mean": 0.0030901968715466564,
+ "weighted_relative_frobenius": 0.005367400629039276,
+ "weighted_mean_cosine": 0.9999913956326278,
+ "max_relative_norm_change": 0.0030120548461313864
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.826677,
+ "warm_p50_ms": 0.11443500000000001,
+ "warm_p95_ms": 0.12597964999999997,
+ "warm_p99_ms": 0.14579478999999998,
+ "samples_ms": [
+ 0.137658,
+ 0.116355,
+ 0.116426,
+ 0.113562,
+ 0.113562,
+ 0.115867,
+ 0.114959,
+ 0.114959,
+ 0.115937,
+ 0.11447,
+ 0.116356,
+ 0.115029,
+ 0.11454,
+ 0.113562,
+ 0.113562,
+ 0.113073,
+ 0.113562,
+ 0.113562,
+ 0.1144,
+ 0.113003,
+ 0.112514,
+ 0.113492,
+ 0.113563,
+ 0.113981,
+ 0.114051,
+ 0.112235,
+ 0.114959,
+ 0.113143,
+ 0.125365,
+ 0.116356,
+ 0.113562,
+ 0.113981,
+ 0.119708,
+ 0.11447,
+ 0.150997,
+ 0.11447,
+ 0.114959,
+ 0.113562,
+ 0.115378,
+ 0.113982
+ ],
+ "warm_rows_per_second": 4474155.634202822
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525349765882355,
+ "character_perplexity": 6.37596196791308,
+ "next_character_accuracy": 0.5016101824873486,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.7812517247741032e-05,
+ "mean_prediction_kl_nats": 3.922783085973939e-06,
+ "prediction_kl_p99_nats": 3.02223781367696e-05,
+ "maximum_prediction_kl_nats": 7.174838273647961e-05,
+ "top1_changed_fraction": 0.0012268057046465265,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.03138601779937744,
+ "maximum_per_target_cross_entropy_increase": 0.03577136993408203
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2802430,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 307,
+ "max_block_bytes": 8192,
+ "raw_blocks": 130,
+ "compressed_blocks": 177,
+ "zlib_blocks": 177,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "825e9083d4fb773e7477d19b942e81d71aa7941df0d31375fb16b12950b6837f",
+ "table_sha256": "c455ed4ebdc5a63b131ba6c6f986de04703fa0b5e64a378a46a5675bbbb32ab2",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2969288320746273,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2803682,
+ "backend_with_key_registry_bytes": 3401203,
+ "backend_with_datrie_key_index_bytes": 3098642,
+ "full_model_serving_artifact_bytes": 4101201,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999423853983437,
+ "mean_cosine": 0.9999882608898237,
+ "p05_cosine": 0.9999771263451054,
+ "relative_l2_mean": 0.0047384062543476555,
+ "relative_l2_p95": 0.006803712913917385,
+ "relative_l2_max": 0.010788400016504664,
+ "weighted_relative_l2_mean": 0.0030901968715466564,
+ "weighted_relative_frobenius": 0.005367400629039276,
+ "weighted_mean_cosine": 0.9999913956326278,
+ "max_relative_norm_change": 0.0030120548461313864
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.771224,
+ "warm_p50_ms": 0.114261,
+ "warm_p95_ms": 0.11787089999999999,
+ "warm_p99_ms": 0.13780088999999998,
+ "samples_ms": [
+ 0.1188,
+ 0.115937,
+ 0.114959,
+ 0.114401,
+ 0.114959,
+ 0.114958,
+ 0.11447,
+ 0.113562,
+ 0.11447,
+ 0.113911,
+ 0.114051,
+ 0.112724,
+ 0.113073,
+ 0.114121,
+ 0.114959,
+ 0.113493,
+ 0.114051,
+ 0.11447,
+ 0.11447,
+ 0.114539,
+ 0.113073,
+ 0.117822,
+ 0.113632,
+ 0.149949,
+ 0.11447,
+ 0.114051,
+ 0.115867,
+ 0.113562,
+ 0.11447,
+ 0.114401,
+ 0.113073,
+ 0.113143,
+ 0.114051,
+ 0.11454,
+ 0.113562,
+ 0.113492,
+ 0.113004,
+ 0.113003,
+ 0.114959,
+ 0.113563
+ ],
+ "warm_rows_per_second": 4480969.009548315
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525349765882355,
+ "character_perplexity": 6.37596196791308,
+ "next_character_accuracy": 0.5016101824873486,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.7812517247741032e-05,
+ "mean_prediction_kl_nats": 3.922783085973939e-06,
+ "prediction_kl_p99_nats": 3.02223781367696e-05,
+ "maximum_prediction_kl_nats": 7.174838273647961e-05,
+ "top1_changed_fraction": 0.0012268057046465265,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.03138601779937744,
+ "maximum_per_target_cross_entropy_increase": 0.03577136993408203
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2159553,
+ "unique_rows": 27372,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 229,
+ "max_block_bytes": 8192,
+ "raw_blocks": 58,
+ "compressed_blocks": 171,
+ "zlib_blocks": 171,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27372
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "bd98a4599f4a90cead3e90ca3b7aabff1968d2f9f4a281441f297c467992bbef",
+ "table_sha256": "489b9f9de01d8abe5f50495510ba00a43233e4069b495e950ff93e3c268d70e5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21825729892589152,
+ "representatives_sha256": "ee823010b22da96364634aa24fdb0332bd7e63e2a18483df332a03c9f0e4e861",
+ "structurally_merged_rows": 9406,
+ "backend_artifact_bytes": 2160803,
+ "backend_with_key_registry_bytes": 2758324,
+ "backend_with_datrie_key_index_bytes": 2455763,
+ "full_model_serving_artifact_bytes": 3458322,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9901936823909471,
+ "mean_cosine": 0.9997707398036456,
+ "p05_cosine": 0.9988471744883699,
+ "relative_l2_mean": 0.01370286698712146,
+ "relative_l2_p95": 0.04818096490309023,
+ "relative_l2_max": 0.13970731928662025,
+ "weighted_relative_l2_mean": 0.005253370717849569,
+ "weighted_relative_frobenius": 0.012475970273467448,
+ "weighted_mean_cosine": 0.9999434089752636,
+ "max_relative_norm_change": 0.01100592058332153
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900201496082667,
+ "mean_cosine": 0.999777006668141,
+ "p05_cosine": 0.9988339063736171,
+ "relative_l2_mean": 0.01018436070719559,
+ "relative_l2_p95": 0.04850662653422283,
+ "relative_l2_max": 0.14093234220233247,
+ "weighted_relative_l2_mean": 0.0025508588944308784,
+ "weighted_relative_frobenius": 0.011543854934588026,
+ "weighted_mean_cosine": 0.9999501476197958,
+ "max_relative_norm_change": 0.010088626688142067
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.299713,
+ "warm_p50_ms": 0.109057,
+ "warm_p95_ms": 0.11356884999999999,
+ "warm_p99_ms": 0.13028083,
+ "samples_ms": [
+ 0.113492,
+ 0.111676,
+ 0.109651,
+ 0.109791,
+ 0.109232,
+ 0.110699,
+ 0.10979,
+ 0.109232,
+ 0.110769,
+ 0.109302,
+ 0.11021,
+ 0.108882,
+ 0.108324,
+ 0.108394,
+ 0.108813,
+ 0.11028,
+ 0.108323,
+ 0.109791,
+ 0.108813,
+ 0.107835,
+ 0.109721,
+ 0.107835,
+ 0.107836,
+ 0.11021,
+ 0.11028,
+ 0.108813,
+ 0.108324,
+ 0.109791,
+ 0.108813,
+ 0.108324,
+ 0.108324,
+ 0.109791,
+ 0.107416,
+ 0.140032,
+ 0.115029,
+ 0.107416,
+ 0.108324,
+ 0.108254,
+ 0.108394,
+ 0.108743
+ ],
+ "warm_rows_per_second": 4694792.6313762525
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.852623154194073,
+ "character_perplexity": 6.376524209762526,
+ "next_character_accuracy": 0.5016101824873486,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00010599558157896495,
+ "mean_prediction_kl_nats": 1.0218319596541971e-05,
+ "prediction_kl_p99_nats": 0.00012016061647095144,
+ "maximum_prediction_kl_nats": 0.000566337115838136,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.23667526245117188,
+ "maximum_per_target_cross_entropy_increase": 0.06543993949890137
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 8.818149359684829e-05,
+ "mean_prediction_kl_nats": 6.195150129714561e-06,
+ "prediction_kl_p99_nats": 0.0001115058126726823,
+ "maximum_prediction_kl_nats": 0.0005962701972969846,
+ "top1_changed_fraction": 0.0006134028523232633,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.23351335525512695,
+ "maximum_per_target_cross_entropy_increase": 0.06547760963439941
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2158281,
+ "unique_rows": 27354,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 54,
+ "compressed_blocks": 174,
+ "zlib_blocks": 174,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27354
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "121b91d816b17acd3e8c8ac010f3e4a7ebaa64edf50d3f714101c59e309c0e86",
+ "table_sha256": "88d3e18197a954cafaa5e43f8509d538120223f6adb2098aec19f4ff18320483",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22241418599151075,
+ "representatives_sha256": "c55543f234ecf51aa35da166c68b43c7be495da8eef77faab73ad37db8ca1643",
+ "structurally_merged_rows": 9424,
+ "backend_artifact_bytes": 2159531,
+ "backend_with_key_registry_bytes": 2757052,
+ "backend_with_datrie_key_index_bytes": 2454491,
+ "full_model_serving_artifact_bytes": 3457050,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9803841974692542,
+ "mean_cosine": 0.99976350324125,
+ "p05_cosine": 0.9988431736660357,
+ "relative_l2_mean": 0.01378569874972075,
+ "relative_l2_p95": 0.04830936413209999,
+ "relative_l2_max": 0.19914099783926265,
+ "weighted_relative_l2_mean": 0.005332148767566714,
+ "weighted_relative_frobenius": 0.014300160532346706,
+ "weighted_mean_cosine": 0.9999364576676586,
+ "max_relative_norm_change": 0.01100592058332153
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.980402063783619,
+ "mean_cosine": 0.9997697652136252,
+ "p05_cosine": 0.9988288715392164,
+ "relative_l2_mean": 0.010269748863580546,
+ "relative_l2_p95": 0.048580357032820674,
+ "relative_l2_max": 0.1991238828124796,
+ "weighted_relative_l2_mean": 0.002632430675154758,
+ "weighted_relative_frobenius": 0.013510576247234417,
+ "weighted_mean_cosine": 0.999943160807779,
+ "max_relative_norm_change": 0.010088626688142067
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.15165,
+ "warm_p50_ms": 0.11045450000000001,
+ "warm_p95_ms": 0.11361089999999999,
+ "warm_p99_ms": 0.12399804999999998,
+ "samples_ms": [
+ 0.113562,
+ 0.112236,
+ 0.113562,
+ 0.111746,
+ 0.11014,
+ 0.11021,
+ 0.11028,
+ 0.110629,
+ 0.111117,
+ 0.111187,
+ 0.108813,
+ 0.110629,
+ 0.11028,
+ 0.10979,
+ 0.110698,
+ 0.130045,
+ 0.110768,
+ 0.110279,
+ 0.111746,
+ 0.111187,
+ 0.111118,
+ 0.110769,
+ 0.11454,
+ 0.110279,
+ 0.107905,
+ 0.110768,
+ 0.109302,
+ 0.109232,
+ 0.110769,
+ 0.108813,
+ 0.110209,
+ 0.107835,
+ 0.108882,
+ 0.110699,
+ 0.109721,
+ 0.109791,
+ 0.112165,
+ 0.110209,
+ 0.10979,
+ 0.109791
+ ],
+ "warm_rows_per_second": 4635392.854071133
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525341141760467,
+ "character_perplexity": 6.375956469208134,
+ "next_character_accuracy": 0.5018402085569698,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.6950090069191802e-05,
+ "mean_prediction_kl_nats": 2.6299175439423748e-05,
+ "prediction_kl_p99_nats": 0.00013393714038067466,
+ "maximum_prediction_kl_nats": 0.03291117008090324,
+ "top1_changed_fraction": 0.001993559270050606,
+ "accuracy_change": 7.667535654032331e-05,
+ "maximum_absolute_logit_change": 0.35282444953918457,
+ "maximum_per_target_cross_entropy_increase": 0.2796969413757324
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -8.624118168238675e-07,
+ "mean_prediction_kl_nats": 2.1991440496527458e-05,
+ "prediction_kl_p99_nats": 0.00011574609366350268,
+ "maximum_prediction_kl_nats": 0.0329851882669773,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 0.35393357276916504,
+ "maximum_per_target_cross_entropy_increase": 0.28632402420043945
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2157615,
+ "unique_rows": 27344,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 57,
+ "compressed_blocks": 171,
+ "zlib_blocks": 171,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27344
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "7c3e9a58c57a7130ed4bc6689968425ec271ab2704ee9b2a41e1f66c4e7717c5",
+ "table_sha256": "2feedd457e94ec3fb748a014d1832bf8ab40ce1343fe543ca6ba8db74bbbd404",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21872844803147018,
+ "representatives_sha256": "ed6c53830fcf065d41f20f69996586e097443ff3194041894a9d84046c3704e4",
+ "structurally_merged_rows": 9434,
+ "backend_artifact_bytes": 2158865,
+ "backend_with_key_registry_bytes": 2756386,
+ "backend_with_datrie_key_index_bytes": 2453825,
+ "full_model_serving_artifact_bytes": 3456384,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9709833361988419,
+ "mean_cosine": 0.9997537322133385,
+ "p05_cosine": 0.9988391669088803,
+ "relative_l2_mean": 0.013869420282903912,
+ "relative_l2_p95": 0.048450114957389066,
+ "relative_l2_max": 0.2415260799662704,
+ "weighted_relative_l2_mean": 0.005368595776662369,
+ "weighted_relative_frobenius": 0.015593640901344548,
+ "weighted_mean_cosine": 0.9999321681835488,
+ "max_relative_norm_change": 0.01100592058332153
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9709638081310338,
+ "mean_cosine": 0.999759985730253,
+ "p05_cosine": 0.9988257447836826,
+ "relative_l2_mean": 0.010354703942669713,
+ "relative_l2_p95": 0.04866831910058303,
+ "relative_l2_max": 0.24158341386608242,
+ "weighted_relative_l2_mean": 0.0026695630853409,
+ "weighted_relative_frobenius": 0.014868980379677798,
+ "weighted_mean_cosine": 0.9999388807722678,
+ "max_relative_norm_change": 0.010088626688142067
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.157237,
+ "warm_p50_ms": 0.11024500000000001,
+ "warm_p95_ms": 0.11379949999999998,
+ "warm_p99_ms": 0.1209106,
+ "samples_ms": [
+ 0.118312,
+ 0.113492,
+ 0.113562,
+ 0.111676,
+ 0.111187,
+ 0.111188,
+ 0.111187,
+ 0.10986,
+ 0.110628,
+ 0.111188,
+ 0.11028,
+ 0.110699,
+ 0.109721,
+ 0.110768,
+ 0.122572,
+ 0.111258,
+ 0.108743,
+ 0.110768,
+ 0.111117,
+ 0.109791,
+ 0.11021,
+ 0.111187,
+ 0.109301,
+ 0.112165,
+ 0.109232,
+ 0.109721,
+ 0.10979,
+ 0.109232,
+ 0.111258,
+ 0.108813,
+ 0.107905,
+ 0.10979,
+ 0.108882,
+ 0.109791,
+ 0.11014,
+ 0.110349,
+ 0.108743,
+ 0.109861,
+ 0.108324,
+ 0.108394
+ ],
+ "warm_rows_per_second": 4644201.551090752
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525229937301986,
+ "character_perplexity": 6.375885566123726,
+ "next_character_accuracy": 0.5018402085569698,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 5.829517561522718e-06,
+ "mean_prediction_kl_nats": 3.125478532769563e-05,
+ "prediction_kl_p99_nats": 0.00013693641873771206,
+ "maximum_prediction_kl_nats": 0.04956129816142537,
+ "top1_changed_fraction": 0.001993559270050606,
+ "accuracy_change": 7.667535654032331e-05,
+ "maximum_absolute_logit_change": 0.737525463104248,
+ "maximum_per_target_cross_entropy_increase": 0.2796969413757324
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -1.1982786242631691e-05,
+ "mean_prediction_kl_nats": 2.6956319617727833e-05,
+ "prediction_kl_p99_nats": 0.00012046420390634976,
+ "maximum_prediction_kl_nats": 0.049577409481285674,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 0.741180419921875,
+ "maximum_per_target_cross_entropy_increase": 0.28632402420043945
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2156785,
+ "unique_rows": 27332,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 53,
+ "compressed_blocks": 175,
+ "zlib_blocks": 175,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27332
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "ea8e9c923a52ba3548854b19f21826cba3cfc333f25192406250623d413a6dc5",
+ "table_sha256": "93986663112648d6711b10d225f692e834dfe6ebcea5624bfc857368f33a39ad",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2166988591197878,
+ "representatives_sha256": "bdfc2ff5485bc8159dfea7f07b84d9fe9fea0769fdc0dc012e2f75b5d5e28045",
+ "structurally_merged_rows": 9446,
+ "backend_artifact_bytes": 2158035,
+ "backend_with_key_registry_bytes": 2755556,
+ "backend_with_datrie_key_index_bytes": 2452995,
+ "full_model_serving_artifact_bytes": 3455554,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9511665284852492,
+ "mean_cosine": 0.9997352426815083,
+ "p05_cosine": 0.9988325331385304,
+ "relative_l2_mean": 0.013997645638866652,
+ "relative_l2_p95": 0.04860716272042506,
+ "relative_l2_max": 0.31180209459341773,
+ "weighted_relative_l2_mean": 0.00548975913955002,
+ "weighted_relative_frobenius": 0.01916368741803261,
+ "weighted_mean_cosine": 0.9999153705346492,
+ "max_relative_norm_change": 0.01100592058332153
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9508873098418595,
+ "mean_cosine": 0.9997414245304309,
+ "p05_cosine": 0.9988216621684859,
+ "relative_l2_mean": 0.010484797159271605,
+ "relative_l2_p95": 0.04876521079590782,
+ "relative_l2_max": 0.3126990872022993,
+ "weighted_relative_l2_mean": 0.002793320906186573,
+ "weighted_relative_frobenius": 0.018629382528077235,
+ "weighted_mean_cosine": 0.9999218470723966,
+ "max_relative_norm_change": 0.010088626688142067
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.028589,
+ "warm_p50_ms": 0.11,
+ "warm_p95_ms": 0.11365629999999999,
+ "warm_p99_ms": 0.13219127999999997,
+ "samples_ms": [
+ 0.115448,
+ 0.111118,
+ 0.112585,
+ 0.111188,
+ 0.109791,
+ 0.113073,
+ 0.111118,
+ 0.110698,
+ 0.110279,
+ 0.111117,
+ 0.11021,
+ 0.11035,
+ 0.108324,
+ 0.108813,
+ 0.109791,
+ 0.111606,
+ 0.111746,
+ 0.110209,
+ 0.109721,
+ 0.108813,
+ 0.108812,
+ 0.109302,
+ 0.109721,
+ 0.108813,
+ 0.108813,
+ 0.10979,
+ 0.109232,
+ 0.109372,
+ 0.109302,
+ 0.110699,
+ 0.109791,
+ 0.110629,
+ 0.10979,
+ 0.109721,
+ 0.142896,
+ 0.113562,
+ 0.11035,
+ 0.111607,
+ 0.108394,
+ 0.108813
+ ],
+ "warm_rows_per_second": 4654545.454545454
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8521772767588605,
+ "character_perplexity": 6.3736816952559385,
+ "next_character_accuracy": 0.5020702346265911,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0003398297155654553,
+ "mean_prediction_kl_nats": 0.00013342371873271348,
+ "prediction_kl_p99_nats": 0.00014980814582011523,
+ "maximum_prediction_kl_nats": 0.12502342394746385,
+ "top1_changed_fraction": 0.002376936052752645,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.8871172070503235,
+ "maximum_per_target_cross_entropy_increase": 0.2796969413757324
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.0003576358624182286,
+ "mean_prediction_kl_nats": 0.00012999157531167803,
+ "prediction_kl_p99_nats": 0.00014307763126692507,
+ "maximum_prediction_kl_nats": 0.12617057049907654,
+ "top1_changed_fraction": 0.0013034810611869344,
+ "accuracy_change": 0.000460052139242495,
+ "maximum_absolute_logit_change": 0.8878851532936096,
+ "maximum_per_target_cross_entropy_increase": 0.28632402420043945
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2155118,
+ "unique_rows": 27307,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 52,
+ "compressed_blocks": 176,
+ "zlib_blocks": 176,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27307
+ },
+ "source_float32_sha256": "23fef795d559ef2b26b5103ab065e1c40248974ddd5afae0b8d0d0e33b95942a",
+ "decoded_float32_sha256": "56eea6ba9d9e03d7f3c3f40ee73d05807f60a8149404c5e42586e401ba51e21b",
+ "table_sha256": "3238854efa156d42be7bdcb48de9dccedcb24f755e8e3c2ce154d71cbdd17de1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21731388196349144,
+ "representatives_sha256": "ae1c126285926ec3a37278faaedcc33c201df3c5bea4cbfac3adbc753b12b51c",
+ "structurally_merged_rows": 9471,
+ "backend_artifact_bytes": 2156368,
+ "backend_with_key_registry_bytes": 2753889,
+ "backend_with_datrie_key_index_bytes": 2451328,
+ "full_model_serving_artifact_bytes": 3453887,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9002791692838276,
+ "mean_cosine": 0.9996697757800215,
+ "p05_cosine": 0.9988203220304905,
+ "relative_l2_mean": 0.014320318924229457,
+ "relative_l2_p95": 0.04885864608503015,
+ "relative_l2_max": 0.445566953706329,
+ "weighted_relative_l2_mean": 0.005858387832668077,
+ "weighted_relative_frobenius": 0.024652605165551348,
+ "weighted_mean_cosine": 0.9998401080404868,
+ "max_relative_norm_change": 0.0113042873169545
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9007489979693729,
+ "mean_cosine": 0.9996758896490945,
+ "p05_cosine": 0.9988132913388617,
+ "relative_l2_mean": 0.010810970940916554,
+ "relative_l2_p95": 0.048995893370770364,
+ "relative_l2_max": 0.44674244837039334,
+ "weighted_relative_l2_mean": 0.0031668512815220123,
+ "weighted_relative_frobenius": 0.024233944025561935,
+ "weighted_mean_cosine": 0.9998466319447397,
+ "max_relative_norm_change": 0.010088626688142067
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.211713,
+ "warm_p50_ms": 0.1097205,
+ "warm_p95_ms": 0.1126783,
+ "warm_p99_ms": 0.13611279999999998,
+ "samples_ms": [
+ 0.112584,
+ 0.110699,
+ 0.111188,
+ 0.111257,
+ 0.111258,
+ 0.14995,
+ 0.110769,
+ 0.110699,
+ 0.11021,
+ 0.111187,
+ 0.10986,
+ 0.112165,
+ 0.108323,
+ 0.11447,
+ 0.108882,
+ 0.109302,
+ 0.11021,
+ 0.110769,
+ 0.109301,
+ 0.109232,
+ 0.109791,
+ 0.108812,
+ 0.107905,
+ 0.109301,
+ 0.109302,
+ 0.110629,
+ 0.109791,
+ 0.110768,
+ 0.108813,
+ 0.108813,
+ 0.107835,
+ 0.107835,
+ 0.108324,
+ 0.109302,
+ 0.109651,
+ 0.108324,
+ 0.108813,
+ 0.10979,
+ 0.109232,
+ 0.109302
+ ],
+ "warm_rows_per_second": 4666402.358720567
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.852207039291068,
+ "character_perplexity": 6.373871394985632,
+ "next_character_accuracy": 0.5019935592700506,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0003100768547929411,
+ "mean_prediction_kl_nats": 0.00023291380840749926,
+ "prediction_kl_p99_nats": 0.00037058985748242746,
+ "maximum_prediction_kl_nats": 0.12502342394746385,
+ "top1_changed_fraction": 0.002836988191995093,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 1.0748310089111328,
+ "maximum_per_target_cross_entropy_increase": 0.6846189498901367
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.00032788353160962004,
+ "mean_prediction_kl_nats": 0.00022920314508750436,
+ "prediction_kl_p99_nats": 0.0003665996871860343,
+ "maximum_prediction_kl_nats": 0.12617057049907654,
+ "top1_changed_fraction": 0.001763533200429382,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 1.0692188739776611,
+ "maximum_per_target_cross_entropy_increase": 0.6830034255981445
+ }
+ }
+ ],
+ "baseline_quality": {
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8525171642296288,
+ "character_perplexity": 6.375848398003521,
+ "next_character_accuracy": 0.5017635332004294,
+ "unknown_targets": 0
+ },
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7264914504842404,
+ "character_perplexity": 5.620898072320001,
+ "next_character_accuracy": 0.5251092540059802,
+ "unknown_targets": 0
+ }
+ },
+ "selection": {
+ "fp32": {
+ "feasible": true,
+ "name": "fp32-joint-0.97",
+ "backend_with_key_registry_bytes": 7425813,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7263920138538618,
+ "character_perplexity": 5.620339176943785,
+ "next_character_accuracy": 0.5251859234838611,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -9.943168672067081e-05,
+ "mean_prediction_kl_nats": 3.599514161387495e-05,
+ "prediction_kl_p99_nats": 0.00012654597219580023,
+ "maximum_prediction_kl_nats": 0.04607604998831831,
+ "top1_changed_fraction": 0.0013033811239745456,
+ "accuracy_change": 7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.7387614250183105,
+ "maximum_per_target_cross_entropy_increase": 0.36521947383880615
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00019782683373067314,
+ 4.029123812401072e-06
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "fp16": {
+ "feasible": true,
+ "name": "fp16-joint-0.97",
+ "backend_with_key_registry_bytes": 4134227,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.726390375275437,
+ "character_perplexity": 5.620329967584816,
+ "next_character_accuracy": 0.5251859234838611,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -0.00010107010087641331,
+ "mean_prediction_kl_nats": 3.6004410187967044e-05,
+ "prediction_kl_p99_nats": 0.00012695645761842848,
+ "maximum_prediction_kl_nats": 0.04610446029724359,
+ "top1_changed_fraction": 0.0013033811239745456,
+ "accuracy_change": 7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.739013671875,
+ "maximum_per_target_cross_entropy_increase": 0.36529219150543213
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0001983225788790956,
+ 1.5536949280217424e-06
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "int8": {
+ "feasible": true,
+ "name": "int8-joint-0.97",
+ "backend_with_key_registry_bytes": 2756386,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7263646895656914,
+ "character_perplexity": 5.6201856072746,
+ "next_character_accuracy": 0.5252625929617419,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -0.00012675288472309662,
+ "mean_prediction_kl_nats": 3.98149611035077e-05,
+ "prediction_kl_p99_nats": 0.0001327648061789036,
+ "maximum_prediction_kl_nats": 0.045853842864706854,
+ "top1_changed_fraction": 0.0018400674691405351,
+ "accuracy_change": 0.00015333895576175038,
+ "maximum_absolute_logit_change": 0.7336316108703613,
+ "maximum_per_target_cross_entropy_increase": 0.36336004734039307
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0002373574082006347,
+ -1.4517555775323528e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ }
+ },
+ "fixed_quantization_test_controls": {
+ "fp16": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7264897442538625,
+ "character_perplexity": 5.620888481781141,
+ "next_character_accuracy": 0.5251092540059802,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -1.70622892227715e-06,
+ "mean_prediction_kl_nats": 3.9210284471012125e-09,
+ "prediction_kl_p99_nats": 2.915575772916698e-08,
+ "maximum_prediction_kl_nats": 9.294681676850027e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0011017322540283203,
+ "maximum_per_target_cross_entropy_increase": 0.0009131431579589844
+ }
+ },
+ "int8": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.726475467567206,
+ "character_perplexity": 5.620808234690389,
+ "next_character_accuracy": 0.5250325845280993,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -1.5982789308212916e-05,
+ "mean_prediction_kl_nats": 3.616042494524373e-06,
+ "prediction_kl_p99_nats": 2.9782866738580935e-05,
+ "maximum_prediction_kl_nats": 9.914208261234641e-05,
+ "top1_changed_fraction": 0.0005366863451659894,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.031708717346191406,
+ "maximum_per_target_cross_entropy_increase": 0.021615982055664062
+ }
+ }
+ },
+ "checkpoint_sha256": "f773d277a5444631a7ca1865880469f92d01fd664f4da9363621026ccd307845",
+ "registry_sha256": "e22a7692100a89679e3088226c3399678ae46a06d73d93bef7d898aa00f6fb7c"
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 11,
+ "rows": 36778,
+ "dim": 64,
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "source_float32_bytes": 9415168,
+ "key_registry_bytes": 597521,
+ "unchanged_backbone_artifact_bytes": 699998,
+ "alternative_datrie_key_index_bytes": 294960,
+ "query_rows": 512,
+ "query_ids_sha256": "c0a9c7f836d519eda638b9a28093508b2d62873b170686d55b9ec727411573db",
+ "frequency_trace_sha256": "efa6b79677263433708f5d18746f076a51a377d034e599be829419ac6115bd49",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 9415168,
+ "unique_rows": 36778,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "table_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.04457491612993181,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9415972,
+ "backend_with_key_registry_bytes": 10013493,
+ "backend_with_datrie_key_index_bytes": 9710932,
+ "full_model_serving_artifact_bytes": 10713491,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.31226,
+ "warm_p50_ms": 0.047771,
+ "warm_p95_ms": 0.0491715,
+ "warm_p99_ms": 0.049834579999999996,
+ "samples_ms": [
+ 0.049238,
+ 0.04833,
+ 0.048749,
+ 0.04875,
+ 0.047283,
+ 0.047213,
+ 0.046793,
+ 0.04826,
+ 0.046863,
+ 0.046794,
+ 0.048191,
+ 0.046794,
+ 0.04826,
+ 0.04826,
+ 0.045467,
+ 0.04826,
+ 0.04826,
+ 0.047771,
+ 0.047282,
+ 0.048261,
+ 0.047353,
+ 0.048331,
+ 0.047283,
+ 0.047213,
+ 0.045886,
+ 0.046793,
+ 0.047702,
+ 0.047772,
+ 0.047283,
+ 0.046305,
+ 0.047842,
+ 0.04875,
+ 0.047772,
+ 0.050216,
+ 0.046863,
+ 0.047771,
+ 0.049168,
+ 0.047353,
+ 0.047283,
+ 0.047772
+ ],
+ "warm_rows_per_second": 10717799.501789788
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853825927869258,
+ "character_perplexity": 6.384198339416614,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 9081798,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1150,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1150,
+ "zlib_blocks": 1150,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "table_sha256": "960fa1563e78af85b221076d89eb9f0ecd65b332bd4ee2659401acca84f40fbb",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.7745489759836346,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9083008,
+ "backend_with_key_registry_bytes": 9680529,
+ "backend_with_datrie_key_index_bytes": 9377968,
+ "full_model_serving_artifact_bytes": 10380527,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.499553,
+ "warm_p50_ms": 0.056292,
+ "warm_p95_ms": 0.060066549999999996,
+ "warm_p99_ms": 0.07278662,
+ "samples_ms": [
+ 0.060134,
+ 0.058807,
+ 0.05727,
+ 0.058178,
+ 0.055873,
+ 0.056712,
+ 0.055803,
+ 0.055803,
+ 0.056222,
+ 0.056292,
+ 0.054965,
+ 0.056851,
+ 0.05727,
+ 0.059644,
+ 0.058666,
+ 0.057759,
+ 0.055803,
+ 0.059156,
+ 0.058178,
+ 0.055803,
+ 0.056292,
+ 0.054406,
+ 0.056292,
+ 0.055804,
+ 0.055804,
+ 0.055803,
+ 0.055803,
+ 0.056781,
+ 0.055873,
+ 0.056292,
+ 0.055384,
+ 0.055454,
+ 0.056781,
+ 0.056292,
+ 0.057689,
+ 0.056711,
+ 0.055873,
+ 0.060063,
+ 0.080876,
+ 0.058178
+ ],
+ "warm_rows_per_second": 9095430.967100121
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853825927869258,
+ "character_perplexity": 6.384198339416614,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 9081798,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1150,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1150,
+ "zlib_blocks": 1150,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "table_sha256": "960fa1563e78af85b221076d89eb9f0ecd65b332bd4ee2659401acca84f40fbb",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.9328933600336313,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9083007,
+ "backend_with_key_registry_bytes": 9680528,
+ "backend_with_datrie_key_index_bytes": 9377967,
+ "full_model_serving_artifact_bytes": 10380526,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.521344,
+ "warm_p50_ms": 0.056292,
+ "warm_p95_ms": 0.05999669999999998,
+ "warm_p99_ms": 0.09123768999999997,
+ "samples_ms": [
+ 0.059644,
+ 0.059086,
+ 0.059225,
+ 0.056292,
+ 0.056362,
+ 0.056362,
+ 0.056781,
+ 0.055873,
+ 0.056292,
+ 0.058597,
+ 0.056781,
+ 0.056781,
+ 0.056292,
+ 0.106927,
+ 0.066698,
+ 0.057689,
+ 0.056223,
+ 0.056292,
+ 0.055873,
+ 0.054476,
+ 0.056781,
+ 0.057689,
+ 0.056781,
+ 0.056292,
+ 0.056292,
+ 0.054756,
+ 0.055314,
+ 0.054895,
+ 0.054755,
+ 0.054825,
+ 0.055314,
+ 0.055803,
+ 0.055804,
+ 0.05727,
+ 0.056781,
+ 0.055873,
+ 0.056851,
+ 0.056781,
+ 0.055803,
+ 0.056292
+ ],
+ "warm_rows_per_second": 9095430.967100121
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853825927869258,
+ "character_perplexity": 6.384198339416614,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 9081798,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1150,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1150,
+ "zlib_blocks": 1150,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "table_sha256": "960fa1563e78af85b221076d89eb9f0ecd65b332bd4ee2659401acca84f40fbb",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.791637617861852,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 9083007,
+ "backend_with_key_registry_bytes": 9680528,
+ "backend_with_datrie_key_index_bytes": 9377967,
+ "full_model_serving_artifact_bytes": 10380526,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.503324,
+ "warm_p50_ms": 0.0572,
+ "warm_p95_ms": 0.0601584,
+ "warm_p99_ms": 0.08281867999999998,
+ "samples_ms": [
+ 0.060064,
+ 0.057689,
+ 0.058597,
+ 0.056362,
+ 0.056781,
+ 0.055803,
+ 0.060622,
+ 0.058177,
+ 0.05727,
+ 0.057828,
+ 0.05727,
+ 0.057269,
+ 0.0572,
+ 0.056711,
+ 0.055804,
+ 0.056851,
+ 0.0572,
+ 0.057759,
+ 0.057759,
+ 0.055803,
+ 0.055873,
+ 0.055873,
+ 0.055803,
+ 0.09701,
+ 0.058178,
+ 0.0572,
+ 0.060063,
+ 0.056292,
+ 0.058178,
+ 0.055873,
+ 0.055314,
+ 0.058248,
+ 0.060134,
+ 0.060064,
+ 0.058178,
+ 0.055803,
+ 0.056292,
+ 0.056293,
+ 0.055873,
+ 0.056851
+ ],
+ "warm_rows_per_second": 8951048.951048952
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853825927869258,
+ "character_perplexity": 6.384198339416614,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6839176,
+ "unique_rows": 27379,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 856,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 856,
+ "zlib_blocks": 856,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27379
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "7b499bfa05f48b7e1c31f10149e4d29e810350961ea1f67e55e4d8542b296adc",
+ "table_sha256": "666c1328a1d134210be3c20b41c22616a8ee49866e8e7decdd3981323cd472d1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6203860640525818,
+ "representatives_sha256": "d2c58fe5024e67300a190b5313bddd70ad42ec3afe185e3362dbb3ce52f03d93",
+ "structurally_merged_rows": 9399,
+ "backend_artifact_bytes": 6840382,
+ "backend_with_key_registry_bytes": 7437903,
+ "backend_with_datrie_key_index_bytes": 7135342,
+ "full_model_serving_artifact_bytes": 8137901,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900606923999148,
+ "mean_cosine": 0.9997784576355798,
+ "p05_cosine": 0.998848116220339,
+ "relative_l2_mean": 0.01010489966444938,
+ "relative_l2_p95": 0.048224850651301776,
+ "relative_l2_max": 0.14115591128178223,
+ "weighted_relative_l2_mean": 0.002516309229029448,
+ "weighted_relative_frobenius": 0.011360509821225311,
+ "weighted_mean_cosine": 0.9999512639366307,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900606923999148,
+ "mean_cosine": 0.9997784576355798,
+ "p05_cosine": 0.998848116220339,
+ "relative_l2_mean": 0.01010489966444938,
+ "relative_l2_p95": 0.048224850651301776,
+ "relative_l2_max": 0.14115591128178223,
+ "weighted_relative_l2_mean": 0.002516309229029448,
+ "weighted_relative_frobenius": 0.011360509821225311,
+ "weighted_mean_cosine": 0.9999512639366307,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.131,
+ "warm_p50_ms": 0.056292,
+ "warm_p95_ms": 0.06111434999999999,
+ "warm_p99_ms": 0.07000611999999999,
+ "samples_ms": [
+ 0.062508,
+ 0.061041,
+ 0.059156,
+ 0.059156,
+ 0.0572,
+ 0.056781,
+ 0.058108,
+ 0.056292,
+ 0.055803,
+ 0.056711,
+ 0.056223,
+ 0.056292,
+ 0.055803,
+ 0.055314,
+ 0.056362,
+ 0.056292,
+ 0.055873,
+ 0.0748,
+ 0.056781,
+ 0.054895,
+ 0.055314,
+ 0.055873,
+ 0.056292,
+ 0.055874,
+ 0.055873,
+ 0.055873,
+ 0.056781,
+ 0.056293,
+ 0.056292,
+ 0.055314,
+ 0.056292,
+ 0.056781,
+ 0.056292,
+ 0.055873,
+ 0.058667,
+ 0.059644,
+ 0.058178,
+ 0.056781,
+ 0.056292,
+ 0.056362
+ ],
+ "warm_rows_per_second": 9095430.967100121
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8539171709741622,
+ "character_perplexity": 6.384780880071433,
+ "next_character_accuracy": 0.502453611409293,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 9.124726768285843e-05,
+ "mean_prediction_kl_nats": 7.857263138507182e-06,
+ "prediction_kl_p99_nats": 0.00010876512622539034,
+ "maximum_prediction_kl_nats": 0.0054869359999785315,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.5825309753417969,
+ "maximum_per_target_cross_entropy_increase": 0.10682880878448486
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 9.124726768285843e-05,
+ "mean_prediction_kl_nats": 7.857263138507182e-06,
+ "prediction_kl_p99_nats": 0.00010876512622539034,
+ "maximum_prediction_kl_nats": 0.0054869359999785315,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.5825309753417969,
+ "maximum_per_target_cross_entropy_increase": 0.10682880878448486
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6836039,
+ "unique_rows": 27366,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 856,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 856,
+ "zlib_blocks": 856,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27366
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "0b61fee666a805d5857209965801405cb41299b40a5bd43f939b298d1e2be6d1",
+ "table_sha256": "63c82cd7343837b95fd33d1a151079fd56a99eb850266b5aa9b6bde7879075f9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6236524749547243,
+ "representatives_sha256": "11886570fc035d70d87f25334b52f1d14537fddee3ed7a4239958fb4d236f4a7",
+ "structurally_merged_rows": 9412,
+ "backend_artifact_bytes": 6837245,
+ "backend_with_key_registry_bytes": 7434766,
+ "backend_with_datrie_key_index_bytes": 7132205,
+ "full_model_serving_artifact_bytes": 8134764,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9809576293463765,
+ "mean_cosine": 0.9997725672015289,
+ "p05_cosine": 0.9988457547530647,
+ "relative_l2_mean": 0.010172216018393514,
+ "relative_l2_p95": 0.04829450850377814,
+ "relative_l2_max": 0.1950679758979058,
+ "weighted_relative_l2_mean": 0.0025796313919513427,
+ "weighted_relative_frobenius": 0.012457324903073863,
+ "weighted_mean_cosine": 0.9999459192235313,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9809576293463765,
+ "mean_cosine": 0.9997725672015289,
+ "p05_cosine": 0.9988457547530647,
+ "relative_l2_mean": 0.010172216018393514,
+ "relative_l2_p95": 0.04829450850377814,
+ "relative_l2_max": 0.1950679758979058,
+ "weighted_relative_l2_mean": 0.0025796313919513427,
+ "weighted_relative_frobenius": 0.012457324903073863,
+ "weighted_mean_cosine": 0.9999459192235313,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.846187,
+ "warm_p50_ms": 0.055873,
+ "warm_p95_ms": 0.0606649,
+ "warm_p99_ms": 0.06640261,
+ "samples_ms": [
+ 0.061461,
+ 0.060623,
+ 0.059575,
+ 0.060622,
+ 0.058177,
+ 0.059156,
+ 0.058667,
+ 0.0572,
+ 0.056292,
+ 0.055384,
+ 0.069562,
+ 0.058247,
+ 0.0572,
+ 0.057201,
+ 0.055384,
+ 0.055384,
+ 0.055384,
+ 0.055384,
+ 0.054965,
+ 0.055873,
+ 0.055384,
+ 0.055873,
+ 0.055384,
+ 0.057689,
+ 0.057689,
+ 0.056222,
+ 0.055873,
+ 0.056851,
+ 0.055384,
+ 0.055873,
+ 0.055384,
+ 0.055384,
+ 0.055873,
+ 0.055315,
+ 0.055384,
+ 0.055803,
+ 0.055803,
+ 0.054406,
+ 0.055873,
+ 0.054476
+ ],
+ "warm_rows_per_second": 9163638.966942888
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853967702354528,
+ "character_perplexity": 6.3851035200142885,
+ "next_character_accuracy": 0.502453611409293,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.000141784535747201,
+ "mean_prediction_kl_nats": 1.1384415248378461e-05,
+ "prediction_kl_p99_nats": 0.00012439289968876123,
+ "maximum_prediction_kl_nats": 0.00827237447897241,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.5825309753417969,
+ "maximum_per_target_cross_entropy_increase": 0.5608301162719727
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.000141784535747201,
+ "mean_prediction_kl_nats": 1.1384415248378461e-05,
+ "prediction_kl_p99_nats": 0.00012439289968876123,
+ "maximum_prediction_kl_nats": 0.00827237447897241,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.5825309753417969,
+ "maximum_per_target_cross_entropy_increase": 0.5608301162719727
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6833584,
+ "unique_rows": 27356,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 855,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 855,
+ "zlib_blocks": 855,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27356
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "340970ca40a2cbc3bb86a8ec96c731b51e67bc4586f55202758eebaa23a3516f",
+ "table_sha256": "584b01d501784e587d59853bd4ef104247034ea52bb5ab199301e5fadfa9e8f6",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6324292970821261,
+ "representatives_sha256": "37d324b2c4872e3e4580f040ba77d1bb6d678c92d38bfa9a8c0c660a6bc22abf",
+ "structurally_merged_rows": 9422,
+ "backend_artifact_bytes": 6834790,
+ "backend_with_key_registry_bytes": 7432311,
+ "backend_with_datrie_key_index_bytes": 7129750,
+ "full_model_serving_artifact_bytes": 8132309,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9704402461322955,
+ "mean_cosine": 0.9997651021883833,
+ "p05_cosine": 0.9988425531711095,
+ "relative_l2_mean": 0.010238032437592758,
+ "relative_l2_p95": 0.048383933605566955,
+ "relative_l2_max": 0.24438774093788204,
+ "weighted_relative_l2_mean": 0.002622011212296494,
+ "weighted_relative_frobenius": 0.013363174052846639,
+ "weighted_mean_cosine": 0.9999410905985338,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9704402461322955,
+ "mean_cosine": 0.9997651021883833,
+ "p05_cosine": 0.9988425531711095,
+ "relative_l2_mean": 0.010238032437592758,
+ "relative_l2_p95": 0.048383933605566955,
+ "relative_l2_max": 0.24438774093788204,
+ "weighted_relative_l2_mean": 0.002622011212296494,
+ "weighted_relative_frobenius": 0.013363174052846639,
+ "weighted_mean_cosine": 0.9999410905985338,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.973927,
+ "warm_p50_ms": 0.056292,
+ "warm_p95_ms": 0.059645,
+ "warm_p99_ms": 0.07809200999999998,
+ "samples_ms": [
+ 0.058667,
+ 0.059645,
+ 0.056781,
+ 0.056712,
+ 0.0572,
+ 0.057689,
+ 0.056292,
+ 0.054895,
+ 0.089886,
+ 0.056293,
+ 0.056292,
+ 0.056781,
+ 0.056292,
+ 0.055873,
+ 0.055873,
+ 0.056222,
+ 0.056712,
+ 0.056292,
+ 0.056292,
+ 0.056362,
+ 0.055384,
+ 0.055315,
+ 0.054896,
+ 0.056292,
+ 0.055803,
+ 0.054895,
+ 0.055803,
+ 0.056292,
+ 0.055803,
+ 0.056292,
+ 0.056781,
+ 0.055873,
+ 0.055384,
+ 0.055873,
+ 0.054965,
+ 0.056362,
+ 0.053987,
+ 0.056712,
+ 0.059645,
+ 0.056781
+ ],
+ "warm_rows_per_second": 9095430.967100121
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8540159733975723,
+ "character_perplexity": 6.385411743060209,
+ "next_character_accuracy": 0.5025302867658334,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00019006358810967617,
+ "mean_prediction_kl_nats": 3.3202891648341754e-05,
+ "prediction_kl_p99_nats": 0.00013455703572435869,
+ "maximum_prediction_kl_nats": 0.05639065703611837,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 0.5825309753417969,
+ "maximum_per_target_cross_entropy_increase": 0.5608301162719727
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00019006358810967617,
+ "mean_prediction_kl_nats": 3.3202891648341754e-05,
+ "prediction_kl_p99_nats": 0.00013455703572435869,
+ "maximum_prediction_kl_nats": 0.05639065703611837,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 0.5825309753417969,
+ "maximum_per_target_cross_entropy_increase": 0.5608301162719727
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6830054,
+ "unique_rows": 27341,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 855,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 855,
+ "zlib_blocks": 855,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27341
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "9eecfa9f246dfb68f0fd6ae8c5f3beeb4d8f7def790e5936e186c05033587ef8",
+ "table_sha256": "5774c63ec3757799dddb912010bd97b9da8869dbe730b1070b54751e8a856b04",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6180408620275557,
+ "representatives_sha256": "0899384c5982cc54601899a082a05066a00c9ba276bd16ed3eef6f38e8c3dea3",
+ "structurally_merged_rows": 9437,
+ "backend_artifact_bytes": 6831260,
+ "backend_with_key_registry_bytes": 7428781,
+ "backend_with_datrie_key_index_bytes": 7126220,
+ "full_model_serving_artifact_bytes": 8128779,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9518773821592306,
+ "mean_cosine": 0.9997393157385929,
+ "p05_cosine": 0.9988358590976054,
+ "relative_l2_mean": 0.010413511527011282,
+ "relative_l2_p95": 0.04856855841130637,
+ "relative_l2_max": 0.3090281189559274,
+ "weighted_relative_l2_mean": 0.0028881504015720893,
+ "weighted_relative_frobenius": 0.02021922611928201,
+ "weighted_mean_cosine": 0.9998993340489422,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9518773821592306,
+ "mean_cosine": 0.9997393157385929,
+ "p05_cosine": 0.9988358590976054,
+ "relative_l2_mean": 0.010413511527011282,
+ "relative_l2_p95": 0.04856855841130637,
+ "relative_l2_max": 0.3090281189559274,
+ "weighted_relative_l2_mean": 0.0028881504015720893,
+ "weighted_relative_frobenius": 0.02021922611928201,
+ "weighted_mean_cosine": 0.9998993340489422,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.944174,
+ "warm_p50_ms": 0.055803,
+ "warm_p95_ms": 0.05824785,
+ "warm_p99_ms": 0.05991538,
+ "samples_ms": [
+ 0.058178,
+ 0.059575,
+ 0.060133,
+ 0.057759,
+ 0.057689,
+ 0.055873,
+ 0.057759,
+ 0.057689,
+ 0.057689,
+ 0.057689,
+ 0.058178,
+ 0.058178,
+ 0.05727,
+ 0.055803,
+ 0.055314,
+ 0.055803,
+ 0.055384,
+ 0.057689,
+ 0.056222,
+ 0.055384,
+ 0.054406,
+ 0.055803,
+ 0.055314,
+ 0.055803,
+ 0.055734,
+ 0.056292,
+ 0.055384,
+ 0.055803,
+ 0.055314,
+ 0.055454,
+ 0.055873,
+ 0.055384,
+ 0.055454,
+ 0.055873,
+ 0.055384,
+ 0.056851,
+ 0.055385,
+ 0.055803,
+ 0.055803,
+ 0.054895
+ ],
+ "warm_rows_per_second": 9175133.953371683
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853981340930912,
+ "character_perplexity": 6.385190604330218,
+ "next_character_accuracy": 0.5025302867658334,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00015542513888953914,
+ "mean_prediction_kl_nats": 3.9837892344174354e-05,
+ "prediction_kl_p99_nats": 0.00020867086076595511,
+ "maximum_prediction_kl_nats": 0.05639065703611837,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 1.3523645401000977,
+ "maximum_per_target_cross_entropy_increase": 0.592867374420166
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00015542513888953914,
+ "mean_prediction_kl_nats": 3.9837892344174354e-05,
+ "prediction_kl_p99_nats": 0.00020867086076595511,
+ "maximum_prediction_kl_nats": 0.05639065703611837,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 1.3523645401000977,
+ "maximum_per_target_cross_entropy_increase": 0.592867374420166
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 9415168,
+ "artifact_bytes": 6823416,
+ "unique_rows": 27313,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 854,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 854,
+ "zlib_blocks": 854,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27313
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "54b9125f3f3cc8048738b1898b724c2aa7603d1fffe0fd57fcb85ab8242ebe55",
+ "table_sha256": "12a0e3b6a62dfb6b77ad4b37cac8fa834ddca6e66f29758fcc19a46bef85d36b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6208194300998002,
+ "representatives_sha256": "53a34e444549f34f5c926369fe03d9391a93cfbf4bb9038f79ee3acd446c468d",
+ "structurally_merged_rows": 9465,
+ "backend_artifact_bytes": 6824622,
+ "backend_with_key_registry_bytes": 7422143,
+ "backend_with_datrie_key_index_bytes": 7119582,
+ "full_model_serving_artifact_bytes": 8122141,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9053524674923515,
+ "mean_cosine": 0.9996788540343262,
+ "p05_cosine": 0.9988257235328838,
+ "relative_l2_mean": 0.010728835804298021,
+ "relative_l2_p95": 0.048762315495126315,
+ "relative_l2_max": 0.4349381373732949,
+ "weighted_relative_l2_mean": 0.002949611927462981,
+ "weighted_relative_frobenius": 0.02111180791434024,
+ "weighted_mean_cosine": 0.9998876560063008,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9053524674923515,
+ "mean_cosine": 0.9996788540343262,
+ "p05_cosine": 0.9988257235328838,
+ "relative_l2_mean": 0.010728835804298021,
+ "relative_l2_p95": 0.048762315495126315,
+ "relative_l2_max": 0.4349381373732949,
+ "weighted_relative_l2_mean": 0.002949611927462981,
+ "weighted_relative_frobenius": 0.02111180791434024,
+ "weighted_mean_cosine": 0.9998876560063008,
+ "max_relative_norm_change": 0.01009546294416497
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 8.052917,
+ "warm_p50_ms": 0.055873,
+ "warm_p95_ms": 0.0595785,
+ "warm_p99_ms": 0.07523781999999998,
+ "samples_ms": [
+ 0.059575,
+ 0.058177,
+ 0.055873,
+ 0.055873,
+ 0.055873,
+ 0.057758,
+ 0.056362,
+ 0.057759,
+ 0.055384,
+ 0.058178,
+ 0.056362,
+ 0.055873,
+ 0.056781,
+ 0.05727,
+ 0.055315,
+ 0.055315,
+ 0.054895,
+ 0.056362,
+ 0.055803,
+ 0.055384,
+ 0.055385,
+ 0.055314,
+ 0.055314,
+ 0.055873,
+ 0.056292,
+ 0.055384,
+ 0.055385,
+ 0.085207,
+ 0.059645,
+ 0.055803,
+ 0.055873,
+ 0.056293,
+ 0.055384,
+ 0.055803,
+ 0.055384,
+ 0.055873,
+ 0.055803,
+ 0.055385,
+ 0.054337,
+ 0.055873
+ ],
+ "warm_rows_per_second": 9163638.966942888
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8539972901438648,
+ "character_perplexity": 6.385292443907039,
+ "next_character_accuracy": 0.502453611409293,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00017137695796017206,
+ "mean_prediction_kl_nats": 6.565723955188551e-05,
+ "prediction_kl_p99_nats": 0.00029545771466669854,
+ "maximum_prediction_kl_nats": 0.18560528499201612,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 1.3523645401000977,
+ "maximum_per_target_cross_entropy_increase": 0.592867374420166
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00017137695796017206,
+ "mean_prediction_kl_nats": 6.565723955188551e-05,
+ "prediction_kl_p99_nats": 0.00029545771466669854,
+ "maximum_prediction_kl_nats": 0.18560528499201612,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 1.3523645401000977,
+ "maximum_per_target_cross_entropy_increase": 0.592867374420166
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 128,
+ "storage": {
+ "artifact_bytes": 4707584,
+ "unique_rows": 36778,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "add9a6b9c5875f02dc14dc3783c9953d7edb14426c4012649076815ddb6e1c71",
+ "table_sha256": "dc1bc47aaa923557f0743007570dfa78218b7b4f83c5f736f5d9ba4446f6a6c1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.05553986900486052,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4708389,
+ "backend_with_key_registry_bytes": 5305910,
+ "backend_with_datrie_key_index_bytes": 5003349,
+ "full_model_serving_artifact_bytes": 6005908,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999530139869,
+ "mean_cosine": 0.9999999791887941,
+ "p05_cosine": 0.9999999715778723,
+ "relative_l2_mean": 0.00020657158571956225,
+ "relative_l2_p95": 0.00024197280109949766,
+ "relative_l2_max": 0.0003091615715408284,
+ "weighted_relative_l2_mean": 0.0001174621969176169,
+ "weighted_relative_frobenius": 0.00020737521315361893,
+ "weighted_mean_cosine": 0.999999988256516,
+ "max_relative_norm_change": 0.0001890064061433835
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.351511,
+ "warm_p50_ms": 0.13769199999999998,
+ "warm_p95_ms": 0.141945,
+ "warm_p99_ms": 0.17003546999999997,
+ "samples_ms": [
+ 0.140521,
+ 0.139124,
+ 0.137657,
+ 0.136749,
+ 0.137727,
+ 0.182565,
+ 0.141498,
+ 0.137168,
+ 0.137168,
+ 0.136749,
+ 0.138146,
+ 0.137657,
+ 0.138077,
+ 0.137657,
+ 0.138635,
+ 0.136679,
+ 0.138146,
+ 0.137657,
+ 0.137099,
+ 0.138566,
+ 0.13675,
+ 0.137657,
+ 0.138076,
+ 0.137587,
+ 0.136679,
+ 0.137657,
+ 0.137727,
+ 0.138146,
+ 0.138566,
+ 0.137168,
+ 0.137657,
+ 0.137168,
+ 0.138635,
+ 0.150438,
+ 0.139543,
+ 0.139543,
+ 0.137588,
+ 0.138077,
+ 0.139543,
+ 0.137168
+ ],
+ "warm_rows_per_second": 3718444.063562154
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853826562635324,
+ "character_perplexity": 6.3842023918903665,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.347662674319074e-07,
+ "mean_prediction_kl_nats": 4.381973465217552e-09,
+ "prediction_kl_p99_nats": 3.551782450171475e-08,
+ "maximum_prediction_kl_nats": 9.231610310342147e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001482546329498291,
+ "maximum_per_target_cross_entropy_increase": 0.0007162094116210938
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 4650510,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 575,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 575,
+ "zlib_blocks": 575,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "add9a6b9c5875f02dc14dc3783c9953d7edb14426c4012649076815ddb6e1c71",
+ "table_sha256": "45733a4f72c3f601b297dd9e696eb7b2b7a960034ab323d3cec5c151a18b7aff",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4225367580074817,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4651718,
+ "backend_with_key_registry_bytes": 5249239,
+ "backend_with_datrie_key_index_bytes": 4946678,
+ "full_model_serving_artifact_bytes": 5949237,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999530139869,
+ "mean_cosine": 0.9999999791887941,
+ "p05_cosine": 0.9999999715778723,
+ "relative_l2_mean": 0.00020657158571956225,
+ "relative_l2_p95": 0.00024197280109949766,
+ "relative_l2_max": 0.0003091615715408284,
+ "weighted_relative_l2_mean": 0.0001174621969176169,
+ "weighted_relative_frobenius": 0.00020737521315361893,
+ "weighted_mean_cosine": 0.999999988256516,
+ "max_relative_norm_change": 0.0001890064061433835
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.804073,
+ "warm_p50_ms": 0.1471205,
+ "warm_p95_ms": 0.15100389999999997,
+ "warm_p99_ms": 0.16728484,
+ "samples_ms": [
+ 0.150508,
+ 0.147575,
+ 0.148972,
+ 0.147156,
+ 0.14946,
+ 0.14953,
+ 0.147644,
+ 0.17167,
+ 0.147155,
+ 0.147645,
+ 0.147016,
+ 0.147644,
+ 0.147575,
+ 0.147645,
+ 0.146666,
+ 0.146178,
+ 0.14527,
+ 0.160426,
+ 0.147225,
+ 0.146178,
+ 0.145759,
+ 0.147156,
+ 0.146248,
+ 0.145759,
+ 0.147156,
+ 0.147086,
+ 0.146737,
+ 0.146667,
+ 0.147644,
+ 0.145689,
+ 0.145689,
+ 0.14527,
+ 0.148064,
+ 0.1452,
+ 0.146667,
+ 0.146597,
+ 0.147574,
+ 0.145689,
+ 0.145619,
+ 0.146178
+ ],
+ "warm_rows_per_second": 3480140.429104034
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853826562635324,
+ "character_perplexity": 6.3842023918903665,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.347662674319074e-07,
+ "mean_prediction_kl_nats": 4.381973465217552e-09,
+ "prediction_kl_p99_nats": 3.551782450171475e-08,
+ "maximum_prediction_kl_nats": 9.231610310342147e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001482546329498291,
+ "maximum_per_target_cross_entropy_increase": 0.0007162094116210938
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 4650510,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 575,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 575,
+ "zlib_blocks": 575,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "add9a6b9c5875f02dc14dc3783c9953d7edb14426c4012649076815ddb6e1c71",
+ "table_sha256": "45733a4f72c3f601b297dd9e696eb7b2b7a960034ab323d3cec5c151a18b7aff",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5809798981063068,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4651717,
+ "backend_with_key_registry_bytes": 5249238,
+ "backend_with_datrie_key_index_bytes": 4946677,
+ "full_model_serving_artifact_bytes": 5949236,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999530139869,
+ "mean_cosine": 0.9999999791887941,
+ "p05_cosine": 0.9999999715778723,
+ "relative_l2_mean": 0.00020657158571956225,
+ "relative_l2_p95": 0.00024197280109949766,
+ "relative_l2_max": 0.0003091615715408284,
+ "weighted_relative_l2_mean": 0.0001174621969176169,
+ "weighted_relative_frobenius": 0.00020737521315361893,
+ "weighted_mean_cosine": 0.999999988256516,
+ "max_relative_norm_change": 0.0001890064061433835
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.910581,
+ "warm_p50_ms": 0.1471555,
+ "warm_p95_ms": 0.15245644999999997,
+ "warm_p99_ms": 0.17083389999999998,
+ "samples_ms": [
+ 0.150438,
+ 0.14953,
+ 0.148482,
+ 0.147645,
+ 0.165594,
+ 0.14995,
+ 0.148972,
+ 0.149041,
+ 0.148063,
+ 0.147156,
+ 0.148064,
+ 0.147086,
+ 0.147156,
+ 0.147156,
+ 0.146666,
+ 0.146248,
+ 0.145689,
+ 0.146667,
+ 0.147644,
+ 0.14527,
+ 0.14527,
+ 0.147645,
+ 0.146667,
+ 0.147086,
+ 0.146666,
+ 0.147086,
+ 0.14569,
+ 0.146178,
+ 0.146178,
+ 0.145689,
+ 0.174184,
+ 0.151765,
+ 0.147575,
+ 0.146178,
+ 0.145689,
+ 0.145689,
+ 0.147645,
+ 0.147645,
+ 0.147155,
+ 0.145759
+ ],
+ "warm_rows_per_second": 3479312.699831131
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853826562635324,
+ "character_perplexity": 6.3842023918903665,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.347662674319074e-07,
+ "mean_prediction_kl_nats": 4.381973465217552e-09,
+ "prediction_kl_p99_nats": 3.551782450171475e-08,
+ "maximum_prediction_kl_nats": 9.231610310342147e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001482546329498291,
+ "maximum_per_target_cross_entropy_increase": 0.0007162094116210938
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 4650510,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 575,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 575,
+ "zlib_blocks": 575,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "add9a6b9c5875f02dc14dc3783c9953d7edb14426c4012649076815ddb6e1c71",
+ "table_sha256": "45733a4f72c3f601b297dd9e696eb7b2b7a960034ab323d3cec5c151a18b7aff",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.44609333807602525,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 4651717,
+ "backend_with_key_registry_bytes": 5249238,
+ "backend_with_datrie_key_index_bytes": 4946677,
+ "full_model_serving_artifact_bytes": 5949236,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999530139869,
+ "mean_cosine": 0.9999999791887941,
+ "p05_cosine": 0.9999999715778723,
+ "relative_l2_mean": 0.00020657158571956225,
+ "relative_l2_p95": 0.00024197280109949766,
+ "relative_l2_max": 0.0003091615715408284,
+ "weighted_relative_l2_mean": 0.0001174621969176169,
+ "weighted_relative_frobenius": 0.00020737521315361893,
+ "weighted_mean_cosine": 0.999999988256516,
+ "max_relative_norm_change": 0.0001890064061433835
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.812034,
+ "warm_p50_ms": 0.147086,
+ "warm_p95_ms": 0.15106334999999999,
+ "warm_p99_ms": 0.15607305999999999,
+ "samples_ms": [
+ 0.150997,
+ 0.152324,
+ 0.150439,
+ 0.14995,
+ 0.148553,
+ 0.147226,
+ 0.146666,
+ 0.146667,
+ 0.147155,
+ 0.148552,
+ 0.146667,
+ 0.14562,
+ 0.146178,
+ 0.146666,
+ 0.146178,
+ 0.148133,
+ 0.147156,
+ 0.147156,
+ 0.147994,
+ 0.147086,
+ 0.15847,
+ 0.149041,
+ 0.147086,
+ 0.146667,
+ 0.147155,
+ 0.147156,
+ 0.147575,
+ 0.145689,
+ 0.146248,
+ 0.14527,
+ 0.147155,
+ 0.146667,
+ 0.147156,
+ 0.147086,
+ 0.146667,
+ 0.145619,
+ 0.146597,
+ 0.145689,
+ 0.146597,
+ 0.146667
+ ],
+ "warm_rows_per_second": 3480956.719198292
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853826562635324,
+ "character_perplexity": 6.3842023918903665,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.347662674319074e-07,
+ "mean_prediction_kl_nats": 4.381973465217552e-09,
+ "prediction_kl_p99_nats": 3.551782450171475e-08,
+ "maximum_prediction_kl_nats": 9.231610310342147e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001482546329498291,
+ "maximum_per_target_cross_entropy_increase": 0.0007162094116210938
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3542163,
+ "unique_rows": 27383,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27383
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "6686cc166ec018c6f58281b13be9d49e7b50f16bc08937faaa161a38c13674ea",
+ "table_sha256": "72f2f63eea66f5b49160da1c9210dd11dcce2cbeff561847c4c10abd0523a5b4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3577422120142728,
+ "representatives_sha256": "ea3df8f03e9e254c1bba48f32255f8b9d75dc0da14c914a9f93c8706ea9cd7be",
+ "structurally_merged_rows": 9395,
+ "backend_artifact_bytes": 3543369,
+ "backend_with_key_registry_bytes": 4140890,
+ "backend_with_datrie_key_index_bytes": 3838329,
+ "full_model_serving_artifact_bytes": 4840888,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900583488604219,
+ "mean_cosine": 0.9997785803340887,
+ "p05_cosine": 0.9988497309240236,
+ "relative_l2_mean": 0.01025296949475231,
+ "relative_l2_p95": 0.048204921643589006,
+ "relative_l2_max": 0.1411682070021142,
+ "weighted_relative_l2_mean": 0.002618490340327802,
+ "weighted_relative_frobenius": 0.011362454733114946,
+ "weighted_mean_cosine": 0.9999512568607424,
+ "max_relative_norm_change": 0.010060688872288616
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900528053809458,
+ "mean_cosine": 0.9997785929759666,
+ "p05_cosine": 0.9988495161403594,
+ "relative_l2_mean": 0.010099256262850595,
+ "relative_l2_p95": 0.04822410935046492,
+ "relative_l2_max": 0.1412116330222307,
+ "weighted_relative_l2_mean": 0.0025162258786515057,
+ "weighted_relative_frobenius": 0.011360453108480639,
+ "weighted_mean_cosine": 0.9999512684129139,
+ "max_relative_norm_change": 0.01005837789888315
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.222644,
+ "warm_p50_ms": 0.146143,
+ "warm_p95_ms": 0.1562882,
+ "warm_p99_ms": 0.16832364,
+ "samples_ms": [
+ 0.150857,
+ 0.148064,
+ 0.147155,
+ 0.147645,
+ 0.146108,
+ 0.151416,
+ 0.147574,
+ 0.146108,
+ 0.146178,
+ 0.145689,
+ 0.14527,
+ 0.1452,
+ 0.146597,
+ 0.145689,
+ 0.145689,
+ 0.147085,
+ 0.15861,
+ 0.147225,
+ 0.146248,
+ 0.150438,
+ 0.156166,
+ 0.146178,
+ 0.146667,
+ 0.146667,
+ 0.14527,
+ 0.144851,
+ 0.174534,
+ 0.149531,
+ 0.144711,
+ 0.146108,
+ 0.146666,
+ 0.1452,
+ 0.144711,
+ 0.144781,
+ 0.145829,
+ 0.145829,
+ 0.145759,
+ 0.145759,
+ 0.14527,
+ 0.145201
+ ],
+ "warm_rows_per_second": 3503417.8852220085
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8539192812693612,
+ "character_perplexity": 6.384794353858088,
+ "next_character_accuracy": 0.502453611409293,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 9.335775766754529e-05,
+ "mean_prediction_kl_nats": 7.860089630404762e-06,
+ "prediction_kl_p99_nats": 0.0001081220717595102,
+ "maximum_prediction_kl_nats": 0.005482748176321384,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.582134485244751,
+ "maximum_per_target_cross_entropy_increase": 0.10675567388534546
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 9.272293254272358e-05,
+ "mean_prediction_kl_nats": 7.853801367279266e-06,
+ "prediction_kl_p99_nats": 0.00010800525827647296,
+ "maximum_prediction_kl_nats": 0.005482279349342036,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.5824252367019653,
+ "maximum_per_target_cross_entropy_increase": 0.10677337646484375
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3540837,
+ "unique_rows": 27370,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27370
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "0b6e274cd163a3050cc5cf3956c9740a9c224513f37ec1b53fd7d568214fe286",
+ "table_sha256": "b9f55fad4e5270d7895a8d87d13647f53ccd4630c1a2ba02bb8a45fb38257a1c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3818652050103992,
+ "representatives_sha256": "03b58da9e10c1c74137221fb8637b90f15bb9cd3ffac7696b72ddd6653ba8fef",
+ "structurally_merged_rows": 9408,
+ "backend_artifact_bytes": 3542043,
+ "backend_with_key_registry_bytes": 4139564,
+ "backend_with_datrie_key_index_bytes": 3837003,
+ "full_model_serving_artifact_bytes": 4839562,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9809678573304468,
+ "mean_cosine": 0.9997726904847208,
+ "p05_cosine": 0.998846390275071,
+ "relative_l2_mean": 0.010320209232536749,
+ "relative_l2_p95": 0.04829327747273836,
+ "relative_l2_max": 0.1950152044633947,
+ "weighted_relative_l2_mean": 0.0026817322477937677,
+ "weighted_relative_frobenius": 0.012459100118378013,
+ "weighted_mean_cosine": 0.9999459122723975,
+ "max_relative_norm_change": 0.010060688872288616
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9809711354383118,
+ "mean_cosine": 0.9997727034164549,
+ "p05_cosine": 0.9988463346832401,
+ "relative_l2_mean": 0.01016656703669211,
+ "relative_l2_p95": 0.04830014203664504,
+ "relative_l2_max": 0.19499627551793236,
+ "weighted_relative_l2_mean": 0.002579537373571873,
+ "weighted_relative_frobenius": 0.012456934608601724,
+ "weighted_mean_cosine": 0.9999459255339357,
+ "max_relative_norm_change": 0.01005837789888315
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.80157,
+ "warm_p50_ms": 0.1475745,
+ "warm_p95_ms": 0.15414704999999998,
+ "warm_p99_ms": 0.16893004,
+ "samples_ms": [
+ 0.14995,
+ 0.149112,
+ 0.150858,
+ 0.162242,
+ 0.150508,
+ 0.153721,
+ 0.14995,
+ 0.147645,
+ 0.147574,
+ 0.148133,
+ 0.148064,
+ 0.146667,
+ 0.147156,
+ 0.148623,
+ 0.146667,
+ 0.146178,
+ 0.149531,
+ 0.147086,
+ 0.147575,
+ 0.149111,
+ 0.147086,
+ 0.149041,
+ 0.147644,
+ 0.147575,
+ 0.146597,
+ 0.146597,
+ 0.146248,
+ 0.147155,
+ 0.146737,
+ 0.173206,
+ 0.150997,
+ 0.147155,
+ 0.147156,
+ 0.147156,
+ 0.14527,
+ 0.147156,
+ 0.147574,
+ 0.148134,
+ 0.146178,
+ 0.146178
+ ],
+ "warm_rows_per_second": 3469434.082446493
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8539698061616214,
+ "character_perplexity": 6.385116953054496,
+ "next_character_accuracy": 0.502453611409293,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0001438886433413367,
+ "mean_prediction_kl_nats": 1.1390478563979348e-05,
+ "prediction_kl_p99_nats": 0.00012464599131380375,
+ "maximum_prediction_kl_nats": 0.008272212034854634,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.582134485244751,
+ "maximum_per_target_cross_entropy_increase": 0.5608210563659668
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00014325378614123366,
+ "mean_prediction_kl_nats": 1.1379573729840894e-05,
+ "prediction_kl_p99_nats": 0.00012433828705492853,
+ "maximum_prediction_kl_nats": 0.00826821449514578,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.5824252367019653,
+ "maximum_per_target_cross_entropy_increase": 0.5607891082763672
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3539617,
+ "unique_rows": 27360,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27360
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "52524d8a80bf6126e4a9ec7b09426a0689fb3a44f58eaec989cbf8e43abc5d31",
+ "table_sha256": "ac74792be931d840aca9d47fb9faf3de767caabbb272f56f36c5dc6ec8be24cf",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3639816988725215,
+ "representatives_sha256": "9431822511204c421fc7b57dec7b6e3871f766ba0a2f98650f61532aeeb9d356",
+ "structurally_merged_rows": 9418,
+ "backend_artifact_bytes": 3540823,
+ "backend_with_key_registry_bytes": 4138344,
+ "backend_with_datrie_key_index_bytes": 3835783,
+ "full_model_serving_artifact_bytes": 4838342,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9704420756698033,
+ "mean_cosine": 0.9997652259801274,
+ "p05_cosine": 0.9988430293145306,
+ "relative_l2_mean": 0.010385968950148119,
+ "relative_l2_p95": 0.04835284671843402,
+ "relative_l2_max": 0.24438303843706874,
+ "weighted_relative_l2_mean": 0.002724070911672629,
+ "weighted_relative_frobenius": 0.01336469162681883,
+ "weighted_mean_cosine": 0.9999410843197322,
+ "max_relative_norm_change": 0.010060688872288616
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9704408373854408,
+ "mean_cosine": 0.9997652392798527,
+ "p05_cosine": 0.9988439386618082,
+ "relative_l2_mean": 0.010232378867656454,
+ "relative_l2_p95": 0.04836524448648539,
+ "relative_l2_max": 0.24438888634794073,
+ "weighted_relative_l2_mean": 0.0026219126257521785,
+ "weighted_relative_frobenius": 0.013362577654920042,
+ "weighted_mean_cosine": 0.9999410980709135,
+ "max_relative_norm_change": 0.01005837789888315
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.06557,
+ "warm_p50_ms": 0.146178,
+ "warm_p95_ms": 0.15030519999999997,
+ "warm_p99_ms": 0.17206923999999998,
+ "samples_ms": [
+ 0.181657,
+ 0.149949,
+ 0.147575,
+ 0.146737,
+ 0.147645,
+ 0.147156,
+ 0.147156,
+ 0.146667,
+ 0.145759,
+ 0.147156,
+ 0.146667,
+ 0.146178,
+ 0.147086,
+ 0.146178,
+ 0.146108,
+ 0.144712,
+ 0.145759,
+ 0.146178,
+ 0.147156,
+ 0.14527,
+ 0.146667,
+ 0.145759,
+ 0.1452,
+ 0.145759,
+ 0.145759,
+ 0.145759,
+ 0.157073,
+ 0.146178,
+ 0.146247,
+ 0.145689,
+ 0.146178,
+ 0.146178,
+ 0.145759,
+ 0.146248,
+ 0.146178,
+ 0.146178,
+ 0.14527,
+ 0.145759,
+ 0.144432,
+ 0.145759
+ ],
+ "warm_rows_per_second": 3502579.0474626822
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.854018102350402,
+ "character_perplexity": 6.3854253373150796,
+ "next_character_accuracy": 0.5025302867658334,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00019219294784254167,
+ "mean_prediction_kl_nats": 3.3213587724556626e-05,
+ "prediction_kl_p99_nats": 0.00013499927110225605,
+ "maximum_prediction_kl_nats": 0.056391579973223666,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 0.582134485244751,
+ "maximum_per_target_cross_entropy_increase": 0.5608210563659668
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00019155805998051503,
+ "mean_prediction_kl_nats": 3.3197522686109876e-05,
+ "prediction_kl_p99_nats": 0.00013509073392637801,
+ "maximum_prediction_kl_nats": 0.056395402051164056,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 0.5824252367019653,
+ "maximum_per_target_cross_entropy_increase": 0.5607891082763672
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3537840,
+ "unique_rows": 27345,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 428,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 428,
+ "zlib_blocks": 428,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27345
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "94fa24d32e5a88de9b1c942b7ab9edcef69331fd8ef535bab494839ab721784b",
+ "table_sha256": "57ebd55f4bb04f32efb851034d1e5b42857f2b971ea9218f16c35a29b0c2dd76",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3652330469340086,
+ "representatives_sha256": "2708ef147a95484ad1214882f63d7ea0dcca875a9e02a63b14ca3f819d1186b1",
+ "structurally_merged_rows": 9433,
+ "backend_artifact_bytes": 3539046,
+ "backend_with_key_registry_bytes": 4136567,
+ "backend_with_datrie_key_index_bytes": 3834006,
+ "full_model_serving_artifact_bytes": 4836565,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.951880122571213,
+ "mean_cosine": 0.9997394374201652,
+ "p05_cosine": 0.9988363151404382,
+ "relative_l2_mean": 0.010561375291258746,
+ "relative_l2_p95": 0.04853784508022435,
+ "relative_l2_max": 0.30901836110861,
+ "weighted_relative_l2_mean": 0.0029901262533581266,
+ "weighted_relative_frobenius": 0.02022010224983534,
+ "weighted_mean_cosine": 0.999899326033451,
+ "max_relative_norm_change": 0.010060688872288616
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9518768396606513,
+ "mean_cosine": 0.9997394490597237,
+ "p05_cosine": 0.9988363326785129,
+ "relative_l2_mean": 0.01040787604509202,
+ "relative_l2_p95": 0.048554263540243614,
+ "relative_l2_max": 0.30902503881007914,
+ "weighted_relative_l2_mean": 0.00288806554005823,
+ "weighted_relative_frobenius": 0.02021895115268188,
+ "weighted_mean_cosine": 0.9998993381404068,
+ "max_relative_norm_change": 0.01005837789888315
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.195685,
+ "warm_p50_ms": 0.147156,
+ "warm_p95_ms": 0.15144045,
+ "warm_p99_ms": 0.16545248999999998,
+ "samples_ms": [
+ 0.151905,
+ 0.149041,
+ 0.148622,
+ 0.148064,
+ 0.147645,
+ 0.150438,
+ 0.147086,
+ 0.148552,
+ 0.146248,
+ 0.148133,
+ 0.148133,
+ 0.147645,
+ 0.145619,
+ 0.147575,
+ 0.146178,
+ 0.146108,
+ 0.147156,
+ 0.147644,
+ 0.146667,
+ 0.145758,
+ 0.147156,
+ 0.147574,
+ 0.174114,
+ 0.151416,
+ 0.146667,
+ 0.146667,
+ 0.148063,
+ 0.147156,
+ 0.145689,
+ 0.147645,
+ 0.145759,
+ 0.147575,
+ 0.145759,
+ 0.145759,
+ 0.146736,
+ 0.145759,
+ 0.145759,
+ 0.146248,
+ 0.14527,
+ 0.145759
+ ],
+ "warm_rows_per_second": 3479300.877979831
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8539834781980202,
+ "character_perplexity": 6.385204251202659,
+ "next_character_accuracy": 0.5025302867658334,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00015756274046707167,
+ "mean_prediction_kl_nats": 3.984621007166472e-05,
+ "prediction_kl_p99_nats": 0.000208762134777395,
+ "maximum_prediction_kl_nats": 0.056391579973223666,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 1.3518116474151611,
+ "maximum_per_target_cross_entropy_increase": 0.5926408767700195
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00015692787458711856,
+ "mean_prediction_kl_nats": 3.9835150656727545e-05,
+ "prediction_kl_p99_nats": 0.00020869774222090573,
+ "maximum_prediction_kl_nats": 0.056395402051164056,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 1.351919412612915,
+ "maximum_per_target_cross_entropy_increase": 0.5928516387939453
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 4707584,
+ "artifact_bytes": 3534415,
+ "unique_rows": 27317,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 427,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 427,
+ "zlib_blocks": 427,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27317
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "fa0c7c96c780049c86c8ec04a7d09dc6f2a63808292ecd5be73ea266f0ebdf25",
+ "table_sha256": "f1353fb4a3847262888f7359146087f31e29c9ea150c5ead4e7e488ffa5cc68f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3661111609544605,
+ "representatives_sha256": "a14855332f8c2549544786d698d0785f8c21627102f575e233c074cf72c6ebb3",
+ "structurally_merged_rows": 9461,
+ "backend_artifact_bytes": 3535621,
+ "backend_with_key_registry_bytes": 4133142,
+ "backend_with_datrie_key_index_bytes": 3830581,
+ "full_model_serving_artifact_bytes": 4833140,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9053573483193148,
+ "mean_cosine": 0.9996789749251602,
+ "p05_cosine": 0.9988266659254238,
+ "relative_l2_mean": 0.01087655038235001,
+ "relative_l2_p95": 0.048746361326333955,
+ "relative_l2_max": 0.4349194109717122,
+ "weighted_relative_l2_mean": 0.003051557936101642,
+ "weighted_relative_frobenius": 0.02111267519256833,
+ "weighted_mean_cosine": 0.9998876477320475,
+ "max_relative_norm_change": 0.010100417761457934
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9053566777874397,
+ "mean_cosine": 0.9996789879119171,
+ "p05_cosine": 0.9988261750540566,
+ "relative_l2_mean": 0.010723201142315683,
+ "relative_l2_p95": 0.048746575378579175,
+ "relative_l2_max": 0.4349282171568034,
+ "weighted_relative_l2_mean": 0.0029495295827145374,
+ "weighted_relative_frobenius": 0.02111158170259364,
+ "weighted_mean_cosine": 0.9998876594669917,
+ "max_relative_norm_change": 0.01005837789888315
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 7.52827,
+ "warm_p50_ms": 0.14534,
+ "warm_p95_ms": 0.1558681,
+ "warm_p99_ms": 0.17647416999999999,
+ "samples_ms": [
+ 0.155676,
+ 0.153232,
+ 0.187315,
+ 0.15421,
+ 0.148552,
+ 0.146248,
+ 0.145689,
+ 0.145759,
+ 0.144712,
+ 0.146108,
+ 0.145689,
+ 0.144712,
+ 0.146178,
+ 0.14527,
+ 0.144781,
+ 0.144781,
+ 0.14534,
+ 0.14534,
+ 0.143804,
+ 0.145689,
+ 0.1452,
+ 0.14527,
+ 0.145689,
+ 0.144851,
+ 0.143873,
+ 0.144223,
+ 0.144712,
+ 0.145689,
+ 0.146178,
+ 0.159518,
+ 0.1452,
+ 0.145201,
+ 0.1452,
+ 0.1452,
+ 0.143804,
+ 0.144362,
+ 0.143803,
+ 0.145689,
+ 0.145689,
+ 0.145619
+ ],
+ "warm_rows_per_second": 3522774.1846704283
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853999420891706,
+ "character_perplexity": 6.385306049369624,
+ "next_character_accuracy": 0.502453611409293,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00017350807323292404,
+ "mean_prediction_kl_nats": 6.566471728432845e-05,
+ "prediction_kl_p99_nats": 0.00029385215046018174,
+ "maximum_prediction_kl_nats": 0.18563221670097035,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 1.3518116474151611,
+ "maximum_per_target_cross_entropy_increase": 0.5926408767700195
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00017287319723141798,
+ "mean_prediction_kl_nats": 6.565572079471623e-05,
+ "prediction_kl_p99_nats": 0.00029369762215553655,
+ "maximum_prediction_kl_nats": 0.18562726426991394,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 1.351919412612915,
+ "maximum_per_target_cross_entropy_increase": 0.5928516387939453
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 68,
+ "storage": {
+ "artifact_bytes": 2500904,
+ "unique_rows": 36778,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "fbf67b027d9c0471b313f44fdd647110f6a4ee32fdb01f42ab8203b4476c24a3",
+ "table_sha256": "80c1610db9ec1aeaad5f260db8066dd9509b9548ce61e66c9563ff1efe6127ca",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.05942556192167103,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2501752,
+ "backend_with_key_registry_bytes": 3099273,
+ "backend_with_datrie_key_index_bytes": 2796712,
+ "full_model_serving_artifact_bytes": 3799271,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999328779127546,
+ "mean_cosine": 0.9999882584835896,
+ "p05_cosine": 0.9999770406579376,
+ "relative_l2_mean": 0.004739083006836966,
+ "relative_l2_p95": 0.006820293522058443,
+ "relative_l2_max": 0.01165596953583282,
+ "weighted_relative_l2_mean": 0.0030815487651294883,
+ "weighted_relative_frobenius": 0.005367583367488608,
+ "weighted_mean_cosine": 0.9999914177601257,
+ "max_relative_norm_change": 0.0031225575438659937
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.315613,
+ "warm_p50_ms": 0.0965205,
+ "warm_p95_ms": 0.09947924999999999,
+ "warm_p99_ms": 0.10731937,
+ "samples_ms": [
+ 0.099385,
+ 0.095124,
+ 0.097429,
+ 0.097009,
+ 0.09708,
+ 0.09701,
+ 0.09652,
+ 0.094635,
+ 0.097009,
+ 0.095543,
+ 0.094985,
+ 0.096521,
+ 0.095124,
+ 0.096521,
+ 0.09652,
+ 0.095124,
+ 0.096101,
+ 0.096102,
+ 0.094636,
+ 0.096451,
+ 0.096451,
+ 0.09701,
+ 0.09701,
+ 0.09694,
+ 0.096451,
+ 0.096521,
+ 0.094566,
+ 0.094984,
+ 0.111187,
+ 0.09694,
+ 0.095124,
+ 0.096102,
+ 0.10127,
+ 0.095962,
+ 0.095054,
+ 0.097009,
+ 0.096521,
+ 0.096521,
+ 0.09694,
+ 0.096521
+ ],
+ "warm_rows_per_second": 5304572.60374739
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853797174879323,
+ "character_perplexity": 6.384014777265012,
+ "next_character_accuracy": 0.5022235853396718,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.8752576571929952e-05,
+ "mean_prediction_kl_nats": 3.7250146321206883e-06,
+ "prediction_kl_p99_nats": 2.9429470754635976e-05,
+ "maximum_prediction_kl_nats": 0.0001005267416845651,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.03496575355529785,
+ "maximum_per_target_cross_entropy_increase": 0.021662473678588867
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2802310,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 307,
+ "max_block_bytes": 8192,
+ "raw_blocks": 124,
+ "compressed_blocks": 183,
+ "zlib_blocks": 183,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "fbf67b027d9c0471b313f44fdd647110f6a4ee32fdb01f42ab8203b4476c24a3",
+ "table_sha256": "522dba7f01d79187a2b1c904e92ff5a91853b63da5e6f1a510d50e2cc1994f5a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25204654480330646,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2803563,
+ "backend_with_key_registry_bytes": 3401084,
+ "backend_with_datrie_key_index_bytes": 3098523,
+ "full_model_serving_artifact_bytes": 4101082,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999328779127546,
+ "mean_cosine": 0.9999882584835896,
+ "p05_cosine": 0.9999770406579376,
+ "relative_l2_mean": 0.004739083006836966,
+ "relative_l2_p95": 0.006820293522058443,
+ "relative_l2_max": 0.01165596953583282,
+ "weighted_relative_l2_mean": 0.0030815487651294883,
+ "weighted_relative_frobenius": 0.005367583367488608,
+ "weighted_mean_cosine": 0.9999914177601257,
+ "max_relative_norm_change": 0.0031225575438659937
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.971319,
+ "warm_p50_ms": 0.113981,
+ "warm_p95_ms": 0.12459669999999999,
+ "warm_p99_ms": 0.12754929,
+ "samples_ms": [
+ 0.127251,
+ 0.117823,
+ 0.114959,
+ 0.113981,
+ 0.114051,
+ 0.11447,
+ 0.115448,
+ 0.113981,
+ 0.114889,
+ 0.113004,
+ 0.113632,
+ 0.11454,
+ 0.116915,
+ 0.120128,
+ 0.114959,
+ 0.112026,
+ 0.113493,
+ 0.11447,
+ 0.113562,
+ 0.12774,
+ 0.124457,
+ 0.113562,
+ 0.113562,
+ 0.113562,
+ 0.11454,
+ 0.11447,
+ 0.112514,
+ 0.117753,
+ 0.116356,
+ 0.112165,
+ 0.113004,
+ 0.113003,
+ 0.113562,
+ 0.113981,
+ 0.112165,
+ 0.113562,
+ 0.111606,
+ 0.112095,
+ 0.111606,
+ 0.112584
+ ],
+ "warm_rows_per_second": 4491976.732964266
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853797174879323,
+ "character_perplexity": 6.384014777265012,
+ "next_character_accuracy": 0.5022235853396718,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.8752576571929952e-05,
+ "mean_prediction_kl_nats": 3.7250146321206883e-06,
+ "prediction_kl_p99_nats": 2.9429470754635976e-05,
+ "maximum_prediction_kl_nats": 0.0001005267416845651,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.03496575355529785,
+ "maximum_per_target_cross_entropy_increase": 0.021662473678588867
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2802310,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 307,
+ "max_block_bytes": 8192,
+ "raw_blocks": 124,
+ "compressed_blocks": 183,
+ "zlib_blocks": 183,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "fbf67b027d9c0471b313f44fdd647110f6a4ee32fdb01f42ab8203b4476c24a3",
+ "table_sha256": "522dba7f01d79187a2b1c904e92ff5a91853b63da5e6f1a510d50e2cc1994f5a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.38379638898186386,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2803562,
+ "backend_with_key_registry_bytes": 3401083,
+ "backend_with_datrie_key_index_bytes": 3098522,
+ "full_model_serving_artifact_bytes": 4101081,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999328779127546,
+ "mean_cosine": 0.9999882584835896,
+ "p05_cosine": 0.9999770406579376,
+ "relative_l2_mean": 0.004739083006836966,
+ "relative_l2_p95": 0.006820293522058443,
+ "relative_l2_max": 0.01165596953583282,
+ "weighted_relative_l2_mean": 0.0030815487651294883,
+ "weighted_relative_frobenius": 0.005367583367488608,
+ "weighted_mean_cosine": 0.9999914177601257,
+ "max_relative_norm_change": 0.0031225575438659937
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.961891,
+ "warm_p50_ms": 0.11359649999999999,
+ "warm_p95_ms": 0.12108409999999999,
+ "warm_p99_ms": 0.12954287,
+ "samples_ms": [
+ 0.130115,
+ 0.116845,
+ 0.11454,
+ 0.11447,
+ 0.113981,
+ 0.1144,
+ 0.113632,
+ 0.114051,
+ 0.114051,
+ 0.116844,
+ 0.113143,
+ 0.113562,
+ 0.113631,
+ 0.113143,
+ 0.114051,
+ 0.111746,
+ 0.113143,
+ 0.113562,
+ 0.119289,
+ 0.115866,
+ 0.112584,
+ 0.120686,
+ 0.113143,
+ 0.114051,
+ 0.111746,
+ 0.111677,
+ 0.113562,
+ 0.113143,
+ 0.113562,
+ 0.112724,
+ 0.113143,
+ 0.115378,
+ 0.113003,
+ 0.114959,
+ 0.111607,
+ 0.128648,
+ 0.112165,
+ 0.111187,
+ 0.112235,
+ 0.114121
+ ],
+ "warm_rows_per_second": 4507181.119136593
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853797174879323,
+ "character_perplexity": 6.384014777265012,
+ "next_character_accuracy": 0.5022235853396718,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.8752576571929952e-05,
+ "mean_prediction_kl_nats": 3.7250146321206883e-06,
+ "prediction_kl_p99_nats": 2.9429470754635976e-05,
+ "maximum_prediction_kl_nats": 0.0001005267416845651,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.03496575355529785,
+ "maximum_per_target_cross_entropy_increase": 0.021662473678588867
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2802310,
+ "unique_rows": 36778,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 307,
+ "max_block_bytes": 8192,
+ "raw_blocks": 124,
+ "compressed_blocks": 183,
+ "zlib_blocks": 183,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 36778
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "fbf67b027d9c0471b313f44fdd647110f6a4ee32fdb01f42ab8203b4476c24a3",
+ "table_sha256": "522dba7f01d79187a2b1c904e92ff5a91853b63da5e6f1a510d50e2cc1994f5a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2576393000781536,
+ "representatives_sha256": "c7545f3e40b9c2c56e00c6fd4bdca8947aa24af4f5a61371a76ae6bd044a19f3",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2803562,
+ "backend_with_key_registry_bytes": 3401083,
+ "backend_with_datrie_key_index_bytes": 3098522,
+ "full_model_serving_artifact_bytes": 4101081,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999328779127546,
+ "mean_cosine": 0.9999882584835896,
+ "p05_cosine": 0.9999770406579376,
+ "relative_l2_mean": 0.004739083006836966,
+ "relative_l2_p95": 0.006820293522058443,
+ "relative_l2_max": 0.01165596953583282,
+ "weighted_relative_l2_mean": 0.0030815487651294883,
+ "weighted_relative_frobenius": 0.005367583367488608,
+ "weighted_mean_cosine": 0.9999914177601257,
+ "max_relative_norm_change": 0.0031225575438659937
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.004005,
+ "warm_p50_ms": 0.1130735,
+ "warm_p95_ms": 0.12035764999999998,
+ "warm_p99_ms": 0.12664727,
+ "samples_ms": [
+ 0.117823,
+ 0.11454,
+ 0.113073,
+ 0.114959,
+ 0.115937,
+ 0.120127,
+ 0.12341,
+ 0.119638,
+ 0.113562,
+ 0.11447,
+ 0.113073,
+ 0.113562,
+ 0.113074,
+ 0.116426,
+ 0.112584,
+ 0.112585,
+ 0.113143,
+ 0.113073,
+ 0.113003,
+ 0.120197,
+ 0.112584,
+ 0.112165,
+ 0.112654,
+ 0.113073,
+ 0.113073,
+ 0.113073,
+ 0.128717,
+ 0.112584,
+ 0.120197,
+ 0.112654,
+ 0.112584,
+ 0.113143,
+ 0.111257,
+ 0.118731,
+ 0.112096,
+ 0.111607,
+ 0.113492,
+ 0.113073,
+ 0.112096,
+ 0.113562
+ ],
+ "warm_rows_per_second": 4528028.229425994
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853797174879323,
+ "character_perplexity": 6.384014777265012,
+ "next_character_accuracy": 0.5022235853396718,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.8752576571929952e-05,
+ "mean_prediction_kl_nats": 3.7250146321206883e-06,
+ "prediction_kl_p99_nats": 2.9429470754635976e-05,
+ "maximum_prediction_kl_nats": 0.0001005267416845651,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.03496575355529785,
+ "maximum_per_target_cross_entropy_increase": 0.021662473678588867
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2161673,
+ "unique_rows": 27405,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 229,
+ "max_block_bytes": 8192,
+ "raw_blocks": 52,
+ "compressed_blocks": 177,
+ "zlib_blocks": 177,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27405
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "f26edc2c27db1202b1bd93c6f9887d1355120d2fb8b6911e9be5182e8128ca01",
+ "table_sha256": "d5cfb885300ffd29876b8e66b47d070093f7e9cfc6eea267728af740d86952ba",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.21681339899078012,
+ "representatives_sha256": "5e5affe13c9770ee8e09b268db8739f3b591e94c60fa195def4a4c9afa4d4c54",
+ "structurally_merged_rows": 9373,
+ "backend_artifact_bytes": 2162923,
+ "backend_with_key_registry_bytes": 2760444,
+ "backend_with_datrie_key_index_bytes": 2457883,
+ "full_model_serving_artifact_bytes": 3460442,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9900501912925984,
+ "mean_cosine": 0.9997672440053686,
+ "p05_cosine": 0.9988363234103511,
+ "relative_l2_mean": 0.01374289161617526,
+ "relative_l2_p95": 0.04850331401862334,
+ "relative_l2_max": 0.1411733068123005,
+ "weighted_relative_l2_mean": 0.005256490610976462,
+ "weighted_relative_frobenius": 0.012504361072079999,
+ "weighted_mean_cosine": 0.9999430264833594,
+ "max_relative_norm_change": 0.01171282268094917
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9901436254618594,
+ "mean_cosine": 0.9997735368835137,
+ "p05_cosine": 0.9988227748854258,
+ "relative_l2_mean": 0.01021942420837844,
+ "relative_l2_p95": 0.04874759081758667,
+ "relative_l2_max": 0.1404733245529406,
+ "weighted_relative_l2_mean": 0.00255980257747284,
+ "weighted_relative_frobenius": 0.011554937193940731,
+ "weighted_mean_cosine": 0.9999498965417413,
+ "max_relative_norm_change": 0.010098008799951975
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.389599,
+ "warm_p50_ms": 0.109267,
+ "warm_p95_ms": 0.11188219999999999,
+ "warm_p99_ms": 0.1199661,
+ "samples_ms": [
+ 0.11447,
+ 0.110699,
+ 0.111606,
+ 0.109232,
+ 0.11035,
+ 0.108813,
+ 0.109302,
+ 0.110349,
+ 0.109721,
+ 0.11021,
+ 0.109861,
+ 0.107416,
+ 0.11021,
+ 0.110628,
+ 0.10972,
+ 0.107905,
+ 0.111187,
+ 0.109302,
+ 0.111257,
+ 0.107835,
+ 0.106997,
+ 0.107835,
+ 0.106508,
+ 0.111257,
+ 0.12348,
+ 0.108394,
+ 0.107835,
+ 0.111746,
+ 0.11021,
+ 0.11021,
+ 0.107905,
+ 0.108394,
+ 0.107416,
+ 0.107416,
+ 0.106927,
+ 0.107905,
+ 0.107905,
+ 0.108324,
+ 0.107835,
+ 0.108813
+ ],
+ "warm_rows_per_second": 4685769.720043562
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8539012885009298,
+ "character_perplexity": 6.384679474765296,
+ "next_character_accuracy": 0.5022235853396718,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 7.536347135542887e-05,
+ "mean_prediction_kl_nats": 1.1918478848719526e-05,
+ "prediction_kl_p99_nats": 0.0001182874969071734,
+ "maximum_prediction_kl_nats": 0.005813479238591007,
+ "top1_changed_fraction": 0.002376936052752645,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.5850428342819214,
+ "maximum_per_target_cross_entropy_increase": 0.11060148477554321
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00010411904161807553,
+ "mean_prediction_kl_nats": 8.02457921924959e-06,
+ "prediction_kl_p99_nats": 0.00011312616276467465,
+ "maximum_prediction_kl_nats": 0.005654594889383709,
+ "top1_changed_fraction": 0.000996779635025303,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.5810846090316772,
+ "maximum_per_target_cross_entropy_increase": 0.1088036298751831
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2160841,
+ "unique_rows": 27393,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 229,
+ "max_block_bytes": 8192,
+ "raw_blocks": 52,
+ "compressed_blocks": 177,
+ "zlib_blocks": 177,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27393
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "e0c8f404698ad436b82903190fb44ff1b86d1b272ece6f794e25c18f86e921e0",
+ "table_sha256": "4454d4e7b75dfde20d97502e573db109509ad85ceda581ca2e36fe031b9e1e8e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22280641505494714,
+ "representatives_sha256": "34180d652b027806c369e5c60ee928d969bd24b330aec9f513a07e1ac98d1af9",
+ "structurally_merged_rows": 9385,
+ "backend_artifact_bytes": 2162091,
+ "backend_with_key_registry_bytes": 2759612,
+ "backend_with_datrie_key_index_bytes": 2457051,
+ "full_model_serving_artifact_bytes": 3459610,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9809002275561916,
+ "mean_cosine": 0.9997616418247693,
+ "p05_cosine": 0.9988324180154542,
+ "relative_l2_mean": 0.013804670547443734,
+ "relative_l2_p95": 0.04857311844876219,
+ "relative_l2_max": 0.19525907761117325,
+ "weighted_relative_l2_mean": 0.005316683713011929,
+ "weighted_relative_frobenius": 0.01349805851442408,
+ "weighted_mean_cosine": 0.9999377895885587,
+ "max_relative_norm_change": 0.01171282268094917
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9810075485672025,
+ "mean_cosine": 0.9997679292523991,
+ "p05_cosine": 0.998817713675655,
+ "relative_l2_mean": 0.010282776670718885,
+ "relative_l2_p95": 0.0488810630868605,
+ "relative_l2_max": 0.19477468433844242,
+ "weighted_relative_l2_mean": 0.002621579245133783,
+ "weighted_relative_frobenius": 0.012622164819903855,
+ "weighted_mean_cosine": 0.9999446580075028,
+ "max_relative_norm_change": 0.010098008799951975
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.677694,
+ "warm_p50_ms": 0.108813,
+ "warm_p95_ms": 0.113981,
+ "warm_p99_ms": 0.12173471,
+ "samples_ms": [
+ 0.113981,
+ 0.112096,
+ 0.109791,
+ 0.109721,
+ 0.126692,
+ 0.11028,
+ 0.109302,
+ 0.108812,
+ 0.110769,
+ 0.108813,
+ 0.108324,
+ 0.109302,
+ 0.109301,
+ 0.108743,
+ 0.108812,
+ 0.109232,
+ 0.109791,
+ 0.108813,
+ 0.106927,
+ 0.109232,
+ 0.107416,
+ 0.107347,
+ 0.108813,
+ 0.107835,
+ 0.107416,
+ 0.113981,
+ 0.107416,
+ 0.106439,
+ 0.107905,
+ 0.107835,
+ 0.108394,
+ 0.108743,
+ 0.109372,
+ 0.108324,
+ 0.110279,
+ 0.107835,
+ 0.109371,
+ 0.108394,
+ 0.109372,
+ 0.108813
+ ],
+ "warm_rows_per_second": 4705320.136380764
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8539514263478871,
+ "character_perplexity": 6.3849995968727224,
+ "next_character_accuracy": 0.5022235853396718,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00012550635389250626,
+ "mean_prediction_kl_nats": 1.5419284971761743e-05,
+ "prediction_kl_p99_nats": 0.0001326295120033132,
+ "maximum_prediction_kl_nats": 0.008157706561191278,
+ "top1_changed_fraction": 0.002376936052752645,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.5850428342819214,
+ "maximum_per_target_cross_entropy_increase": 0.5561943054199219
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00015426336593367745,
+ "mean_prediction_kl_nats": 1.1507709268509601e-05,
+ "prediction_kl_p99_nats": 0.00012408187144064502,
+ "maximum_prediction_kl_nats": 0.008095447209669888,
+ "top1_changed_fraction": 0.000996779635025303,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.5810846090316772,
+ "maximum_per_target_cross_entropy_increase": 0.5560650825500488
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2160199,
+ "unique_rows": 27383,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 229,
+ "max_block_bytes": 8192,
+ "raw_blocks": 49,
+ "compressed_blocks": 180,
+ "zlib_blocks": 180,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27383
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "d6a1157c5f60f37187619fbc4ff50a9b918e4d66dfeb79d16fc3c91121228544",
+ "table_sha256": "ef1a9d21b228ad68dfe3d111695162d2548e0825253498343ec06d93910528f4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.22293834504671395,
+ "representatives_sha256": "06cae1a8e19d95f9221a3e4db79c12c38b3a15b1160a2978d286e849085fb4f9",
+ "structurally_merged_rows": 9395,
+ "backend_artifact_bytes": 2161449,
+ "backend_with_key_registry_bytes": 2758970,
+ "backend_with_datrie_key_index_bytes": 2456409,
+ "full_model_serving_artifact_bytes": 3458968,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9707483615619635,
+ "mean_cosine": 0.9997535139232898,
+ "p05_cosine": 0.9988292108161556,
+ "relative_l2_mean": 0.013874160044969629,
+ "relative_l2_p95": 0.048658593607048724,
+ "relative_l2_max": 0.24306846391228168,
+ "weighted_relative_l2_mean": 0.005358170617296778,
+ "weighted_relative_frobenius": 0.014340872424942506,
+ "weighted_mean_cosine": 0.9999329512941609,
+ "max_relative_norm_change": 0.01171282268094917
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.970906149365439,
+ "mean_cosine": 0.9997597891769698,
+ "p05_cosine": 0.9988121313510618,
+ "relative_l2_mean": 0.010353640332799964,
+ "relative_l2_p95": 0.04898206383172248,
+ "relative_l2_max": 0.24246087148998058,
+ "weighted_relative_l2_mean": 0.002664026215183575,
+ "weighted_relative_frobenius": 0.013518784453951401,
+ "weighted_mean_cosine": 0.9999398142624463,
+ "max_relative_norm_change": 0.010098008799951975
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.554285,
+ "warm_p50_ms": 0.11021,
+ "warm_p95_ms": 0.11684755000000001,
+ "warm_p99_ms": 0.13651185999999998,
+ "samples_ms": [
+ 0.112584,
+ 0.111118,
+ 0.113074,
+ 0.112095,
+ 0.110768,
+ 0.113562,
+ 0.111187,
+ 0.116844,
+ 0.110698,
+ 0.110698,
+ 0.110768,
+ 0.108743,
+ 0.149041,
+ 0.109232,
+ 0.109302,
+ 0.109232,
+ 0.110768,
+ 0.112095,
+ 0.110209,
+ 0.11021,
+ 0.11021,
+ 0.116426,
+ 0.108813,
+ 0.116915,
+ 0.111746,
+ 0.11021,
+ 0.109721,
+ 0.108324,
+ 0.109791,
+ 0.110279,
+ 0.108813,
+ 0.109302,
+ 0.109232,
+ 0.109302,
+ 0.109302,
+ 0.110768,
+ 0.108813,
+ 0.108813,
+ 0.110209,
+ 0.109302
+ ],
+ "warm_rows_per_second": 4645676.43589511
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853999535347123,
+ "character_perplexity": 6.385306780202532,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00017362254851536576,
+ "mean_prediction_kl_nats": 3.727564581693561e-05,
+ "prediction_kl_p99_nats": 0.00014125358045744517,
+ "maximum_prediction_kl_nats": 0.05768568200967636,
+ "top1_changed_fraction": 0.002453611409293053,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.5850428342819214,
+ "maximum_per_target_cross_entropy_increase": 0.5561943054199219
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00020238094406088654,
+ "mean_prediction_kl_nats": 3.349076802780913e-05,
+ "prediction_kl_p99_nats": 0.00013267263580251393,
+ "maximum_prediction_kl_nats": 0.05675248248867579,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.5810846090316772,
+ "maximum_per_target_cross_entropy_increase": 0.5560650825500488
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2159238,
+ "unique_rows": 27369,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 229,
+ "max_block_bytes": 8192,
+ "raw_blocks": 54,
+ "compressed_blocks": 175,
+ "zlib_blocks": 175,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27369
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "71d8ae71b3626295e60df1ce6558160481731a519f86e15592346beb286ac512",
+ "table_sha256": "40a8d067484b995af8774ab44788662bb13f8816f15edf611320c5a507b9110e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.27339977002702653,
+ "representatives_sha256": "abbd52b531a81904fa1dfdc8657b8ded428c5c5471ba2ce2fb606fdb12a85199",
+ "structurally_merged_rows": 9409,
+ "backend_artifact_bytes": 2160488,
+ "backend_with_key_registry_bytes": 2758009,
+ "backend_with_datrie_key_index_bytes": 2455448,
+ "full_model_serving_artifact_bytes": 3458007,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.9519423666731748,
+ "mean_cosine": 0.9997315251747293,
+ "p05_cosine": 0.9988213978259789,
+ "relative_l2_mean": 0.014023950416281409,
+ "relative_l2_p95": 0.04884440100473431,
+ "relative_l2_max": 0.3088439239490936,
+ "weighted_relative_l2_mean": 0.005432842809384321,
+ "weighted_relative_frobenius": 0.017576692562555705,
+ "weighted_mean_cosine": 0.9999215989916698,
+ "max_relative_norm_change": 0.01171282268094917
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9522378764573203,
+ "mean_cosine": 0.9997378001424586,
+ "p05_cosine": 0.9988051283088357,
+ "relative_l2_mean": 0.01050519037324541,
+ "relative_l2_p95": 0.04912235471432704,
+ "relative_l2_max": 0.30808090618982736,
+ "weighted_relative_l2_mean": 0.0027396680528886347,
+ "weighted_relative_frobenius": 0.016913122642540572,
+ "weighted_mean_cosine": 0.999928465481976,
+ "max_relative_norm_change": 0.010098008799951975
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.420888,
+ "warm_p50_ms": 0.1093015,
+ "warm_p95_ms": 0.11358295,
+ "warm_p99_ms": 0.13072367,
+ "samples_ms": [
+ 0.113981,
+ 0.111607,
+ 0.112095,
+ 0.111117,
+ 0.110209,
+ 0.109791,
+ 0.109721,
+ 0.108882,
+ 0.109302,
+ 0.108883,
+ 0.109301,
+ 0.108394,
+ 0.10979,
+ 0.106927,
+ 0.108324,
+ 0.107835,
+ 0.106927,
+ 0.113562,
+ 0.108813,
+ 0.109232,
+ 0.106927,
+ 0.107835,
+ 0.108813,
+ 0.110769,
+ 0.141428,
+ 0.113003,
+ 0.110768,
+ 0.110209,
+ 0.108743,
+ 0.108324,
+ 0.109232,
+ 0.109302,
+ 0.109231,
+ 0.108813,
+ 0.111188,
+ 0.109302,
+ 0.108394,
+ 0.107975,
+ 0.110699,
+ 0.11028
+ ],
+ "warm_rows_per_second": 4684290.700493589
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8541026324868122,
+ "character_perplexity": 6.385965121003553,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00027674290380804155,
+ "mean_prediction_kl_nats": 4.089533978135044e-05,
+ "prediction_kl_p99_nats": 0.00015913884361144735,
+ "maximum_prediction_kl_nats": 0.05768568200967636,
+ "top1_changed_fraction": 0.002453611409293053,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 1.3615862131118774,
+ "maximum_per_target_cross_entropy_increase": 0.5845675468444824
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00030550426441472715,
+ "mean_prediction_kl_nats": 3.712668694624218e-05,
+ "prediction_kl_p99_nats": 0.00014993490974764533,
+ "maximum_prediction_kl_nats": 0.05675248248867579,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.3547651767730713,
+ "maximum_per_target_cross_entropy_increase": 0.5863943099975586
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 36778,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2500904,
+ "artifact_bytes": 2157420,
+ "unique_rows": 27342,
+ "row_map_bytes": 294224,
+ "n_rows": 36778,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 228,
+ "max_block_bytes": 8192,
+ "raw_blocks": 49,
+ "compressed_blocks": 179,
+ "zlib_blocks": 179,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 27342
+ },
+ "source_float32_sha256": "b64693e53f10f160fa384ab8c75dd6a239d68853eb1740a7493797df4b31c874",
+ "decoded_float32_sha256": "e024b76412db848c22f51f6d98ad53262f492c8cb79762aa1f652a285378830e",
+ "table_sha256": "24e92a181c4c69b9f8656d927bc2461c9d867b8e6d211edb3a15eb2e30957dea",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2204549261368811,
+ "representatives_sha256": "9cbd43f53d4fb91de9d18e8d449a3b318a01ede0ae91fc7cd567d5777161b0d0",
+ "structurally_merged_rows": 9436,
+ "backend_artifact_bytes": 2158670,
+ "backend_with_key_registry_bytes": 2756191,
+ "backend_with_datrie_key_index_bytes": 2453630,
+ "full_model_serving_artifact_bytes": 3456189,
+ "fidelity_vs_float32": {
+ "rows": 36778,
+ "minimum_cosine": 0.8993976984974948,
+ "mean_cosine": 0.9996728223469323,
+ "p05_cosine": 0.9988108110186172,
+ "relative_l2_mean": 0.014325679044866257,
+ "relative_l2_p95": 0.049039959300843365,
+ "relative_l2_max": 0.4471761870867479,
+ "weighted_relative_l2_mean": 0.005501955084736431,
+ "weighted_relative_frobenius": 0.01870043506699296,
+ "weighted_mean_cosine": 0.9999082514088644,
+ "max_relative_norm_change": 0.01171282268094917
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 36778,
+ "minimum_cosine": 0.9002733459621711,
+ "mean_cosine": 0.9996790761147037,
+ "p05_cosine": 0.9987953081996668,
+ "relative_l2_mean": 0.010810707385642847,
+ "relative_l2_p95": 0.04928816315452132,
+ "relative_l2_max": 0.44525560469347747,
+ "weighted_relative_l2_mean": 0.0028097726571443588,
+ "weighted_relative_frobenius": 0.018076661725040872,
+ "weighted_mean_cosine": 0.9999151213083863,
+ "max_relative_norm_change": 0.010098008799951975
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.400425,
+ "warm_p50_ms": 0.11017450000000001,
+ "warm_p95_ms": 0.1145224,
+ "warm_p99_ms": 0.12122638,
+ "samples_ms": [
+ 0.113003,
+ 0.112584,
+ 0.11447,
+ 0.111118,
+ 0.109302,
+ 0.111257,
+ 0.111187,
+ 0.111117,
+ 0.108813,
+ 0.109302,
+ 0.109232,
+ 0.110279,
+ 0.109232,
+ 0.109791,
+ 0.10972,
+ 0.108953,
+ 0.109721,
+ 0.111117,
+ 0.109371,
+ 0.109302,
+ 0.11021,
+ 0.10986,
+ 0.108813,
+ 0.109791,
+ 0.108812,
+ 0.11028,
+ 0.11028,
+ 0.115518,
+ 0.10979,
+ 0.111117,
+ 0.107346,
+ 0.124876,
+ 0.10986,
+ 0.110209,
+ 0.112514,
+ 0.110698,
+ 0.110349,
+ 0.109791,
+ 0.11014,
+ 0.111257
+ ],
+ "warm_rows_per_second": 4647173.34773473
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8540404612293304,
+ "character_perplexity": 6.385568109863214,
+ "next_character_accuracy": 0.5022235853396718,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00021455637399936685,
+ "mean_prediction_kl_nats": 7.767856944016081e-05,
+ "prediction_kl_p99_nats": 0.00018685498443185095,
+ "maximum_prediction_kl_nats": 0.1874569444533677,
+ "top1_changed_fraction": 0.002606962122373869,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.3615862131118774,
+ "maximum_per_target_cross_entropy_increase": 0.5845675468444824
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00024331594653168061,
+ "mean_prediction_kl_nats": 7.382736956145754e-05,
+ "prediction_kl_p99_nats": 0.00017967533745234001,
+ "maximum_prediction_kl_nats": 0.18743622827197054,
+ "top1_changed_fraction": 0.0012268057046465265,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 1.3547651767730713,
+ "maximum_per_target_cross_entropy_increase": 0.5863943099975586
+ }
+ }
+ ],
+ "baseline_quality": {
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.853825927869258,
+ "character_perplexity": 6.384198339416614,
+ "next_character_accuracy": 0.5023002606962123,
+ "unknown_targets": 0
+ },
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7280639176093653,
+ "character_perplexity": 5.629743702661178,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ }
+ },
+ "selection": {
+ "fp32": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 7422143,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7281426950457552,
+ "character_perplexity": 5.630187216906834,
+ "next_character_accuracy": 0.5290193973779038,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 7.878053941357795e-05,
+ "mean_prediction_kl_nats": 7.239235945488115e-05,
+ "prediction_kl_p99_nats": 0.00025523719050292914,
+ "maximum_prediction_kl_nats": 0.15028831154148714,
+ "top1_changed_fraction": 0.0007666947788085563,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 1.3504877090454102,
+ "maximum_per_target_cross_entropy_increase": 0.8670740127563477
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00012532674483123354,
+ 0.000296544972959154
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "fp16": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 4133142,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7281424986147633,
+ "character_perplexity": 5.630186110963683,
+ "next_character_accuracy": 0.5290193973779038,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 7.858409296605416e-05,
+ "mean_prediction_kl_nats": 7.239406989054103e-05,
+ "prediction_kl_p99_nats": 0.0002553027836586025,
+ "maximum_prediction_kl_nats": 0.15030648367116142,
+ "top1_changed_fraction": 0.0007666947788085563,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 1.349926471710205,
+ "maximum_per_target_cross_entropy_increase": 0.867246150970459
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0001246097273672068,
+ 0.00029588735143539364
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "int8": {
+ "feasible": true,
+ "name": "int8-joint-0.9",
+ "backend_with_key_registry_bytes": 2756191,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7281543275316205,
+ "character_perplexity": 5.6302527103609785,
+ "next_character_accuracy": 0.5291727363336656,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 9.041400935520268e-05,
+ "mean_prediction_kl_nats": 9.501836386167312e-05,
+ "prediction_kl_p99_nats": 0.0002800923076272734,
+ "maximum_prediction_kl_nats": 0.14853157432377162,
+ "top1_changed_fraction": 0.0024534232921873803,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 1.359821081161499,
+ "maximum_per_target_cross_entropy_increase": 0.8680667877197266
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00016767168309531454,
+ 0.00033279514032229094
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ }
+ },
+ "fixed_quantization_test_controls": {
+ "fp16": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7280637701909376,
+ "character_perplexity": 5.629742872733274,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -1.4741841682677274e-07,
+ "mean_prediction_kl_nats": 3.973670847115932e-09,
+ "prediction_kl_p99_nats": 2.8094301446586266e-08,
+ "maximum_prediction_kl_nats": 6.983048011522801e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0011174902319908142,
+ "maximum_per_target_cross_entropy_increase": 0.000797271728515625
+ }
+ },
+ "int8": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.728062761668189,
+ "character_perplexity": 5.629737195012381,
+ "next_character_accuracy": 0.5293260752894273,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -1.1559405082255191e-06,
+ "mean_prediction_kl_nats": 3.65344786391907e-06,
+ "prediction_kl_p99_nats": 2.9962297667047917e-05,
+ "maximum_prediction_kl_nats": 0.00014314058747329998,
+ "top1_changed_fraction": 0.0011500421682128345,
+ "accuracy_change": 7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.038996338844299316,
+ "maximum_per_target_cross_entropy_increase": 0.033242225646972656
+ }
+ }
+ },
+ "checkpoint_sha256": "b8b0017a1ced25157ff108b6d38a5aaf5dfe56e59ce071a16755aa916bd279f5",
+ "registry_sha256": "e22a7692100a89679e3088226c3399678ae46a06d73d93bef7d898aa00f6fb7c"
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 3,
+ "rows": 43152,
+ "dim": 64,
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "source_float32_bytes": 11046912,
+ "key_registry_bytes": 736107,
+ "unchanged_backbone_artifact_bytes": 827998,
+ "alternative_datrie_key_index_bytes": 346160,
+ "query_rows": 512,
+ "query_ids_sha256": "43ca78408803e060726877734dd5761b4e85ad9be7de5a6b6d8601f32db85068",
+ "frequency_trace_sha256": "7ba41132ac5296b1ee890e8c97c4f892398df8fb97e4bbb475dab44b5c5af0cc",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 11046912,
+ "unique_rows": 43152,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "table_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.05877554905600846,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 11047717,
+ "backend_with_key_registry_bytes": 11783824,
+ "backend_with_datrie_key_index_bytes": 11393877,
+ "full_model_serving_artifact_bytes": 12611822,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.23802,
+ "warm_p50_ms": 0.046863,
+ "warm_p95_ms": 0.04781735,
+ "warm_p99_ms": 0.0516192,
+ "samples_ms": [
+ 0.053499,
+ 0.047352,
+ 0.046863,
+ 0.046375,
+ 0.047283,
+ 0.047772,
+ 0.047352,
+ 0.048679,
+ 0.047213,
+ 0.047282,
+ 0.047282,
+ 0.046864,
+ 0.046794,
+ 0.046863,
+ 0.046863,
+ 0.047283,
+ 0.047353,
+ 0.046444,
+ 0.047282,
+ 0.046864,
+ 0.045886,
+ 0.046444,
+ 0.045886,
+ 0.045886,
+ 0.045886,
+ 0.046304,
+ 0.046305,
+ 0.046305,
+ 0.046305,
+ 0.046375,
+ 0.046864,
+ 0.046375,
+ 0.046863,
+ 0.046374,
+ 0.047352,
+ 0.046794,
+ 0.046375,
+ 0.046794,
+ 0.047352,
+ 0.047282
+ ],
+ "warm_rows_per_second": 10925463.585344514
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772440363845078,
+ "character_perplexity": 6.535468506423138,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 10654345,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1349,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1349,
+ "zlib_blocks": 1349,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "table_sha256": "8480d0c01978ff0fa6f93d6530acada4a8d45fc316a37a67164b316eef6833a2",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.9206812500488013,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 10655557,
+ "backend_with_key_registry_bytes": 11391664,
+ "backend_with_datrie_key_index_bytes": 11001717,
+ "full_model_serving_artifact_bytes": 12219662,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.194159,
+ "warm_p50_ms": 0.055803,
+ "warm_p95_ms": 0.05771344999999999,
+ "warm_p99_ms": 0.058178,
+ "samples_ms": [
+ 0.058178,
+ 0.057689,
+ 0.05727,
+ 0.0572,
+ 0.058178,
+ 0.055873,
+ 0.056781,
+ 0.055873,
+ 0.055803,
+ 0.056711,
+ 0.055734,
+ 0.055803,
+ 0.055314,
+ 0.055803,
+ 0.055314,
+ 0.055804,
+ 0.055804,
+ 0.054407,
+ 0.055803,
+ 0.055314,
+ 0.054406,
+ 0.054336,
+ 0.053428,
+ 0.054895,
+ 0.055314,
+ 0.055803,
+ 0.055803,
+ 0.055384,
+ 0.054407,
+ 0.054407,
+ 0.054476,
+ 0.056292,
+ 0.055315,
+ 0.055384,
+ 0.055454,
+ 0.055803,
+ 0.055384,
+ 0.055873,
+ 0.055385,
+ 0.055803
+ ],
+ "warm_rows_per_second": 9175133.953371683
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772440363845078,
+ "character_perplexity": 6.535468506423138,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 10654345,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1349,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1349,
+ "zlib_blocks": 1349,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "table_sha256": "8480d0c01978ff0fa6f93d6530acada4a8d45fc316a37a67164b316eef6833a2",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 1.0852725929580629,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 10655556,
+ "backend_with_key_registry_bytes": 11391663,
+ "backend_with_datrie_key_index_bytes": 11001716,
+ "full_model_serving_artifact_bytes": 12219661,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.192204,
+ "warm_p50_ms": 0.0548955,
+ "warm_p95_ms": 0.056892899999999996,
+ "warm_p99_ms": 0.05947874,
+ "samples_ms": [
+ 0.057689,
+ 0.055384,
+ 0.056851,
+ 0.055385,
+ 0.055384,
+ 0.055803,
+ 0.056781,
+ 0.054826,
+ 0.054895,
+ 0.054896,
+ 0.054407,
+ 0.054407,
+ 0.054477,
+ 0.054476,
+ 0.054477,
+ 0.055454,
+ 0.055873,
+ 0.054476,
+ 0.055384,
+ 0.054477,
+ 0.054896,
+ 0.053988,
+ 0.055873,
+ 0.060623,
+ 0.053918,
+ 0.054826,
+ 0.054895,
+ 0.054406,
+ 0.054895,
+ 0.053847,
+ 0.054895,
+ 0.055384,
+ 0.055314,
+ 0.055384,
+ 0.054057,
+ 0.055384,
+ 0.054407,
+ 0.054896,
+ 0.054896,
+ 0.054337
+ ],
+ "warm_rows_per_second": 9326811.851608966
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772440363845078,
+ "character_perplexity": 6.535468506423138,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 10654345,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1349,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1349,
+ "zlib_blocks": 1349,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "table_sha256": "8480d0c01978ff0fa6f93d6530acada4a8d45fc316a37a67164b316eef6833a2",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.9405129118822515,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 10655556,
+ "backend_with_key_registry_bytes": 11391663,
+ "backend_with_datrie_key_index_bytes": 11001716,
+ "full_model_serving_artifact_bytes": 12219661,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.215881,
+ "warm_p50_ms": 0.054826,
+ "warm_p95_ms": 0.05629205,
+ "warm_p99_ms": 0.06579252999999999,
+ "samples_ms": [
+ 0.056293,
+ 0.054336,
+ 0.055384,
+ 0.054895,
+ 0.053987,
+ 0.054895,
+ 0.055873,
+ 0.054407,
+ 0.054896,
+ 0.054896,
+ 0.054407,
+ 0.056292,
+ 0.055803,
+ 0.055803,
+ 0.054407,
+ 0.055873,
+ 0.054756,
+ 0.054337,
+ 0.054896,
+ 0.054895,
+ 0.054895,
+ 0.054895,
+ 0.054965,
+ 0.054476,
+ 0.055873,
+ 0.055384,
+ 0.054477,
+ 0.053987,
+ 0.054406,
+ 0.054406,
+ 0.071866,
+ 0.054337,
+ 0.054826,
+ 0.054477,
+ 0.053499,
+ 0.054476,
+ 0.054407,
+ 0.053429,
+ 0.054337,
+ 0.054826
+ ],
+ "warm_rows_per_second": 9338634.954218801
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772440363845078,
+ "character_perplexity": 6.535468506423138,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7473752,
+ "unique_rows": 29823,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 932,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 932,
+ "zlib_blocks": 932,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29823
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "8a35a6bb7b8b8445e7c478a445afbd7401a67bb01ff5e989208f27b0aa65dfa1",
+ "table_sha256": "573970332d020d8c40ce39f2bd889280932f6197833c1a35716ebef2cfc42ef9",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.7199056609533727,
+ "representatives_sha256": "f84ee038c6582695258bf598118e9c5dda39bfc6220d2752237739d1957b8452",
+ "structurally_merged_rows": 13329,
+ "backend_artifact_bytes": 7474959,
+ "backend_with_key_registry_bytes": 8211066,
+ "backend_with_datrie_key_index_bytes": 7821119,
+ "full_model_serving_artifact_bytes": 9039064,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9901374270621742,
+ "mean_cosine": 0.9997243918755482,
+ "p05_cosine": 0.9987015070238352,
+ "relative_l2_mean": 0.012410471547235846,
+ "relative_l2_p95": 0.05118710693197225,
+ "relative_l2_max": 0.14041193528604085,
+ "weighted_relative_l2_mean": 0.001980608544931826,
+ "weighted_relative_frobenius": 0.012203466418320411,
+ "weighted_mean_cosine": 0.9999605554841995,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9901374270621742,
+ "mean_cosine": 0.9997243918755482,
+ "p05_cosine": 0.9987015070238352,
+ "relative_l2_mean": 0.012410471547235846,
+ "relative_l2_p95": 0.05118710693197225,
+ "relative_l2_max": 0.14041193528604085,
+ "weighted_relative_l2_mean": 0.001980608544931826,
+ "weighted_relative_frobenius": 0.012203466418320411,
+ "weighted_mean_cosine": 0.9999605554841995,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.36228,
+ "warm_p50_ms": 0.054337,
+ "warm_p95_ms": 0.056781,
+ "warm_p99_ms": 0.05793146,
+ "samples_ms": [
+ 0.058667,
+ 0.056781,
+ 0.054826,
+ 0.054337,
+ 0.054477,
+ 0.055873,
+ 0.054895,
+ 0.05301,
+ 0.054476,
+ 0.053499,
+ 0.053429,
+ 0.056781,
+ 0.054826,
+ 0.054406,
+ 0.054896,
+ 0.053359,
+ 0.054407,
+ 0.053988,
+ 0.053988,
+ 0.053359,
+ 0.052939,
+ 0.055384,
+ 0.055803,
+ 0.054337,
+ 0.053987,
+ 0.054476,
+ 0.053987,
+ 0.053917,
+ 0.053009,
+ 0.053428,
+ 0.054825,
+ 0.05301,
+ 0.053499,
+ 0.053848,
+ 0.053429,
+ 0.054406,
+ 0.053359,
+ 0.053917,
+ 0.054476,
+ 0.054476
+ ],
+ "warm_rows_per_second": 9422676.997257853
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8771682882462368,
+ "character_perplexity": 6.534973475600116,
+ "next_character_accuracy": 0.49539947860757555,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -7.574526945309828e-05,
+ "mean_prediction_kl_nats": 9.795638979163037e-06,
+ "prediction_kl_p99_nats": 0.00010671248000819322,
+ "maximum_prediction_kl_nats": 0.01692303822047178,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": 0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.37192267179489136,
+ "maximum_per_target_cross_entropy_increase": 0.18367242813110352
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -7.574526945309828e-05,
+ "mean_prediction_kl_nats": 9.795638979163037e-06,
+ "prediction_kl_p99_nats": 0.00010671248000819322,
+ "maximum_prediction_kl_nats": 0.01692303822047178,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": 0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.37192267179489136,
+ "maximum_per_target_cross_entropy_increase": 0.18367242813110352
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7468886,
+ "unique_rows": 29803,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 932,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 932,
+ "zlib_blocks": 932,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29803
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "4905a6fad18e32c4059a82c759792fae16d46000850d13902df0df023d5dbe7e",
+ "table_sha256": "d8c3b793fc309483474b6defe53c3e1ae02c927f98806219fdfa1163c586939b",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6770146200433373,
+ "representatives_sha256": "a34e8833b8f4861f00f182ac6f1b04124f83345e2b1b6930eb5030e02ba0e016",
+ "structurally_merged_rows": 13349,
+ "backend_artifact_bytes": 7470093,
+ "backend_with_key_registry_bytes": 8206200,
+ "backend_with_datrie_key_index_bytes": 7816253,
+ "full_model_serving_artifact_bytes": 9034198,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9811231151394353,
+ "mean_cosine": 0.9997164520711441,
+ "p05_cosine": 0.9986965485510646,
+ "relative_l2_mean": 0.012499865036091668,
+ "relative_l2_p95": 0.051271698632055766,
+ "relative_l2_max": 0.1939953608517965,
+ "weighted_relative_l2_mean": 0.0020205899641472323,
+ "weighted_relative_frobenius": 0.01315867579936367,
+ "weighted_mean_cosine": 0.9999569719651504,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9811231151394353,
+ "mean_cosine": 0.9997164520711441,
+ "p05_cosine": 0.9986965485510646,
+ "relative_l2_mean": 0.012499865036091668,
+ "relative_l2_p95": 0.051271698632055766,
+ "relative_l2_max": 0.1939953608517965,
+ "weighted_relative_l2_mean": 0.0020205899641472323,
+ "weighted_relative_frobenius": 0.01315867579936367,
+ "weighted_mean_cosine": 0.9999569719651504,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.567194,
+ "warm_p50_ms": 0.055838,
+ "warm_p95_ms": 0.0572035,
+ "warm_p99_ms": 0.05841985,
+ "samples_ms": [
+ 0.059155,
+ 0.05727,
+ 0.055873,
+ 0.055384,
+ 0.056362,
+ 0.055454,
+ 0.056781,
+ 0.056851,
+ 0.05685,
+ 0.056292,
+ 0.056292,
+ 0.056293,
+ 0.055384,
+ 0.056292,
+ 0.055873,
+ 0.055384,
+ 0.054895,
+ 0.056362,
+ 0.056781,
+ 0.055314,
+ 0.056292,
+ 0.055314,
+ 0.055873,
+ 0.055874,
+ 0.055314,
+ 0.055733,
+ 0.055314,
+ 0.055314,
+ 0.056712,
+ 0.055315,
+ 0.054825,
+ 0.054756,
+ 0.053917,
+ 0.055873,
+ 0.055244,
+ 0.055803,
+ 0.055385,
+ 0.055315,
+ 0.0572,
+ 0.054965
+ ],
+ "warm_rows_per_second": 9169382.857552204
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8771938659447684,
+ "character_perplexity": 6.535140627319256,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -5.016918122395854e-05,
+ "mean_prediction_kl_nats": 2.704286829560198e-05,
+ "prediction_kl_p99_nats": 0.00012042148487811871,
+ "maximum_prediction_kl_nats": 0.030118161308665473,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.45440253615379333,
+ "maximum_per_target_cross_entropy_increase": 0.4206380844116211
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -5.016918122395854e-05,
+ "mean_prediction_kl_nats": 2.704286829560198e-05,
+ "prediction_kl_p99_nats": 0.00012042148487811871,
+ "maximum_prediction_kl_nats": 0.030118161308665473,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.45440253615379333,
+ "maximum_per_target_cross_entropy_increase": 0.4206380844116211
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7467319,
+ "unique_rows": 29796,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 932,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 932,
+ "zlib_blocks": 932,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29796
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "76e59af330807446eb3388016584f5b1ffe5b9530a0dc641f5cfadd3b45bf019",
+ "table_sha256": "55c5d6bc568d7c82b224a6314e8d4e023289730c725eaa47455202b3ca76fc4a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6827592109329998,
+ "representatives_sha256": "cab20745ce73829806d07741efde1373edcc1f289f171b2fb91379dfa4cdcc3d",
+ "structurally_merged_rows": 13356,
+ "backend_artifact_bytes": 7468526,
+ "backend_with_key_registry_bytes": 8204633,
+ "backend_with_datrie_key_index_bytes": 7814686,
+ "full_model_serving_artifact_bytes": 9032631,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9701550536775938,
+ "mean_cosine": 0.999708000910412,
+ "p05_cosine": 0.9986934950654631,
+ "relative_l2_mean": 0.012570300126435703,
+ "relative_l2_p95": 0.05136065697318756,
+ "relative_l2_max": 0.24360219263104912,
+ "weighted_relative_l2_mean": 0.002048947266653511,
+ "weighted_relative_frobenius": 0.013740653481808742,
+ "weighted_mean_cosine": 0.9999537070465959,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9701550536775938,
+ "mean_cosine": 0.999708000910412,
+ "p05_cosine": 0.9986934950654631,
+ "relative_l2_mean": 0.012570300126435703,
+ "relative_l2_p95": 0.05136065697318756,
+ "relative_l2_max": 0.24360219263104912,
+ "weighted_relative_l2_mean": 0.002048947266653511,
+ "weighted_relative_frobenius": 0.013740653481808742,
+ "weighted_mean_cosine": 0.9999537070465959,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.579905,
+ "warm_p50_ms": 0.055384,
+ "warm_p95_ms": 0.05738524999999999,
+ "warm_p99_ms": 0.06937342999999999,
+ "samples_ms": [
+ 0.059575,
+ 0.05727,
+ 0.056362,
+ 0.055314,
+ 0.054336,
+ 0.056781,
+ 0.055873,
+ 0.055873,
+ 0.055454,
+ 0.055873,
+ 0.055873,
+ 0.056362,
+ 0.056781,
+ 0.05727,
+ 0.056781,
+ 0.054896,
+ 0.054477,
+ 0.054966,
+ 0.054966,
+ 0.055384,
+ 0.054895,
+ 0.075638,
+ 0.054895,
+ 0.053987,
+ 0.055384,
+ 0.054895,
+ 0.053917,
+ 0.053917,
+ 0.053917,
+ 0.054476,
+ 0.053987,
+ 0.054407,
+ 0.055454,
+ 0.054965,
+ 0.054406,
+ 0.055803,
+ 0.055384,
+ 0.054895,
+ 0.055873,
+ 0.055454
+ ],
+ "warm_rows_per_second": 9244547.161635129
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.877203405332208,
+ "character_perplexity": 6.535202968855021,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.063022686975546e-05,
+ "mean_prediction_kl_nats": 2.8766093224460522e-05,
+ "prediction_kl_p99_nats": 0.00013629267134794706,
+ "maximum_prediction_kl_nats": 0.030118161308665473,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.5846255421638489,
+ "maximum_per_target_cross_entropy_increase": 0.4206380844116211
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.063022686975546e-05,
+ "mean_prediction_kl_nats": 2.8766093224460522e-05,
+ "prediction_kl_p99_nats": 0.00013629267134794706,
+ "maximum_prediction_kl_nats": 0.030118161308665473,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.5846255421638489,
+ "maximum_per_target_cross_entropy_increase": 0.4206380844116211
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7462811,
+ "unique_rows": 29777,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 931,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 931,
+ "zlib_blocks": 931,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29777
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "3331c4ea80f85f10466abdaab1952bd3e75c40cb1a5f35b4f90f614397cf8432",
+ "table_sha256": "6367f0e13a691b1f6c9774871b1373af8ed7f578f39cc3bc277f2594368f6437",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6852218171115965,
+ "representatives_sha256": "44d86fc4452c21f6da8ffbcc5e605d507676bb84feeeeeb1b86615155177f01b",
+ "structurally_merged_rows": 13375,
+ "backend_artifact_bytes": 7464018,
+ "backend_with_key_registry_bytes": 8200125,
+ "backend_with_datrie_key_index_bytes": 7810178,
+ "full_model_serving_artifact_bytes": 9028123,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9510951584267655,
+ "mean_cosine": 0.9996851725896819,
+ "p05_cosine": 0.9986878807415716,
+ "relative_l2_mean": 0.012730953354210144,
+ "relative_l2_p95": 0.05149837886632302,
+ "relative_l2_max": 0.313052181227116,
+ "weighted_relative_l2_mean": 0.002212216378153698,
+ "weighted_relative_frobenius": 0.019003876776819983,
+ "weighted_mean_cosine": 0.9999304818236316,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9510951584267655,
+ "mean_cosine": 0.9996851725896819,
+ "p05_cosine": 0.9986878807415716,
+ "relative_l2_mean": 0.012730953354210144,
+ "relative_l2_p95": 0.05149837886632302,
+ "relative_l2_max": 0.313052181227116,
+ "weighted_relative_l2_mean": 0.002212216378153698,
+ "weighted_relative_frobenius": 0.019003876776819983,
+ "weighted_mean_cosine": 0.9999304818236316,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.525079,
+ "warm_p50_ms": 0.0548605,
+ "warm_p95_ms": 0.05589825,
+ "warm_p99_ms": 0.06893617999999999,
+ "samples_ms": [
+ 0.057689,
+ 0.055804,
+ 0.054895,
+ 0.055384,
+ 0.054895,
+ 0.055454,
+ 0.055384,
+ 0.054406,
+ 0.053987,
+ 0.055384,
+ 0.054336,
+ 0.053428,
+ 0.053987,
+ 0.054406,
+ 0.053499,
+ 0.054476,
+ 0.055384,
+ 0.053429,
+ 0.054337,
+ 0.053499,
+ 0.054895,
+ 0.054407,
+ 0.054477,
+ 0.055385,
+ 0.054966,
+ 0.053918,
+ 0.054337,
+ 0.054826,
+ 0.055384,
+ 0.053498,
+ 0.055314,
+ 0.054406,
+ 0.054895,
+ 0.054477,
+ 0.054407,
+ 0.055803,
+ 0.054896,
+ 0.076127,
+ 0.055384,
+ 0.054895
+ ],
+ "warm_rows_per_second": 9332762.187730698
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8769014268564972,
+ "character_perplexity": 6.533229776169881,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0003425508440682324,
+ "mean_prediction_kl_nats": 4.486066248487181e-05,
+ "prediction_kl_p99_nats": 0.00028717883435742936,
+ "maximum_prediction_kl_nats": 0.030118161308665473,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.9886711835861206,
+ "maximum_per_target_cross_entropy_increase": 0.4206380844116211
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.0003425508440682324,
+ "mean_prediction_kl_nats": 4.486066248487181e-05,
+ "prediction_kl_p99_nats": 0.00028717883435742936,
+ "maximum_prediction_kl_nats": 0.030118161308665473,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.9886711835861206,
+ "maximum_per_target_cross_entropy_increase": 0.4206380844116211
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7454598,
+ "unique_rows": 29742,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 930,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 930,
+ "zlib_blocks": 930,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29742
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "9238f59bb5d2e8d5c0cbe6fed8ee9ae2a6a730e6dfda6a2239dfef8ad277d2ba",
+ "table_sha256": "51d61f3f0511a39080d1da0564e367a9ed2814b1f1eaed94ef2771056c7687bd",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6819237689487636,
+ "representatives_sha256": "064b7af427b129290dc557791b38db34694a99a1a4dbab35ebef74e4b4d53644",
+ "structurally_merged_rows": 13410,
+ "backend_artifact_bytes": 7455805,
+ "backend_with_key_registry_bytes": 8191912,
+ "backend_with_datrie_key_index_bytes": 7801965,
+ "full_model_serving_artifact_bytes": 9019910,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9029576241944695,
+ "mean_cosine": 0.9996150002872196,
+ "p05_cosine": 0.9986763907676827,
+ "relative_l2_mean": 0.013093520889840756,
+ "relative_l2_p95": 0.05173769988680542,
+ "relative_l2_max": 0.4406349148093082,
+ "weighted_relative_l2_mean": 0.002310199870706357,
+ "weighted_relative_frobenius": 0.022844354660961463,
+ "weighted_mean_cosine": 0.9999118104322755,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9029576241944695,
+ "mean_cosine": 0.9996150002872196,
+ "p05_cosine": 0.9986763907676827,
+ "relative_l2_mean": 0.013093520889840756,
+ "relative_l2_p95": 0.05173769988680542,
+ "relative_l2_max": 0.4406349148093082,
+ "weighted_relative_l2_mean": 0.002310199870706357,
+ "weighted_relative_frobenius": 0.022844354660961463,
+ "weighted_mean_cosine": 0.9999118104322755,
+ "max_relative_norm_change": 0.010088294764666801
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.67461,
+ "warm_p50_ms": 0.0544415,
+ "warm_p95_ms": 0.05587305,
+ "warm_p99_ms": 0.057023239999999996,
+ "samples_ms": [
+ 0.057758,
+ 0.055384,
+ 0.055384,
+ 0.055384,
+ 0.055454,
+ 0.055873,
+ 0.055385,
+ 0.055874,
+ 0.05294,
+ 0.054407,
+ 0.054895,
+ 0.054337,
+ 0.053429,
+ 0.053917,
+ 0.055384,
+ 0.055384,
+ 0.054896,
+ 0.05308,
+ 0.054895,
+ 0.054336,
+ 0.054825,
+ 0.05294,
+ 0.054476,
+ 0.053428,
+ 0.054965,
+ 0.055384,
+ 0.053917,
+ 0.054336,
+ 0.053429,
+ 0.054407,
+ 0.054896,
+ 0.05308,
+ 0.053918,
+ 0.053499,
+ 0.053428,
+ 0.054895,
+ 0.054895,
+ 0.053498,
+ 0.053987,
+ 0.053429
+ ],
+ "warm_rows_per_second": 9404590.24824812
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8769683562712904,
+ "character_perplexity": 6.533667056048789,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.00027564211694667493,
+ "mean_prediction_kl_nats": 7.812557434837714e-05,
+ "prediction_kl_p99_nats": 0.0004873318513749142,
+ "maximum_prediction_kl_nats": 0.10622121741628762,
+ "top1_changed_fraction": 0.0016101824873485662,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 1.421809434890747,
+ "maximum_per_target_cross_entropy_increase": 0.9145984649658203
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.00027564211694667493,
+ "mean_prediction_kl_nats": 7.812557434837714e-05,
+ "prediction_kl_p99_nats": 0.0004873318513749142,
+ "maximum_prediction_kl_nats": 0.10622121741628762,
+ "top1_changed_fraction": 0.0016101824873485662,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 1.421809434890747,
+ "maximum_per_target_cross_entropy_increase": 0.9145984649658203
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 128,
+ "storage": {
+ "artifact_bytes": 5523456,
+ "unique_rows": 43152,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "212e57b5fa9fcef9bbaf640d7bbb586eb6dec1a60e9e380099be530a0d30a4a2",
+ "table_sha256": "90c4d98b998a5e102a57ac4c4bbec828fa37973f1ce190c7e5052edf80b98b29",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.0683542201295495,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5524261,
+ "backend_with_key_registry_bytes": 6260368,
+ "backend_with_datrie_key_index_bytes": 5870421,
+ "full_model_serving_artifact_bytes": 7088366,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.999999952237527,
+ "mean_cosine": 0.9999999791989146,
+ "p05_cosine": 0.9999999715381719,
+ "relative_l2_mean": 0.00020643723533316717,
+ "relative_l2_p95": 0.00024222825928337357,
+ "relative_l2_max": 0.0003103547968601321,
+ "weighted_relative_l2_mean": 7.796723903184389e-05,
+ "weighted_relative_frobenius": 0.00020660422892998176,
+ "weighted_mean_cosine": 0.9999999922088666,
+ "max_relative_norm_change": 0.00017841050618007463
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.317848,
+ "warm_p50_ms": 0.139613,
+ "warm_p95_ms": 0.14421844999999994,
+ "warm_p99_ms": 0.17171321999999997,
+ "samples_ms": [
+ 0.161403,
+ 0.142406,
+ 0.139543,
+ 0.140032,
+ 0.14094,
+ 0.139054,
+ 0.140451,
+ 0.139124,
+ 0.139613,
+ 0.139543,
+ 0.139473,
+ 0.138635,
+ 0.140032,
+ 0.139543,
+ 0.140031,
+ 0.138565,
+ 0.139054,
+ 0.139613,
+ 0.139613,
+ 0.140101,
+ 0.138146,
+ 0.138146,
+ 0.140101,
+ 0.140451,
+ 0.139613,
+ 0.139124,
+ 0.138635,
+ 0.138076,
+ 0.178305,
+ 0.143314,
+ 0.140032,
+ 0.140102,
+ 0.141359,
+ 0.140102,
+ 0.140032,
+ 0.138635,
+ 0.140032,
+ 0.140521,
+ 0.139543,
+ 0.139123
+ ],
+ "warm_rows_per_second": 3667280.267596857
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.877244241302235,
+ "character_perplexity": 6.535469845656627,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.049177481175951e-07,
+ "mean_prediction_kl_nats": 4.309312088974973e-09,
+ "prediction_kl_p99_nats": 3.498371872602242e-08,
+ "maximum_prediction_kl_nats": 7.336224652209635e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001119852066040039,
+ "maximum_per_target_cross_entropy_increase": 0.0007524490356445312
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 5455215,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 675,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 675,
+ "zlib_blocks": 675,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "212e57b5fa9fcef9bbaf640d7bbb586eb6dec1a60e9e380099be530a0d30a4a2",
+ "table_sha256": "1d743e28d52c9816eba61c77df349b65b520ac69ed444b0daef85dd12dcce1d4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5009870349895209,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5456423,
+ "backend_with_key_registry_bytes": 6192530,
+ "backend_with_datrie_key_index_bytes": 5802583,
+ "full_model_serving_artifact_bytes": 7020528,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.999999952237527,
+ "mean_cosine": 0.9999999791989146,
+ "p05_cosine": 0.9999999715381719,
+ "relative_l2_mean": 0.00020643723533316717,
+ "relative_l2_p95": 0.00024222825928337357,
+ "relative_l2_max": 0.0003103547968601321,
+ "weighted_relative_l2_mean": 7.796723903184389e-05,
+ "weighted_relative_frobenius": 0.00020660422892998176,
+ "weighted_mean_cosine": 0.9999999922088666,
+ "max_relative_norm_change": 0.00017841050618007463
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.90355,
+ "warm_p50_ms": 0.1474005,
+ "warm_p95_ms": 0.1495797,
+ "warm_p99_ms": 0.17641738999999998,
+ "samples_ms": [
+ 0.149461,
+ 0.148553,
+ 0.147155,
+ 0.148553,
+ 0.147156,
+ 0.148133,
+ 0.147645,
+ 0.148553,
+ 0.149111,
+ 0.147644,
+ 0.148553,
+ 0.147575,
+ 0.147155,
+ 0.147156,
+ 0.147575,
+ 0.147644,
+ 0.147226,
+ 0.147156,
+ 0.146666,
+ 0.145759,
+ 0.147156,
+ 0.192134,
+ 0.151835,
+ 0.148133,
+ 0.148064,
+ 0.148064,
+ 0.146597,
+ 0.147575,
+ 0.145689,
+ 0.146597,
+ 0.146666,
+ 0.147645,
+ 0.147156,
+ 0.147225,
+ 0.146178,
+ 0.147155,
+ 0.146667,
+ 0.146177,
+ 0.147156,
+ 0.147575
+ ],
+ "warm_rows_per_second": 3473529.6013242835
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.877244241302235,
+ "character_perplexity": 6.535469845656627,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.049177481175951e-07,
+ "mean_prediction_kl_nats": 4.309312088974973e-09,
+ "prediction_kl_p99_nats": 3.498371872602242e-08,
+ "maximum_prediction_kl_nats": 7.336224652209635e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001119852066040039,
+ "maximum_per_target_cross_entropy_increase": 0.0007524490356445312
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 5455215,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 675,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 675,
+ "zlib_blocks": 675,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "212e57b5fa9fcef9bbaf640d7bbb586eb6dec1a60e9e380099be530a0d30a4a2",
+ "table_sha256": "1d743e28d52c9816eba61c77df349b65b520ac69ed444b0daef85dd12dcce1d4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6599411331117153,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5456422,
+ "backend_with_key_registry_bytes": 6192529,
+ "backend_with_datrie_key_index_bytes": 5802582,
+ "full_model_serving_artifact_bytes": 7020527,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.999999952237527,
+ "mean_cosine": 0.9999999791989146,
+ "p05_cosine": 0.9999999715381719,
+ "relative_l2_mean": 0.00020643723533316717,
+ "relative_l2_p95": 0.00024222825928337357,
+ "relative_l2_max": 0.0003103547968601321,
+ "weighted_relative_l2_mean": 7.796723903184389e-05,
+ "weighted_relative_frobenius": 0.00020660422892998176,
+ "weighted_mean_cosine": 0.9999999922088666,
+ "max_relative_norm_change": 0.00017841050618007463
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.925829,
+ "warm_p50_ms": 0.147575,
+ "warm_p95_ms": 0.16536259999999997,
+ "warm_p99_ms": 0.19463553999999997,
+ "samples_ms": [
+ 0.150439,
+ 0.148064,
+ 0.160775,
+ 0.1496,
+ 0.149041,
+ 0.21106,
+ 0.153791,
+ 0.148063,
+ 0.147645,
+ 0.148064,
+ 0.148553,
+ 0.147644,
+ 0.146248,
+ 0.165174,
+ 0.168946,
+ 0.149461,
+ 0.146667,
+ 0.147644,
+ 0.146597,
+ 0.145759,
+ 0.147644,
+ 0.147156,
+ 0.146248,
+ 0.147575,
+ 0.147156,
+ 0.147017,
+ 0.145689,
+ 0.146666,
+ 0.159378,
+ 0.148971,
+ 0.147574,
+ 0.147226,
+ 0.147575,
+ 0.146667,
+ 0.146667,
+ 0.146667,
+ 0.145689,
+ 0.147575,
+ 0.146178,
+ 0.146667
+ ],
+ "warm_rows_per_second": 3469422.327630018
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.877244241302235,
+ "character_perplexity": 6.535469845656627,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.049177481175951e-07,
+ "mean_prediction_kl_nats": 4.309312088974973e-09,
+ "prediction_kl_p99_nats": 3.498371872602242e-08,
+ "maximum_prediction_kl_nats": 7.336224652209635e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001119852066040039,
+ "maximum_per_target_cross_entropy_increase": 0.0007524490356445312
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 5455215,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 675,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 675,
+ "zlib_blocks": 675,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "212e57b5fa9fcef9bbaf640d7bbb586eb6dec1a60e9e380099be530a0d30a4a2",
+ "table_sha256": "1d743e28d52c9816eba61c77df349b65b520ac69ed444b0daef85dd12dcce1d4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5223842598497868,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5456422,
+ "backend_with_key_registry_bytes": 6192529,
+ "backend_with_datrie_key_index_bytes": 5802582,
+ "full_model_serving_artifact_bytes": 7020527,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.999999952237527,
+ "mean_cosine": 0.9999999791989146,
+ "p05_cosine": 0.9999999715381719,
+ "relative_l2_mean": 0.00020643723533316717,
+ "relative_l2_p95": 0.00024222825928337357,
+ "relative_l2_max": 0.0003103547968601321,
+ "weighted_relative_l2_mean": 7.796723903184389e-05,
+ "weighted_relative_frobenius": 0.00020660422892998176,
+ "weighted_mean_cosine": 0.9999999922088666,
+ "max_relative_norm_change": 0.00017841050618007463
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.927156,
+ "warm_p50_ms": 0.1472255,
+ "warm_p95_ms": 0.15281984999999995,
+ "warm_p99_ms": 0.17174175,
+ "samples_ms": [
+ 0.152324,
+ 0.14995,
+ 0.148553,
+ 0.148064,
+ 0.148064,
+ 0.147993,
+ 0.162241,
+ 0.151276,
+ 0.148552,
+ 0.148552,
+ 0.147156,
+ 0.147225,
+ 0.147156,
+ 0.147156,
+ 0.147644,
+ 0.147575,
+ 0.146597,
+ 0.146597,
+ 0.147226,
+ 0.146667,
+ 0.147645,
+ 0.147086,
+ 0.147156,
+ 0.146248,
+ 0.147156,
+ 0.147086,
+ 0.146667,
+ 0.147644,
+ 0.147645,
+ 0.145758,
+ 0.147575,
+ 0.147086,
+ 0.147645,
+ 0.177816,
+ 0.148552,
+ 0.146597,
+ 0.146667,
+ 0.147156,
+ 0.147156,
+ 0.147155
+ ],
+ "warm_rows_per_second": 3477658.4219445675
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.877244241302235,
+ "character_perplexity": 6.535469845656627,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.049177481175951e-07,
+ "mean_prediction_kl_nats": 4.309312088974973e-09,
+ "prediction_kl_p99_nats": 3.498371872602242e-08,
+ "maximum_prediction_kl_nats": 7.336224652209635e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001119852066040039,
+ "maximum_per_target_cross_entropy_increase": 0.0007524490356445312
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3882363,
+ "unique_rows": 29824,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29824
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "9640ced718266675406375e98d479e25432f8c5824aa00d1cbc9d9466572d157",
+ "table_sha256": "b8b2ee3f87c9066969a1eb6381b347b672ba6cb101bfc84a86e152bce94815eb",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.406329584075138,
+ "representatives_sha256": "4d39c273c7801f244bdc864f10b3b9cec557ac92c039ad051a7df02e48959e14",
+ "structurally_merged_rows": 13328,
+ "backend_artifact_bytes": 3883569,
+ "backend_with_key_registry_bytes": 4619676,
+ "backend_with_datrie_key_index_bytes": 4229729,
+ "full_model_serving_artifact_bytes": 5447674,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9901359988769479,
+ "mean_cosine": 0.9997243911150577,
+ "p05_cosine": 0.998701582408625,
+ "relative_l2_mean": 0.012552533723385375,
+ "relative_l2_p95": 0.051191088644676495,
+ "relative_l2_max": 0.1404184277793818,
+ "weighted_relative_l2_mean": 0.002046682131465092,
+ "weighted_relative_frobenius": 0.012205143139227746,
+ "weighted_mean_cosine": 0.9999605527530744,
+ "max_relative_norm_change": 0.010145906005057173
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9901350207672447,
+ "mean_cosine": 0.9997243975370762,
+ "p05_cosine": 0.9987013964404386,
+ "relative_l2_mean": 0.012410022358831409,
+ "relative_l2_p95": 0.051198839042667935,
+ "relative_l2_max": 0.14042520696577482,
+ "weighted_relative_l2_mean": 0.0019803323994328197,
+ "weighted_relative_frobenius": 0.012203608015228266,
+ "weighted_mean_cosine": 0.9999605594959479,
+ "max_relative_norm_change": 0.010085298327435002
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.270578,
+ "warm_p50_ms": 0.1476095,
+ "warm_p95_ms": 0.15102145,
+ "warm_p99_ms": 0.15583103,
+ "samples_ms": [
+ 0.150997,
+ 0.14953,
+ 0.151486,
+ 0.149531,
+ 0.148064,
+ 0.148063,
+ 0.149041,
+ 0.147645,
+ 0.147645,
+ 0.147086,
+ 0.147016,
+ 0.148553,
+ 0.148552,
+ 0.147644,
+ 0.147994,
+ 0.147156,
+ 0.148063,
+ 0.147086,
+ 0.147645,
+ 0.147574,
+ 0.146597,
+ 0.146667,
+ 0.147574,
+ 0.146248,
+ 0.146666,
+ 0.148134,
+ 0.146178,
+ 0.146667,
+ 0.158609,
+ 0.148622,
+ 0.146667,
+ 0.148063,
+ 0.147086,
+ 0.147575,
+ 0.147086,
+ 0.147645,
+ 0.147156,
+ 0.146666,
+ 0.146667,
+ 0.147156
+ ],
+ "warm_rows_per_second": 3468611.437610723
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.877168853361512,
+ "character_perplexity": 6.5349771686144935,
+ "next_character_accuracy": 0.49539947860757555,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -7.518019682302672e-05,
+ "mean_prediction_kl_nats": 9.795772424569903e-06,
+ "prediction_kl_p99_nats": 0.00010697166204515648,
+ "maximum_prediction_kl_nats": 0.016904514918556646,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": 0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.3721555471420288,
+ "maximum_per_target_cross_entropy_increase": 0.1838548183441162
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -7.538509912339956e-05,
+ "mean_prediction_kl_nats": 9.794082950592873e-06,
+ "prediction_kl_p99_nats": 0.00010678432447086124,
+ "maximum_prediction_kl_nats": 0.016922180537672565,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": 0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.37213319540023804,
+ "maximum_per_target_cross_entropy_increase": 0.1836996078491211
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3879963,
+ "unique_rows": 29804,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29804
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "1d0186d6610372f7e7cf9c0bdb919c93be38e0923ea8bebbb8c1e10e9693ecdf",
+ "table_sha256": "ba1abb815031f6d09a15c3d915199b9e5fb476c2d95d3f92f044aa5d5579f8c7",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4629533898551017,
+ "representatives_sha256": "13e2eaff298ad79ed375c10a562953bc213cf6eb7194aa52470ea8ba63509403",
+ "structurally_merged_rows": 13348,
+ "backend_artifact_bytes": 3881169,
+ "backend_with_key_registry_bytes": 4617276,
+ "backend_with_datrie_key_index_bytes": 4227329,
+ "full_model_serving_artifact_bytes": 5445274,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9811277047037894,
+ "mean_cosine": 0.9997164512755508,
+ "p05_cosine": 0.9986966362509793,
+ "relative_l2_mean": 0.012641831401540108,
+ "relative_l2_p95": 0.05127963928816258,
+ "relative_l2_max": 0.19396570003323146,
+ "weighted_relative_l2_mean": 0.0020866221878897153,
+ "weighted_relative_frobenius": 0.013160371895786201,
+ "weighted_mean_cosine": 0.9999569690400654,
+ "max_relative_norm_change": 0.010145906005057173
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9811243907892742,
+ "mean_cosine": 0.9997164577430483,
+ "p05_cosine": 0.9986972952572108,
+ "relative_l2_mean": 0.012499417060502187,
+ "relative_l2_p95": 0.051283627134803604,
+ "relative_l2_max": 0.19396372523399272,
+ "weighted_relative_l2_mean": 0.0020203178548211255,
+ "weighted_relative_frobenius": 0.013159117282672982,
+ "weighted_mean_cosine": 0.9999569754682313,
+ "max_relative_norm_change": 0.010085298327435002
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.178317,
+ "warm_p50_ms": 0.147575,
+ "warm_p95_ms": 0.1491109,
+ "warm_p99_ms": 0.1555935,
+ "samples_ms": [
+ 0.150439,
+ 0.148552,
+ 0.147575,
+ 0.147645,
+ 0.147575,
+ 0.148064,
+ 0.148553,
+ 0.148133,
+ 0.148622,
+ 0.149041,
+ 0.147645,
+ 0.147575,
+ 0.147155,
+ 0.147575,
+ 0.158889,
+ 0.148064,
+ 0.146178,
+ 0.146178,
+ 0.146667,
+ 0.146667,
+ 0.148064,
+ 0.146178,
+ 0.147575,
+ 0.145759,
+ 0.147645,
+ 0.147226,
+ 0.146248,
+ 0.146667,
+ 0.145759,
+ 0.147575,
+ 0.147156,
+ 0.147086,
+ 0.145689,
+ 0.147645,
+ 0.147085,
+ 0.146597,
+ 0.146178,
+ 0.146108,
+ 0.147575,
+ 0.147156
+ ],
+ "warm_rows_per_second": 3469422.327630018
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8771944360852009,
+ "character_perplexity": 6.535144353268222,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.9599069232447145e-05,
+ "mean_prediction_kl_nats": 2.7040531143815635e-05,
+ "prediction_kl_p99_nats": 0.0001212205948734834,
+ "maximum_prediction_kl_nats": 0.030108628146649247,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.4547811448574066,
+ "maximum_per_target_cross_entropy_increase": 0.42043161392211914
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.980397677484597e-05,
+ "mean_prediction_kl_nats": 2.7045709011130913e-05,
+ "prediction_kl_p99_nats": 0.0001210610821354744,
+ "maximum_prediction_kl_nats": 0.0301333351480087,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.4548949599266052,
+ "maximum_per_target_cross_entropy_increase": 0.4206087589263916
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3879148,
+ "unique_rows": 29797,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29797
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "137c54c851ee9aef9c2f59e5cee561bc162292c80f0e5da5cbb5b2bf4309a815",
+ "table_sha256": "7653a1ca0100a0c56b2af12aa58e4fa8dd0b6e891d4d809e7bce133c1becfc59",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3883936258498579,
+ "representatives_sha256": "60f10fc23f54f0ee0daca046aaba8f2642f7fb595eed113e2b98c6fc86038a5f",
+ "structurally_merged_rows": 13355,
+ "backend_artifact_bytes": 3880354,
+ "backend_with_key_registry_bytes": 4616461,
+ "backend_with_datrie_key_index_bytes": 4226514,
+ "full_model_serving_artifact_bytes": 5444459,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9701552429340661,
+ "mean_cosine": 0.9997080003704789,
+ "p05_cosine": 0.998692743504631,
+ "relative_l2_mean": 0.01271223511424447,
+ "relative_l2_p95": 0.05137353966363325,
+ "relative_l2_max": 0.24360749548388305,
+ "weighted_relative_l2_mean": 0.002114968820646803,
+ "weighted_relative_frobenius": 0.013742242501495193,
+ "weighted_mean_cosine": 0.9999537044720979,
+ "max_relative_norm_change": 0.010145906005057173
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9701453061361723,
+ "mean_cosine": 0.9997080063633795,
+ "p05_cosine": 0.9986928810695538,
+ "relative_l2_mean": 0.012569856943483957,
+ "relative_l2_p95": 0.0513781474701145,
+ "relative_l2_max": 0.24365314010630398,
+ "weighted_relative_l2_mean": 0.0020486727955327188,
+ "weighted_relative_frobenius": 0.013741028017087598,
+ "weighted_mean_cosine": 0.9999537109984574,
+ "max_relative_norm_change": 0.010085298327435002
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.226089,
+ "warm_p50_ms": 0.147156,
+ "warm_p95_ms": 0.16056194999999995,
+ "warm_p99_ms": 0.24260743999999992,
+ "samples_ms": [
+ 0.14995,
+ 0.148134,
+ 0.148134,
+ 0.147225,
+ 0.146737,
+ 0.148552,
+ 0.148133,
+ 0.148553,
+ 0.146248,
+ 0.148063,
+ 0.147156,
+ 0.147644,
+ 0.181727,
+ 0.151765,
+ 0.147644,
+ 0.147574,
+ 0.146667,
+ 0.147156,
+ 0.145759,
+ 0.146736,
+ 0.147156,
+ 0.146667,
+ 0.146248,
+ 0.147575,
+ 0.146178,
+ 0.146667,
+ 0.147156,
+ 0.146178,
+ 0.146597,
+ 0.146667,
+ 0.147086,
+ 0.147016,
+ 0.146178,
+ 0.147575,
+ 0.145759,
+ 0.147645,
+ 0.146667,
+ 0.145759,
+ 0.281531,
+ 0.159448
+ ],
+ "warm_rows_per_second": 3479300.877979831
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772039686603894,
+ "character_perplexity": 6.535206650320062,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.006692141793789e-05,
+ "mean_prediction_kl_nats": 2.8764179419291013e-05,
+ "prediction_kl_p99_nats": 0.0001373004646596293,
+ "maximum_prediction_kl_nats": 0.030108628146649247,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.5845257639884949,
+ "maximum_per_target_cross_entropy_increase": 0.42043161392211914
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.0271830913642586e-05,
+ "mean_prediction_kl_nats": 2.8768364376307013e-05,
+ "prediction_kl_p99_nats": 0.00013665960831430523,
+ "maximum_prediction_kl_nats": 0.0301333351480087,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.5844752788543701,
+ "maximum_per_target_cross_entropy_increase": 0.4206087589263916
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3876935,
+ "unique_rows": 29778,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29778
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "e434feffabfbc8116628e58406f42b8cc8207c53c795caf18a5f3dce23f1fbdc",
+ "table_sha256": "09dcfe5cdd90d9a99c06c8eb3588fd6cc523f743071e4fbc422417216eaed5f1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3968824229668826,
+ "representatives_sha256": "c5bc80eab280d3eae987b35d58945c2c403667bb392c94616c14def6b5313cfe",
+ "structurally_merged_rows": 13374,
+ "backend_artifact_bytes": 3878141,
+ "backend_with_key_registry_bytes": 4614248,
+ "backend_with_datrie_key_index_bytes": 4224301,
+ "full_model_serving_artifact_bytes": 5442246,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.95109107112859,
+ "mean_cosine": 0.9996851734688298,
+ "p05_cosine": 0.9986880269709896,
+ "relative_l2_mean": 0.012872793154547231,
+ "relative_l2_p95": 0.05149540003059486,
+ "relative_l2_max": 0.3130204723240637,
+ "weighted_relative_l2_mean": 0.002278146478787276,
+ "weighted_relative_frobenius": 0.019003676722386505,
+ "weighted_mean_cosine": 0.9999304849711296,
+ "max_relative_norm_change": 0.010145906005057173
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9511049942344023,
+ "mean_cosine": 0.9996851795755632,
+ "p05_cosine": 0.9986874967079526,
+ "relative_l2_mean": 0.012730502151193564,
+ "relative_l2_p95": 0.05150844694982633,
+ "relative_l2_max": 0.31304083713182396,
+ "weighted_relative_l2_mean": 0.0022119122255464248,
+ "weighted_relative_frobenius": 0.01900275506586697,
+ "weighted_mean_cosine": 0.9999304925324197,
+ "max_relative_norm_change": 0.010085298327435002
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.034513,
+ "warm_p50_ms": 0.1476445,
+ "warm_p95_ms": 0.15286874999999997,
+ "warm_p99_ms": 0.17421059,
+ "samples_ms": [
+ 0.151835,
+ 0.14946,
+ 0.149041,
+ 0.148552,
+ 0.148553,
+ 0.148553,
+ 0.148552,
+ 0.149041,
+ 0.147993,
+ 0.148553,
+ 0.148064,
+ 0.181238,
+ 0.152324,
+ 0.148972,
+ 0.147645,
+ 0.146736,
+ 0.147645,
+ 0.147156,
+ 0.147644,
+ 0.146667,
+ 0.147505,
+ 0.148063,
+ 0.147156,
+ 0.147156,
+ 0.146178,
+ 0.146248,
+ 0.147086,
+ 0.147575,
+ 0.146109,
+ 0.147645,
+ 0.147575,
+ 0.147156,
+ 0.146248,
+ 0.147155,
+ 0.147226,
+ 0.146178,
+ 0.146667,
+ 0.147644,
+ 0.163219,
+ 0.148552
+ ],
+ "warm_rows_per_second": 3467789.1828005784
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8769020492525026,
+ "character_perplexity": 6.5332338424272605,
+ "next_character_accuracy": 0.49524612789449474,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0003419286610716499,
+ "mean_prediction_kl_nats": 4.486075688066997e-05,
+ "prediction_kl_p99_nats": 0.00028680683746140797,
+ "maximum_prediction_kl_nats": 0.030108628146649247,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.9884049892425537,
+ "maximum_per_target_cross_entropy_increase": 0.42043161392211914
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.0003421335087105394,
+ "mean_prediction_kl_nats": 4.486310455467269e-05,
+ "prediction_kl_p99_nats": 0.00028727213267286655,
+ "maximum_prediction_kl_nats": 0.0301333351480087,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.9884847402572632,
+ "maximum_per_target_cross_entropy_increase": 0.4206087589263916
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3872672,
+ "unique_rows": 29743,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 465,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 465,
+ "zlib_blocks": 465,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29743
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "53ebd25388250f9b4d6fc0dea0b9513e4484d137ff218e9794f89bd8bef08e9b",
+ "table_sha256": "dd0a70a5f8029e1e7e09ceaf44f3084c81a74c19029ed8926cda58a2d3edc03d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4126182380132377,
+ "representatives_sha256": "8870b968e8cd56e4698498505b13e7f0d1419dd3859ad36a5917ef111bdcfc5a",
+ "structurally_merged_rows": 13409,
+ "backend_artifact_bytes": 3873878,
+ "backend_with_key_registry_bytes": 4609985,
+ "backend_with_datrie_key_index_bytes": 4220038,
+ "full_model_serving_artifact_bytes": 5437983,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9029568391232163,
+ "mean_cosine": 0.9996149997015428,
+ "p05_cosine": 0.9986758813379707,
+ "relative_l2_mean": 0.013235200765704624,
+ "relative_l2_p95": 0.05174014608774899,
+ "relative_l2_max": 0.4406441354927482,
+ "weighted_relative_l2_mean": 0.002376078832953385,
+ "weighted_relative_frobenius": 0.02284375735876039,
+ "weighted_mean_cosine": 0.9999118142079768,
+ "max_relative_norm_change": 0.010145906005057173
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9029504707467081,
+ "mean_cosine": 0.9996150045673875,
+ "p05_cosine": 0.9986755277169472,
+ "relative_l2_mean": 0.013093074009532956,
+ "relative_l2_p95": 0.05173732092541074,
+ "relative_l2_max": 0.4406531373114421,
+ "weighted_relative_l2_mean": 0.002309894378455299,
+ "weighted_relative_frobenius": 0.0228432212925786,
+ "weighted_mean_cosine": 0.9999118214769647,
+ "max_relative_norm_change": 0.010085298327435002
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.968235,
+ "warm_p50_ms": 0.1475745,
+ "warm_p95_ms": 0.15090589999999998,
+ "warm_p99_ms": 0.17322221,
+ "samples_ms": [
+ 0.150857,
+ 0.149042,
+ 0.148064,
+ 0.148553,
+ 0.148064,
+ 0.147086,
+ 0.147156,
+ 0.147505,
+ 0.147156,
+ 0.147156,
+ 0.148133,
+ 0.147156,
+ 0.147086,
+ 0.146667,
+ 0.147994,
+ 0.186896,
+ 0.150508,
+ 0.147645,
+ 0.147575,
+ 0.147994,
+ 0.147575,
+ 0.151835,
+ 0.146178,
+ 0.145759,
+ 0.147574,
+ 0.147086,
+ 0.147645,
+ 0.147644,
+ 0.147645,
+ 0.147156,
+ 0.146178,
+ 0.147645,
+ 0.146667,
+ 0.147156,
+ 0.147575,
+ 0.146667,
+ 0.146737,
+ 0.147644,
+ 0.146667,
+ 0.146178
+ ],
+ "warm_rows_per_second": 3469434.082446493
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8769689418982365,
+ "character_perplexity": 6.5336708823413945,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.00027505665125251234,
+ "mean_prediction_kl_nats": 7.810417882368735e-05,
+ "prediction_kl_p99_nats": 0.0004862708800758725,
+ "maximum_prediction_kl_nats": 0.10616205347563132,
+ "top1_changed_fraction": 0.0016101824873485662,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 1.4215319156646729,
+ "maximum_per_target_cross_entropy_increase": 0.9142379760742188
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.0002752615125946606,
+ "mean_prediction_kl_nats": 7.81205352925241e-05,
+ "prediction_kl_p99_nats": 0.0004863201053157216,
+ "maximum_prediction_kl_nats": 0.10619663094734044,
+ "top1_changed_fraction": 0.0016101824873485662,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 1.421834945678711,
+ "maximum_per_target_cross_entropy_increase": 0.9144940376281738
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 68,
+ "storage": {
+ "artifact_bytes": 2934336,
+ "unique_rows": 43152,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "2ac1cbfa580f67daaec9296dbc36834c852d35f0c2ac598a3062f05020c54a6d",
+ "table_sha256": "c6b56112cce3e1b0060ebf715a19f2aa46b58a9555a1fddb0277cd94bbec29c3",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.05933483806438744,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2935184,
+ "backend_with_key_registry_bytes": 3671291,
+ "backend_with_datrie_key_index_bytes": 3281344,
+ "full_model_serving_artifact_bytes": 4499289,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999211334377017,
+ "mean_cosine": 0.9999884214594844,
+ "p05_cosine": 0.9999770988080822,
+ "relative_l2_mean": 0.004703892852436545,
+ "relative_l2_p95": 0.006808280335029292,
+ "relative_l2_max": 0.012559680173836917,
+ "weighted_relative_l2_mean": 0.0020427094037580255,
+ "weighted_relative_frobenius": 0.005389228558747884,
+ "weighted_mean_cosine": 0.9999943438751884,
+ "max_relative_norm_change": 0.0031613003512570913
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.275314,
+ "warm_p50_ms": 0.09746350000000001,
+ "warm_p95_ms": 0.10044209999999996,
+ "warm_p99_ms": 0.13248015,
+ "samples_ms": [
+ 0.097987,
+ 0.097918,
+ 0.097918,
+ 0.09701,
+ 0.097428,
+ 0.097429,
+ 0.095962,
+ 0.097009,
+ 0.1452,
+ 0.112585,
+ 0.098895,
+ 0.099803,
+ 0.097918,
+ 0.097918,
+ 0.098407,
+ 0.097499,
+ 0.095682,
+ 0.096102,
+ 0.09701,
+ 0.09694,
+ 0.09708,
+ 0.097988,
+ 0.097009,
+ 0.09701,
+ 0.09701,
+ 0.097498,
+ 0.096521,
+ 0.096521,
+ 0.096939,
+ 0.097917,
+ 0.097499,
+ 0.097918,
+ 0.097499,
+ 0.098406,
+ 0.09701,
+ 0.097499,
+ 0.09701,
+ 0.097009,
+ 0.09694,
+ 0.098407
+ ],
+ "warm_rows_per_second": 5253248.652059489
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772194514216143,
+ "character_perplexity": 6.535307834147485,
+ "next_character_accuracy": 0.4953228032510351,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.458466068578227e-05,
+ "mean_prediction_kl_nats": 3.7257925306687268e-06,
+ "prediction_kl_p99_nats": 2.9096108127972065e-05,
+ "maximum_prediction_kl_nats": 0.00010730878408645676,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.04188680648803711,
+ "maximum_per_target_cross_entropy_increase": 0.021907806396484375
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 3288614,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 360,
+ "max_block_bytes": 8192,
+ "raw_blocks": 187,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "2ac1cbfa580f67daaec9296dbc36834c852d35f0c2ac598a3062f05020c54a6d",
+ "table_sha256": "c1f6b7272b67c0a95e089109a7f335761d0d41f274542bf2a0963570fd069540",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2841149701271206,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3289867,
+ "backend_with_key_registry_bytes": 4025974,
+ "backend_with_datrie_key_index_bytes": 3636027,
+ "full_model_serving_artifact_bytes": 4853972,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999211334377017,
+ "mean_cosine": 0.9999884214594844,
+ "p05_cosine": 0.9999770988080822,
+ "relative_l2_mean": 0.004703892852436545,
+ "relative_l2_p95": 0.006808280335029292,
+ "relative_l2_max": 0.012559680173836917,
+ "weighted_relative_l2_mean": 0.0020427094037580255,
+ "weighted_relative_frobenius": 0.005389228558747884,
+ "weighted_mean_cosine": 0.9999943438751884,
+ "max_relative_norm_change": 0.0031613003512570913
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.226962,
+ "warm_p50_ms": 0.112165,
+ "warm_p95_ms": 0.11365624999999999,
+ "warm_p99_ms": 0.12324402,
+ "samples_ms": [
+ 0.115447,
+ 0.112654,
+ 0.113073,
+ 0.112095,
+ 0.112165,
+ 0.112585,
+ 0.112585,
+ 0.112584,
+ 0.113562,
+ 0.113073,
+ 0.112654,
+ 0.112166,
+ 0.112655,
+ 0.111677,
+ 0.112096,
+ 0.11021,
+ 0.112655,
+ 0.110839,
+ 0.112166,
+ 0.128229,
+ 0.111606,
+ 0.112165,
+ 0.111676,
+ 0.111677,
+ 0.111676,
+ 0.112585,
+ 0.110699,
+ 0.11021,
+ 0.112235,
+ 0.112585,
+ 0.112654,
+ 0.110699,
+ 0.111118,
+ 0.110699,
+ 0.113073,
+ 0.111606,
+ 0.110768,
+ 0.11021,
+ 0.112096,
+ 0.112096
+ ],
+ "warm_rows_per_second": 4564703.784603041
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772194514216143,
+ "character_perplexity": 6.535307834147485,
+ "next_character_accuracy": 0.4953228032510351,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.458466068578227e-05,
+ "mean_prediction_kl_nats": 3.7257925306687268e-06,
+ "prediction_kl_p99_nats": 2.9096108127972065e-05,
+ "maximum_prediction_kl_nats": 0.00010730878408645676,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.04188680648803711,
+ "maximum_per_target_cross_entropy_increase": 0.021907806396484375
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 3288614,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 360,
+ "max_block_bytes": 8192,
+ "raw_blocks": 187,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "2ac1cbfa580f67daaec9296dbc36834c852d35f0c2ac598a3062f05020c54a6d",
+ "table_sha256": "c1f6b7272b67c0a95e089109a7f335761d0d41f274542bf2a0963570fd069540",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4498774129897356,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3289866,
+ "backend_with_key_registry_bytes": 4025973,
+ "backend_with_datrie_key_index_bytes": 3636026,
+ "full_model_serving_artifact_bytes": 4853971,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999211334377017,
+ "mean_cosine": 0.9999884214594844,
+ "p05_cosine": 0.9999770988080822,
+ "relative_l2_mean": 0.004703892852436545,
+ "relative_l2_p95": 0.006808280335029292,
+ "relative_l2_max": 0.012559680173836917,
+ "weighted_relative_l2_mean": 0.0020427094037580255,
+ "weighted_relative_frobenius": 0.005389228558747884,
+ "weighted_mean_cosine": 0.9999943438751884,
+ "max_relative_norm_change": 0.0031613003512570913
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.159285,
+ "warm_p50_ms": 0.11171149999999999,
+ "warm_p95_ms": 0.1150044,
+ "warm_p99_ms": 0.13465499999999997,
+ "samples_ms": [
+ 0.114959,
+ 0.113003,
+ 0.112514,
+ 0.113562,
+ 0.112026,
+ 0.146667,
+ 0.115867,
+ 0.113073,
+ 0.111746,
+ 0.112515,
+ 0.111118,
+ 0.110699,
+ 0.112654,
+ 0.111188,
+ 0.112096,
+ 0.110629,
+ 0.11021,
+ 0.112095,
+ 0.112165,
+ 0.110769,
+ 0.112165,
+ 0.112515,
+ 0.110629,
+ 0.112654,
+ 0.111677,
+ 0.111257,
+ 0.113073,
+ 0.111188,
+ 0.110699,
+ 0.110699,
+ 0.112165,
+ 0.11021,
+ 0.110768,
+ 0.111677,
+ 0.111257,
+ 0.111257,
+ 0.112095,
+ 0.111606,
+ 0.110768,
+ 0.111257
+ ],
+ "warm_rows_per_second": 4583234.492420208
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772194514216143,
+ "character_perplexity": 6.535307834147485,
+ "next_character_accuracy": 0.4953228032510351,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.458466068578227e-05,
+ "mean_prediction_kl_nats": 3.7257925306687268e-06,
+ "prediction_kl_p99_nats": 2.9096108127972065e-05,
+ "maximum_prediction_kl_nats": 0.00010730878408645676,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.04188680648803711,
+ "maximum_per_target_cross_entropy_increase": 0.021907806396484375
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 3288614,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 360,
+ "max_block_bytes": 8192,
+ "raw_blocks": 187,
+ "compressed_blocks": 173,
+ "zlib_blocks": 173,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "2ac1cbfa580f67daaec9296dbc36834c852d35f0c2ac598a3062f05020c54a6d",
+ "table_sha256": "c1f6b7272b67c0a95e089109a7f335761d0d41f274542bf2a0963570fd069540",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2968744949903339,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3289866,
+ "backend_with_key_registry_bytes": 4025973,
+ "backend_with_datrie_key_index_bytes": 3636026,
+ "full_model_serving_artifact_bytes": 4853971,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999211334377017,
+ "mean_cosine": 0.9999884214594844,
+ "p05_cosine": 0.9999770988080822,
+ "relative_l2_mean": 0.004703892852436545,
+ "relative_l2_p95": 0.006808280335029292,
+ "relative_l2_max": 0.012559680173836917,
+ "weighted_relative_l2_mean": 0.0020427094037580255,
+ "weighted_relative_frobenius": 0.005389228558747884,
+ "weighted_mean_cosine": 0.9999943438751884,
+ "max_relative_norm_change": 0.0031613003512570913
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.164453,
+ "warm_p50_ms": 0.111607,
+ "warm_p95_ms": 0.11356535,
+ "warm_p99_ms": 0.12420416,
+ "samples_ms": [
+ 0.114959,
+ 0.113143,
+ 0.113073,
+ 0.113073,
+ 0.110629,
+ 0.113004,
+ 0.112584,
+ 0.113003,
+ 0.113492,
+ 0.113073,
+ 0.111676,
+ 0.112654,
+ 0.111607,
+ 0.111607,
+ 0.111677,
+ 0.111188,
+ 0.110279,
+ 0.111607,
+ 0.110699,
+ 0.11021,
+ 0.111676,
+ 0.110768,
+ 0.110629,
+ 0.111257,
+ 0.111676,
+ 0.111676,
+ 0.111606,
+ 0.112095,
+ 0.110768,
+ 0.11028,
+ 0.111118,
+ 0.111607,
+ 0.130115,
+ 0.11021,
+ 0.110768,
+ 0.111606,
+ 0.111187,
+ 0.110698,
+ 0.108813,
+ 0.112096
+ ],
+ "warm_rows_per_second": 4587525.872033116
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772194514216143,
+ "character_perplexity": 6.535307834147485,
+ "next_character_accuracy": 0.4953228032510351,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -2.458466068578227e-05,
+ "mean_prediction_kl_nats": 3.7257925306687268e-06,
+ "prediction_kl_p99_nats": 2.9096108127972065e-05,
+ "maximum_prediction_kl_nats": 0.00010730878408645676,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.04188680648803711,
+ "maximum_per_target_cross_entropy_increase": 0.021907806396484375
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2379583,
+ "unique_rows": 29847,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 71,
+ "compressed_blocks": 178,
+ "zlib_blocks": 178,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29847
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "6d66c0bfd5543334bcb73d019de8b8b4b1b38fd95f9880f4375fb90ea837117c",
+ "table_sha256": "2e10776d273d205120149583a3e4758e4a4912fc03dbce073a78c5e772eef25e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.23807862401008606,
+ "representatives_sha256": "c5951879742dcb40fbdb0518d2f79b502806f8e52e309338331414737bac02e5",
+ "structurally_merged_rows": 13305,
+ "backend_artifact_bytes": 2380833,
+ "backend_with_key_registry_bytes": 3116940,
+ "backend_with_datrie_key_index_bytes": 2726993,
+ "full_model_serving_artifact_bytes": 3944938,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9901475114542572,
+ "mean_cosine": 0.9997133396421275,
+ "p05_cosine": 0.9986894419312247,
+ "relative_l2_mean": 0.015794454168486457,
+ "relative_l2_p95": 0.05149892399538544,
+ "relative_l2_max": 0.14034170133988702,
+ "weighted_relative_l2_mean": 0.003753597272702232,
+ "weighted_relative_frobenius": 0.0133093761146089,
+ "weighted_mean_cosine": 0.9999552172022823,
+ "max_relative_norm_change": 0.01184111110318379
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.990135508727166,
+ "mean_cosine": 0.9997185262944867,
+ "p05_cosine": 0.9986738463742685,
+ "relative_l2_mean": 0.012552149296026199,
+ "relative_l2_p95": 0.051718263226385855,
+ "relative_l2_max": 0.1404126776934458,
+ "weighted_relative_l2_mean": 0.0020091129867621264,
+ "weighted_relative_frobenius": 0.01244834613064305,
+ "weighted_mean_cosine": 0.9999594602328237,
+ "max_relative_norm_change": 0.010095277078878997
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.017077,
+ "warm_p50_ms": 0.108778,
+ "warm_p95_ms": 0.11601279999999996,
+ "warm_p99_ms": 0.12795555,
+ "samples_ms": [
+ 0.113492,
+ 0.111117,
+ 0.110349,
+ 0.128718,
+ 0.109302,
+ 0.110698,
+ 0.11028,
+ 0.109721,
+ 0.107416,
+ 0.108813,
+ 0.10979,
+ 0.108743,
+ 0.108324,
+ 0.108324,
+ 0.10979,
+ 0.108324,
+ 0.107416,
+ 0.108324,
+ 0.107905,
+ 0.107905,
+ 0.108394,
+ 0.109302,
+ 0.108883,
+ 0.115447,
+ 0.109721,
+ 0.109302,
+ 0.109232,
+ 0.107905,
+ 0.108394,
+ 0.108393,
+ 0.107835,
+ 0.109302,
+ 0.107835,
+ 0.107835,
+ 0.107835,
+ 0.110698,
+ 0.108394,
+ 0.108324,
+ 0.107905,
+ 0.126763
+ ],
+ "warm_rows_per_second": 4706834.102483958
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.877132831222215,
+ "character_perplexity": 6.534741768996447,
+ "next_character_accuracy": 0.49562950467719674,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.00011119897922794746,
+ "mean_prediction_kl_nats": 1.3663462342206782e-05,
+ "prediction_kl_p99_nats": 0.00011834122541509815,
+ "maximum_prediction_kl_nats": 0.01791453215246488,
+ "top1_changed_fraction": 0.002376936052752645,
+ "accuracy_change": 0.0004600521392424395,
+ "maximum_absolute_logit_change": 0.3684694766998291,
+ "maximum_per_target_cross_entropy_increase": 0.17836976051330566
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -8.661644797814854e-05,
+ "mean_prediction_kl_nats": 9.926255770447644e-06,
+ "prediction_kl_p99_nats": 0.00011290883215839557,
+ "maximum_prediction_kl_nats": 0.01800184768005272,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.3697659969329834,
+ "maximum_per_target_cross_entropy_increase": 0.18305516242980957
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2378155,
+ "unique_rows": 29827,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 71,
+ "compressed_blocks": 178,
+ "zlib_blocks": 178,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29827
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "b729a52ed75f57d8d872a3a339100a2bf16cc4392b747efe0a684192c9635936",
+ "table_sha256": "a60156342c173dde8c6dcb645547d787177fa54a3e90961c78478c2af5728767",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25508715701289475,
+ "representatives_sha256": "5b810087fc04119dec569a0ed5e0d67ffe99962113fca6aec0bbb3278a241f73",
+ "structurally_merged_rows": 13325,
+ "backend_artifact_bytes": 2379405,
+ "backend_with_key_registry_bytes": 3115512,
+ "backend_with_datrie_key_index_bytes": 2725565,
+ "full_model_serving_artifact_bytes": 3943510,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9810367637945125,
+ "mean_cosine": 0.9997053753890165,
+ "p05_cosine": 0.9986822306235927,
+ "relative_l2_mean": 0.015881604219005543,
+ "relative_l2_p95": 0.051603757393284606,
+ "relative_l2_max": 0.19497180693208166,
+ "weighted_relative_l2_mean": 0.0037924838223366837,
+ "weighted_relative_frobenius": 0.014189580581151816,
+ "weighted_mean_cosine": 0.9999516331953032,
+ "max_relative_norm_change": 0.01184111110318379
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9810632185328769,
+ "mean_cosine": 0.9997105512825292,
+ "p05_cosine": 0.9986681096578764,
+ "relative_l2_mean": 0.012641758762971559,
+ "relative_l2_p95": 0.05181365555230398,
+ "relative_l2_max": 0.194871458842179,
+ "weighted_relative_l2_mean": 0.0020492382753159143,
+ "weighted_relative_frobenius": 0.013392380945672376,
+ "weighted_mean_cosine": 0.9999558576885209,
+ "max_relative_norm_change": 0.010095277078878997
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.125889,
+ "warm_p50_ms": 0.106927,
+ "warm_p95_ms": 0.1098994,
+ "warm_p99_ms": 0.11791423,
+ "samples_ms": [
+ 0.110629,
+ 0.107835,
+ 0.107835,
+ 0.106927,
+ 0.108324,
+ 0.109861,
+ 0.108743,
+ 0.108813,
+ 0.106857,
+ 0.107835,
+ 0.106438,
+ 0.106508,
+ 0.106927,
+ 0.106997,
+ 0.106927,
+ 0.106997,
+ 0.106019,
+ 0.107416,
+ 0.107835,
+ 0.10553,
+ 0.122572,
+ 0.108394,
+ 0.107416,
+ 0.106857,
+ 0.106927,
+ 0.107346,
+ 0.10602,
+ 0.105112,
+ 0.106927,
+ 0.106019,
+ 0.106927,
+ 0.106438,
+ 0.106369,
+ 0.107416,
+ 0.106927,
+ 0.10602,
+ 0.10595,
+ 0.106508,
+ 0.106927,
+ 0.107486
+ ],
+ "warm_rows_per_second": 4788313.522309613
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8771582425836568,
+ "character_perplexity": 6.534907827791349,
+ "next_character_accuracy": 0.49547615396411593,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -8.579012066808811e-05,
+ "mean_prediction_kl_nats": 3.137851799195357e-05,
+ "prediction_kl_p99_nats": 0.00013102815635868638,
+ "maximum_prediction_kl_nats": 0.03130812199619916,
+ "top1_changed_fraction": 0.002683637478914277,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.46176108717918396,
+ "maximum_per_target_cross_entropy_increase": 0.4255211353302002
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -6.120696473476546e-05,
+ "mean_prediction_kl_nats": 2.7501241235541264e-05,
+ "prediction_kl_p99_nats": 0.00012287957135383222,
+ "maximum_prediction_kl_nats": 0.030504890121537204,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.4579961895942688,
+ "maximum_per_target_cross_entropy_increase": 0.4258437156677246
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2377769,
+ "unique_rows": 29821,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 73,
+ "compressed_blocks": 176,
+ "zlib_blocks": 176,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29821
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "115f6261b3cbe6655b7e135b2f32fdca5578e89cf4214ced50a0406053548c76",
+ "table_sha256": "58effb10cfc7ec7e38a869e5a2bfb52f9e666911643bd558330de3f31fda30f1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24021192686632276,
+ "representatives_sha256": "dbb73c2f0f3c858d4a43facc7c2385f6331eff0e78a56521e5d8fb196e4fac1f",
+ "structurally_merged_rows": 13331,
+ "backend_artifact_bytes": 2379019,
+ "backend_with_key_registry_bytes": 3115126,
+ "backend_with_datrie_key_index_bytes": 2725179,
+ "full_model_serving_artifact_bytes": 3943124,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9718663746106818,
+ "mean_cosine": 0.9996986927185848,
+ "p05_cosine": 0.9986778996457422,
+ "relative_l2_mean": 0.01593737000666218,
+ "relative_l2_p95": 0.05164042874828437,
+ "relative_l2_max": 0.23754541086938025,
+ "weighted_relative_l2_mean": 0.0038149122698589153,
+ "weighted_relative_frobenius": 0.014632505319454783,
+ "weighted_mean_cosine": 0.9999490403153539,
+ "max_relative_norm_change": 0.01184111110318379
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.971988892125288,
+ "mean_cosine": 0.9997038738568005,
+ "p05_cosine": 0.9986650205358348,
+ "relative_l2_mean": 0.012698022992704637,
+ "relative_l2_p95": 0.051902060540808904,
+ "relative_l2_max": 0.2370694668141154,
+ "weighted_relative_l2_mean": 0.0020718489413359344,
+ "weighted_relative_frobenius": 0.01386121950714379,
+ "weighted_mean_cosine": 0.9999532611190162,
+ "max_relative_norm_change": 0.010095277078878997
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.002899,
+ "warm_p50_ms": 0.1073465,
+ "warm_p95_ms": 0.11318825,
+ "warm_p99_ms": 0.13591303999999998,
+ "samples_ms": [
+ 0.10986,
+ 0.107835,
+ 0.106439,
+ 0.106438,
+ 0.113073,
+ 0.109791,
+ 0.108324,
+ 0.106508,
+ 0.106438,
+ 0.108812,
+ 0.107835,
+ 0.106857,
+ 0.149042,
+ 0.115378,
+ 0.107835,
+ 0.106927,
+ 0.107835,
+ 0.106927,
+ 0.107416,
+ 0.105949,
+ 0.107835,
+ 0.106857,
+ 0.107835,
+ 0.108324,
+ 0.106438,
+ 0.106928,
+ 0.107835,
+ 0.107346,
+ 0.107347,
+ 0.107416,
+ 0.106438,
+ 0.106438,
+ 0.107346,
+ 0.106997,
+ 0.105949,
+ 0.106928,
+ 0.105949,
+ 0.106927,
+ 0.108813,
+ 0.107905
+ ],
+ "warm_rows_per_second": 4769601.244567825
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8771634736025458,
+ "character_perplexity": 6.534942012107043,
+ "next_character_accuracy": 0.49547615396411593,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -8.05595368682254e-05,
+ "mean_prediction_kl_nats": 3.282020287449531e-05,
+ "prediction_kl_p99_nats": 0.00014358032176465654,
+ "maximum_prediction_kl_nats": 0.03130812199619916,
+ "top1_changed_fraction": 0.002683637478914277,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.5839344263076782,
+ "maximum_per_target_cross_entropy_increase": 0.4255211353302002
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -5.597625233961335e-05,
+ "mean_prediction_kl_nats": 2.8922105282300183e-05,
+ "prediction_kl_p99_nats": 0.00012901098501385112,
+ "maximum_prediction_kl_nats": 0.030504890121537204,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.5850692987442017,
+ "maximum_per_target_cross_entropy_increase": 0.4258437156677246
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2376392,
+ "unique_rows": 29802,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 77,
+ "compressed_blocks": 172,
+ "zlib_blocks": 172,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29802
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "5d7ec9791c81359ad1c407a5e62ea64b81ea827f1ba565257771dcefc50ee1eb",
+ "table_sha256": "28680da4d7eec2caecfcc6d9ff90c6c462f723ee96724f0e1a4689c7f173fd25",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.29219344910234213,
+ "representatives_sha256": "4a40c9f9526483f4a76ea25a46a74bbfd0333639fe41ce63a88008fc05d6e3c5",
+ "structurally_merged_rows": 13350,
+ "backend_artifact_bytes": 2377642,
+ "backend_with_key_registry_bytes": 3113749,
+ "backend_with_datrie_key_index_bytes": 2723802,
+ "full_model_serving_artifact_bytes": 3941747,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9508160845312954,
+ "mean_cosine": 0.9996751721259401,
+ "p05_cosine": 0.998671347364355,
+ "relative_l2_mean": 0.016100586464595373,
+ "relative_l2_p95": 0.05178025972616924,
+ "relative_l2_max": 0.31401738190123984,
+ "weighted_relative_l2_mean": 0.003975753280806931,
+ "weighted_relative_frobenius": 0.01963455685845534,
+ "weighted_mean_cosine": 0.9999259144129621,
+ "max_relative_norm_change": 0.01184111110318379
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.950807329596665,
+ "mean_cosine": 0.9996803199105813,
+ "p05_cosine": 0.9986574164925269,
+ "relative_l2_mean": 0.01286334642602158,
+ "relative_l2_p95": 0.052039602686126125,
+ "relative_l2_max": 0.3140450860026669,
+ "weighted_relative_l2_mean": 0.002234223796431955,
+ "weighted_relative_frobenius": 0.019059819155137006,
+ "weighted_mean_cosine": 0.9999301556627495,
+ "max_relative_norm_change": 0.010095277078878997
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.699159,
+ "warm_p50_ms": 0.108813,
+ "warm_p95_ms": 0.11120795,
+ "warm_p99_ms": 0.13388807999999996,
+ "samples_ms": [
+ 0.111606,
+ 0.111187,
+ 0.109372,
+ 0.108813,
+ 0.108813,
+ 0.11021,
+ 0.11028,
+ 0.109231,
+ 0.108813,
+ 0.108813,
+ 0.108813,
+ 0.108813,
+ 0.108883,
+ 0.108813,
+ 0.108324,
+ 0.108813,
+ 0.108813,
+ 0.108324,
+ 0.108812,
+ 0.108883,
+ 0.109301,
+ 0.108883,
+ 0.108324,
+ 0.108883,
+ 0.108813,
+ 0.107905,
+ 0.109651,
+ 0.108813,
+ 0.108394,
+ 0.108324,
+ 0.110699,
+ 0.148134,
+ 0.108813,
+ 0.108393,
+ 0.109302,
+ 0.108323,
+ 0.109302,
+ 0.109302,
+ 0.108393,
+ 0.108324
+ ],
+ "warm_rows_per_second": 4705320.136380764
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8768621728990524,
+ "character_perplexity": 6.532973326079643,
+ "next_character_accuracy": 0.49547615396411593,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0003817905848742269,
+ "mean_prediction_kl_nats": 4.888279063807856e-05,
+ "prediction_kl_p99_nats": 0.0002892620817657262,
+ "maximum_prediction_kl_nats": 0.03130812199619916,
+ "top1_changed_fraction": 0.002683637478914277,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.9851020574569702,
+ "maximum_per_target_cross_entropy_increase": 0.4255211353302002
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.00035721470619078837,
+ "mean_prediction_kl_nats": 4.502350122816192e-05,
+ "prediction_kl_p99_nats": 0.00027518929094581124,
+ "maximum_prediction_kl_nats": 0.030504890121537204,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.00015335071308081316,
+ "maximum_absolute_logit_change": 0.9834340810775757,
+ "maximum_per_target_cross_entropy_increase": 0.4258437156677246
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2374029,
+ "unique_rows": 29766,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 80,
+ "compressed_blocks": 169,
+ "zlib_blocks": 169,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29766
+ },
+ "source_float32_sha256": "530a17bab5f26cb25db40425b1f57616f527687e9c87188ad04b27a87bdf7499",
+ "decoded_float32_sha256": "5fb1b69083eeb75566bd08b09655b73f2bcd4e9326a9a4b2108e81cd81172d94",
+ "table_sha256": "67ef7eb622861792ba7891ca23e20d3617c4bf8423ec6edeca09fe0e6b2b459e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2369166729040444,
+ "representatives_sha256": "27ad7f329d093b6bd2bf2d359586473ab4ceee5466007315ad1c08b3678b61fc",
+ "structurally_merged_rows": 13386,
+ "backend_artifact_bytes": 2375279,
+ "backend_with_key_registry_bytes": 3111386,
+ "backend_with_datrie_key_index_bytes": 2721439,
+ "full_model_serving_artifact_bytes": 3939384,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9028863763223173,
+ "mean_cosine": 0.9996009633335711,
+ "p05_cosine": 0.9986613733646358,
+ "relative_l2_mean": 0.01647700353742353,
+ "relative_l2_p95": 0.05197424947071407,
+ "relative_l2_max": 0.44076765865474365,
+ "weighted_relative_l2_mean": 0.004060801898233982,
+ "weighted_relative_frobenius": 0.023043644896926015,
+ "weighted_mean_cosine": 0.9999097692978524,
+ "max_relative_norm_change": 0.01184111110318379
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9030375219130348,
+ "mean_cosine": 0.999606075750037,
+ "p05_cosine": 0.9986471498976491,
+ "relative_l2_mean": 0.013243753300374974,
+ "relative_l2_p95": 0.052220655036914476,
+ "relative_l2_max": 0.44045396153669425,
+ "weighted_relative_l2_mean": 0.0023202173505376656,
+ "weighted_relative_frobenius": 0.022549180092356604,
+ "weighted_mean_cosine": 0.9999140294942955,
+ "max_relative_norm_change": 0.010095277078878997
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.68114,
+ "warm_p50_ms": 0.109092,
+ "warm_p95_ms": 0.11626874999999998,
+ "warm_p99_ms": 0.12742882,
+ "samples_ms": [
+ 0.111677,
+ 0.110699,
+ 0.111118,
+ 0.130534,
+ 0.109651,
+ 0.109372,
+ 0.113073,
+ 0.111188,
+ 0.108812,
+ 0.11028,
+ 0.109302,
+ 0.108882,
+ 0.109721,
+ 0.108324,
+ 0.108813,
+ 0.109302,
+ 0.110209,
+ 0.108813,
+ 0.109301,
+ 0.108813,
+ 0.108324,
+ 0.11454,
+ 0.108743,
+ 0.109301,
+ 0.109651,
+ 0.108883,
+ 0.108743,
+ 0.115937,
+ 0.108393,
+ 0.108813,
+ 0.113981,
+ 0.108324,
+ 0.108813,
+ 0.107346,
+ 0.108812,
+ 0.108324,
+ 0.108324,
+ 0.108324,
+ 0.122572,
+ 0.108813
+ ],
+ "warm_rows_per_second": 4693286.400469329
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8769360013925565,
+ "character_perplexity": 6.533455663463312,
+ "next_character_accuracy": 0.49539947860757555,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0003079875540441543,
+ "mean_prediction_kl_nats": 8.006796692217607e-05,
+ "prediction_kl_p99_nats": 0.00045658951732107804,
+ "maximum_prediction_kl_nats": 0.10787614291341004,
+ "top1_changed_fraction": 0.002913663548535501,
+ "accuracy_change": 0.0002300260696212475,
+ "maximum_absolute_logit_change": 1.4234192371368408,
+ "maximum_per_target_cross_entropy_increase": 0.9190587997436523
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.0002834098608936371,
+ "mean_prediction_kl_nats": 7.585275770922256e-05,
+ "prediction_kl_p99_nats": 0.00045597743366669087,
+ "maximum_prediction_kl_nats": 0.10639203174675102,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.4199752807617188,
+ "maximum_per_target_cross_entropy_increase": 0.9124231338500977
+ }
+ }
+ ],
+ "baseline_quality": {
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8772440363845078,
+ "character_perplexity": 6.535468506423138,
+ "next_character_accuracy": 0.4951694525379543,
+ "unknown_targets": 0
+ },
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7466553011190267,
+ "character_perplexity": 5.73538741541858,
+ "next_character_accuracy": 0.5228091696695546,
+ "unknown_targets": 0
+ }
+ },
+ "selection": {
+ "fp32": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 8191912,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467407476364325,
+ "character_perplexity": 5.735877505237137,
+ "next_character_accuracy": 0.522579161235912,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 8.545016806342631e-05,
+ "mean_prediction_kl_nats": 7.555171199507978e-05,
+ "prediction_kl_p99_nats": 0.0002332713868702684,
+ "maximum_prediction_kl_nats": 0.07069631943714265,
+ "top1_changed_fraction": 0.0017633979912596795,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 1.421870231628418,
+ "maximum_per_target_cross_entropy_increase": 0.5875310897827148
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00011373525728950033,
+ 0.0002831974613994154
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "fp16": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 4609985,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467389892458764,
+ "character_perplexity": 5.735867419333168,
+ "next_character_accuracy": 0.522579161235912,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 8.369162879872298e-05,
+ "mean_prediction_kl_nats": 7.55530360250437e-05,
+ "prediction_kl_p99_nats": 0.00023711060763044312,
+ "maximum_prediction_kl_nats": 0.07071264041728009,
+ "top1_changed_fraction": 0.0017633979912596795,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 1.4215927124023438,
+ "maximum_per_target_cross_entropy_increase": 0.5875155925750732
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00011714467232062939,
+ 0.0002800202357481195
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "int8": {
+ "feasible": true,
+ "name": "int8-joint-0.9",
+ "backend_with_key_registry_bytes": 3111386,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467207411986971,
+ "character_perplexity": 5.735762751908877,
+ "next_character_accuracy": 0.5227325001916737,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 6.544222091906036e-05,
+ "mean_prediction_kl_nats": 7.611414126180771e-05,
+ "prediction_kl_p99_nats": 0.00023912776528949614,
+ "maximum_prediction_kl_nats": 0.07220868673265157,
+ "top1_changed_fraction": 0.002300084336425669,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 1.4234991073608398,
+ "maximum_per_target_cross_entropy_increase": 0.597667932510376
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00014768831565003417,
+ 0.00027000310437506474
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ }
+ },
+ "fixed_quantization_test_controls": {
+ "fp16": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7466544953558614,
+ "character_perplexity": 5.735382794056524,
+ "next_character_accuracy": 0.5227325001916737,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -8.057628406771341e-07,
+ "mean_prediction_kl_nats": 3.845799707402909e-09,
+ "prediction_kl_p99_nats": 2.6922498327991973e-08,
+ "maximum_prediction_kl_nats": 8.843846398377895e-08,
+ "top1_changed_fraction": 7.666947788085564e-05,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.0010104477405548096,
+ "maximum_per_target_cross_entropy_increase": 0.0008120536804199219
+ }
+ },
+ "int8": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7466401825079891,
+ "character_perplexity": 5.735300704982569,
+ "next_character_accuracy": 0.5226558307137928,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -1.5118496752060422e-05,
+ "mean_prediction_kl_nats": 3.526689483178103e-06,
+ "prediction_kl_p99_nats": 2.802159804706195e-05,
+ "maximum_prediction_kl_nats": 0.00010887467438350836,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": -0.00015333895576175038,
+ "maximum_absolute_logit_change": 0.04378390312194824,
+ "maximum_per_target_cross_entropy_increase": 0.02187180519104004
+ }
+ }
+ },
+ "checkpoint_sha256": "b1dc4a4b7a8ae2ed1ef9688aa0cccf047c766581b08fe1608fc5625cfc93d0d6",
+ "registry_sha256": "abab407b9c927abc758d3d85a11cec2c2535254fe5af55b6cce10b650a992d4d"
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 7,
+ "rows": 43152,
+ "dim": 64,
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "source_float32_bytes": 11046912,
+ "key_registry_bytes": 736107,
+ "unchanged_backbone_artifact_bytes": 827998,
+ "alternative_datrie_key_index_bytes": 346160,
+ "query_rows": 512,
+ "query_ids_sha256": "43ca78408803e060726877734dd5761b4e85ad9be7de5a6b6d8601f32db85068",
+ "frequency_trace_sha256": "7ba41132ac5296b1ee890e8c97c4f892398df8fb97e4bbb475dab44b5c5af0cc",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 11046912,
+ "unique_rows": 43152,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "table_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.05895057111047208,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 11047717,
+ "backend_with_key_registry_bytes": 11783824,
+ "backend_with_datrie_key_index_bytes": 11393877,
+ "full_model_serving_artifact_bytes": 12611822,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.240883,
+ "warm_p50_ms": 0.0468635,
+ "warm_p95_ms": 0.0497059,
+ "warm_p99_ms": 0.07372593999999998,
+ "samples_ms": [
+ 0.049238,
+ 0.046305,
+ 0.088489,
+ 0.049657,
+ 0.046793,
+ 0.046374,
+ 0.046375,
+ 0.04875,
+ 0.046794,
+ 0.046235,
+ 0.047283,
+ 0.046794,
+ 0.048191,
+ 0.046794,
+ 0.047283,
+ 0.047282,
+ 0.046794,
+ 0.046724,
+ 0.046305,
+ 0.046793,
+ 0.047283,
+ 0.046375,
+ 0.046375,
+ 0.047282,
+ 0.047282,
+ 0.050635,
+ 0.049238,
+ 0.046444,
+ 0.047841,
+ 0.047352,
+ 0.04826,
+ 0.046794,
+ 0.046864,
+ 0.047772,
+ 0.046374,
+ 0.046863,
+ 0.046864,
+ 0.046375,
+ 0.046374,
+ 0.047352
+ ],
+ "warm_rows_per_second": 10925347.01846853
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.873996702168374,
+ "character_perplexity": 6.514280077498528,
+ "next_character_accuracy": 0.5035270664008588,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 10651825,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1349,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1349,
+ "zlib_blocks": 1349,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "table_sha256": "135de811416f938de0eb92c8b6a88aa9a431a4329bf7d25c64da79dc7640765d",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.912386961048469,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 10653037,
+ "backend_with_key_registry_bytes": 11389144,
+ "backend_with_datrie_key_index_bytes": 10999197,
+ "full_model_serving_artifact_bytes": 12217142,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.196884,
+ "warm_p50_ms": 0.0549305,
+ "warm_p95_ms": 0.05731885,
+ "warm_p99_ms": 0.06519124,
+ "samples_ms": [
+ 0.05727,
+ 0.055873,
+ 0.069631,
+ 0.056781,
+ 0.058247,
+ 0.056292,
+ 0.055803,
+ 0.054895,
+ 0.055803,
+ 0.056223,
+ 0.056781,
+ 0.056362,
+ 0.054896,
+ 0.053988,
+ 0.054965,
+ 0.053918,
+ 0.054407,
+ 0.055873,
+ 0.056711,
+ 0.055384,
+ 0.055873,
+ 0.055315,
+ 0.053987,
+ 0.056362,
+ 0.055803,
+ 0.055384,
+ 0.053498,
+ 0.053987,
+ 0.05294,
+ 0.054336,
+ 0.054895,
+ 0.05308,
+ 0.054476,
+ 0.053498,
+ 0.054476,
+ 0.053498,
+ 0.053429,
+ 0.05294,
+ 0.05294,
+ 0.053499
+ ],
+ "warm_rows_per_second": 9320869.09822412
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.873996702168374,
+ "character_perplexity": 6.514280077498528,
+ "next_character_accuracy": 0.5035270664008588,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 10651825,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1349,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1349,
+ "zlib_blocks": 1349,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "table_sha256": "135de811416f938de0eb92c8b6a88aa9a431a4329bf7d25c64da79dc7640765d",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 1.0875776370521635,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 10653036,
+ "backend_with_key_registry_bytes": 11389143,
+ "backend_with_datrie_key_index_bytes": 10999196,
+ "full_model_serving_artifact_bytes": 12217141,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.188851,
+ "warm_p50_ms": 0.055384,
+ "warm_p95_ms": 0.0569173,
+ "warm_p99_ms": 0.06571843,
+ "samples_ms": [
+ 0.058177,
+ 0.055384,
+ 0.056292,
+ 0.054895,
+ 0.055804,
+ 0.055873,
+ 0.053987,
+ 0.053499,
+ 0.054896,
+ 0.053499,
+ 0.053918,
+ 0.053428,
+ 0.07054,
+ 0.055384,
+ 0.055384,
+ 0.056851,
+ 0.055385,
+ 0.055734,
+ 0.055733,
+ 0.055803,
+ 0.055314,
+ 0.054826,
+ 0.054895,
+ 0.055804,
+ 0.055734,
+ 0.055803,
+ 0.056292,
+ 0.054825,
+ 0.055803,
+ 0.055803,
+ 0.054826,
+ 0.054896,
+ 0.056851,
+ 0.055384,
+ 0.054406,
+ 0.056292,
+ 0.053847,
+ 0.052939,
+ 0.053498,
+ 0.055384
+ ],
+ "warm_rows_per_second": 9244547.161635129
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.873996702168374,
+ "character_perplexity": 6.514280077498528,
+ "next_character_accuracy": 0.5035270664008588,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 10651825,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1349,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1349,
+ "zlib_blocks": 1349,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "table_sha256": "135de811416f938de0eb92c8b6a88aa9a431a4329bf7d25c64da79dc7640765d",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.9435423500835896,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 10653036,
+ "backend_with_key_registry_bytes": 11389143,
+ "backend_with_datrie_key_index_bytes": 10999196,
+ "full_model_serving_artifact_bytes": 12217141,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.176001,
+ "warm_p50_ms": 0.055873,
+ "warm_p95_ms": 0.05729095,
+ "warm_p99_ms": 0.05858326,
+ "samples_ms": [
+ 0.059155,
+ 0.056781,
+ 0.054825,
+ 0.054895,
+ 0.055803,
+ 0.056223,
+ 0.055314,
+ 0.055314,
+ 0.056292,
+ 0.056292,
+ 0.055385,
+ 0.056292,
+ 0.056292,
+ 0.055873,
+ 0.056292,
+ 0.054407,
+ 0.055873,
+ 0.056362,
+ 0.056362,
+ 0.056362,
+ 0.056223,
+ 0.056222,
+ 0.055384,
+ 0.055314,
+ 0.056711,
+ 0.055385,
+ 0.057689,
+ 0.056711,
+ 0.05727,
+ 0.056711,
+ 0.055733,
+ 0.055384,
+ 0.055315,
+ 0.055734,
+ 0.056362,
+ 0.055384,
+ 0.055384,
+ 0.055873,
+ 0.055385,
+ 0.054407
+ ],
+ "warm_rows_per_second": 9163638.966942888
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.873996702168374,
+ "character_perplexity": 6.514280077498528,
+ "next_character_accuracy": 0.5035270664008588,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7466080,
+ "unique_rows": 29797,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 932,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 932,
+ "zlib_blocks": 932,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29797
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "e6f1a59b18732b2ccca059d7eb3e4d64ff74d96378a19c7ee039c9349d3c48a6",
+ "table_sha256": "bc350fda73574e114a7cadc5cf7f0dff8aa38624fe0814b623f43ec25628ed63",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6810067519545555,
+ "representatives_sha256": "3652f5e1dbffd87708445a47cdbb024b4e5d7094650aad799d8e62ad245812dd",
+ "structurally_merged_rows": 13355,
+ "backend_artifact_bytes": 7467287,
+ "backend_with_key_registry_bytes": 8203394,
+ "backend_with_datrie_key_index_bytes": 7813447,
+ "full_model_serving_artifact_bytes": 9031392,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9900334924016594,
+ "mean_cosine": 0.9997199486281269,
+ "p05_cosine": 0.9986676323241817,
+ "relative_l2_mean": 0.012526903122645183,
+ "relative_l2_p95": 0.05181032951262053,
+ "relative_l2_max": 0.14116478475441507,
+ "weighted_relative_l2_mean": 0.002007054021764998,
+ "weighted_relative_frobenius": 0.012525398130393038,
+ "weighted_mean_cosine": 0.9999602049532175,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9900334924016594,
+ "mean_cosine": 0.9997199486281269,
+ "p05_cosine": 0.9986676323241817,
+ "relative_l2_mean": 0.012526903122645183,
+ "relative_l2_p95": 0.05181032951262053,
+ "relative_l2_max": 0.14116478475441507,
+ "weighted_relative_l2_mean": 0.002007054021764998,
+ "weighted_relative_frobenius": 0.012525398130393038,
+ "weighted_mean_cosine": 0.9999602049532175,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.74934,
+ "warm_p50_ms": 0.0558385,
+ "warm_p95_ms": 0.058321149999999995,
+ "warm_p99_ms": 0.06943399,
+ "samples_ms": [
+ 0.058178,
+ 0.05727,
+ 0.056781,
+ 0.057759,
+ 0.056293,
+ 0.055803,
+ 0.055314,
+ 0.056292,
+ 0.054825,
+ 0.056292,
+ 0.055315,
+ 0.055803,
+ 0.056711,
+ 0.054336,
+ 0.055803,
+ 0.054756,
+ 0.0748,
+ 0.061041,
+ 0.055873,
+ 0.055873,
+ 0.054896,
+ 0.055874,
+ 0.056292,
+ 0.055873,
+ 0.055873,
+ 0.055803,
+ 0.056292,
+ 0.056362,
+ 0.056781,
+ 0.055803,
+ 0.055804,
+ 0.054826,
+ 0.054895,
+ 0.055873,
+ 0.055733,
+ 0.055733,
+ 0.055315,
+ 0.055804,
+ 0.054826,
+ 0.055803
+ ],
+ "warm_rows_per_second": 9169300.751273762
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874021198972967,
+ "character_perplexity": 6.514439658499254,
+ "next_character_accuracy": 0.5036037417573992,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.4497104641860765e-05,
+ "mean_prediction_kl_nats": 7.042508406295861e-06,
+ "prediction_kl_p99_nats": 0.00011149122424128026,
+ "maximum_prediction_kl_nats": 0.006084268086667243,
+ "top1_changed_fraction": 0.0006134028523232633,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.360864520072937,
+ "maximum_per_target_cross_entropy_increase": 0.08151006698608398
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 2.4497104641860765e-05,
+ "mean_prediction_kl_nats": 7.042508406295861e-06,
+ "prediction_kl_p99_nats": 0.00011149122424128026,
+ "maximum_prediction_kl_nats": 0.006084268086667243,
+ "top1_changed_fraction": 0.0006134028523232633,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.360864520072937,
+ "maximum_per_target_cross_entropy_increase": 0.08151006698608398
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7461116,
+ "unique_rows": 29777,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 931,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 931,
+ "zlib_blocks": 931,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29777
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "9f3eed8a373ac6e2b6ca62816734806b6987c4b9bd3514d1bb7aa8f33095cb6e",
+ "table_sha256": "1113a1cbe20f280a6e6c0b9db01883080e7aa1c4b5b231399c21c5e1f5d95fe4",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6789274350740016,
+ "representatives_sha256": "f1d3dab05a26df4964bf133a26202cb3e62b81bf9d91e50daad9713a0a5a6bc6",
+ "structurally_merged_rows": 13375,
+ "backend_artifact_bytes": 7462323,
+ "backend_with_key_registry_bytes": 8198430,
+ "backend_with_datrie_key_index_bytes": 7808483,
+ "full_model_serving_artifact_bytes": 9026428,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9801761387231435,
+ "mean_cosine": 0.999709929561454,
+ "p05_cosine": 0.9986587196347765,
+ "relative_l2_mean": 0.01263662077792843,
+ "relative_l2_p95": 0.05195302074014606,
+ "relative_l2_max": 0.1987708157070306,
+ "weighted_relative_l2_mean": 0.0020355184953020798,
+ "weighted_relative_frobenius": 0.013253868377612358,
+ "weighted_mean_cosine": 0.9999578304533462,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9801761387231435,
+ "mean_cosine": 0.999709929561454,
+ "p05_cosine": 0.9986587196347765,
+ "relative_l2_mean": 0.01263662077792843,
+ "relative_l2_p95": 0.05195302074014606,
+ "relative_l2_max": 0.1987708157070306,
+ "weighted_relative_l2_mean": 0.0020355184953020798,
+ "weighted_relative_frobenius": 0.013253868377612358,
+ "weighted_mean_cosine": 0.9999578304533462,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.488692,
+ "warm_p50_ms": 0.0544415,
+ "warm_p95_ms": 0.055803,
+ "warm_p99_ms": 0.05665517,
+ "samples_ms": [
+ 0.0572,
+ 0.055734,
+ 0.054826,
+ 0.055803,
+ 0.054825,
+ 0.055803,
+ 0.055384,
+ 0.054407,
+ 0.053848,
+ 0.055803,
+ 0.053918,
+ 0.054895,
+ 0.054895,
+ 0.053918,
+ 0.054476,
+ 0.05252,
+ 0.054825,
+ 0.053847,
+ 0.05301,
+ 0.054407,
+ 0.054966,
+ 0.054896,
+ 0.05287,
+ 0.054337,
+ 0.053917,
+ 0.054406,
+ 0.054406,
+ 0.054965,
+ 0.053428,
+ 0.053359,
+ 0.053499,
+ 0.054476,
+ 0.054896,
+ 0.054896,
+ 0.053988,
+ 0.054407,
+ 0.054337,
+ 0.055315,
+ 0.054826,
+ 0.053429
+ ],
+ "warm_rows_per_second": 9404590.24824812
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874016033993057,
+ "character_perplexity": 6.5144060116361855,
+ "next_character_accuracy": 0.5036804171139396,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.9332011543706106e-05,
+ "mean_prediction_kl_nats": 1.4418529893776803e-05,
+ "prediction_kl_p99_nats": 0.00012110718802732223,
+ "maximum_prediction_kl_nats": 0.022813439426192204,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.360864520072937,
+ "maximum_per_target_cross_entropy_increase": 0.353954553604126
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 1.9332011543706106e-05,
+ "mean_prediction_kl_nats": 1.4418529893776803e-05,
+ "prediction_kl_p99_nats": 0.00012110718802732223,
+ "maximum_prediction_kl_nats": 0.022813439426192204,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.360864520072937,
+ "maximum_per_target_cross_entropy_increase": 0.353954553604126
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7458808,
+ "unique_rows": 29767,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 931,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 931,
+ "zlib_blocks": 931,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29767
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "040815345c80f0bb2398856e2464c38ef258c14a58bc8cd72ad18a329d881b2b",
+ "table_sha256": "2acf8ed5238bb7eab2d52639b9d31c3aec424c86375473c0ac44b15c568f4f9f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6844950481317937,
+ "representatives_sha256": "dbdaaf7bd53b24b9909da72b1886fb7acf31042b9137425933ea1675db39a68b",
+ "structurally_merged_rows": 13385,
+ "backend_artifact_bytes": 7460015,
+ "backend_with_key_registry_bytes": 8196122,
+ "backend_with_datrie_key_index_bytes": 7806175,
+ "full_model_serving_artifact_bytes": 9024120,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9700479192000054,
+ "mean_cosine": 0.9997034923496534,
+ "p05_cosine": 0.9986566517495193,
+ "relative_l2_mean": 0.012692843287476942,
+ "relative_l2_p95": 0.05202228637995757,
+ "relative_l2_max": 0.24495158161554725,
+ "weighted_relative_l2_mean": 0.002078217974107309,
+ "weighted_relative_frobenius": 0.01413688195204179,
+ "weighted_mean_cosine": 0.9999530586187222,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9700479192000054,
+ "mean_cosine": 0.9997034923496534,
+ "p05_cosine": 0.9986566517495193,
+ "relative_l2_mean": 0.012692843287476942,
+ "relative_l2_p95": 0.05202228637995757,
+ "relative_l2_max": 0.24495158161554725,
+ "weighted_relative_l2_mean": 0.002078217974107309,
+ "weighted_relative_frobenius": 0.01413688195204179,
+ "weighted_mean_cosine": 0.9999530586187222,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.694933,
+ "warm_p50_ms": 0.0544415,
+ "warm_p95_ms": 0.055873,
+ "warm_p99_ms": 0.055873,
+ "samples_ms": [
+ 0.055873,
+ 0.055803,
+ 0.055384,
+ 0.054407,
+ 0.055315,
+ 0.055385,
+ 0.054965,
+ 0.054476,
+ 0.054407,
+ 0.055873,
+ 0.053987,
+ 0.054476,
+ 0.054825,
+ 0.054336,
+ 0.053429,
+ 0.053499,
+ 0.054477,
+ 0.054966,
+ 0.053499,
+ 0.053429,
+ 0.053498,
+ 0.055873,
+ 0.055873,
+ 0.055384,
+ 0.055734,
+ 0.055314,
+ 0.054336,
+ 0.055314,
+ 0.05301,
+ 0.054895,
+ 0.053569,
+ 0.053429,
+ 0.053848,
+ 0.053498,
+ 0.053568,
+ 0.053917,
+ 0.053499,
+ 0.053499,
+ 0.055315,
+ 0.053499
+ ],
+ "warm_rows_per_second": 9404590.24824812
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8738604000223773,
+ "character_perplexity": 6.513392227653635,
+ "next_character_accuracy": 0.5038337678270204,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.00013629285728137262,
+ "mean_prediction_kl_nats": 3.678425188654329e-05,
+ "prediction_kl_p99_nats": 0.0001588043345496865,
+ "maximum_prediction_kl_nats": 0.03321836029867732,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.40522003173828125,
+ "maximum_per_target_cross_entropy_increase": 0.373732328414917
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.00013629285728137262,
+ "mean_prediction_kl_nats": 3.678425188654329e-05,
+ "prediction_kl_p99_nats": 0.0001588043345496865,
+ "maximum_prediction_kl_nats": 0.03321836029867732,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.40522003173828125,
+ "maximum_per_target_cross_entropy_increase": 0.373732328414917
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7455090,
+ "unique_rows": 29751,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 930,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 930,
+ "zlib_blocks": 930,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29751
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "21aeab503f5167197e19e2bacc9b24daf83bde215cdf92ecf53cd8c7869438c7",
+ "table_sha256": "471e28d4ce98ba96d6e521559d47a96b0f95383516d8c2c9a2c8e6300b3f2181",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6882852690760046,
+ "representatives_sha256": "9472bfcc8138076a97955f7e03d674e66f41b00a37eec5d1b506a610ab62477e",
+ "structurally_merged_rows": 13401,
+ "backend_artifact_bytes": 7456297,
+ "backend_with_key_registry_bytes": 8192404,
+ "backend_with_datrie_key_index_bytes": 7802457,
+ "full_model_serving_artifact_bytes": 9020402,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9511890967280789,
+ "mean_cosine": 0.9996853054841821,
+ "p05_cosine": 0.998652364283855,
+ "relative_l2_mean": 0.012823303720665598,
+ "relative_l2_p95": 0.05210850837998779,
+ "relative_l2_max": 0.31190380849243704,
+ "weighted_relative_l2_mean": 0.0021191932688392873,
+ "weighted_relative_frobenius": 0.01549617638905431,
+ "weighted_mean_cosine": 0.9999475656981209,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9511890967280789,
+ "mean_cosine": 0.9996853054841821,
+ "p05_cosine": 0.998652364283855,
+ "relative_l2_mean": 0.012823303720665598,
+ "relative_l2_p95": 0.05210850837998779,
+ "relative_l2_max": 0.31190380849243704,
+ "weighted_relative_l2_mean": 0.0021191932688392873,
+ "weighted_relative_frobenius": 0.01549617638905431,
+ "weighted_mean_cosine": 0.9999475656981209,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.810381,
+ "warm_p50_ms": 0.0553845,
+ "warm_p95_ms": 0.058111499999999996,
+ "warm_p99_ms": 0.07896801999999997,
+ "samples_ms": [
+ 0.058178,
+ 0.057759,
+ 0.058108,
+ 0.055873,
+ 0.055803,
+ 0.055873,
+ 0.053987,
+ 0.055314,
+ 0.053987,
+ 0.055384,
+ 0.055384,
+ 0.057689,
+ 0.055384,
+ 0.055873,
+ 0.054895,
+ 0.055803,
+ 0.055804,
+ 0.055804,
+ 0.053848,
+ 0.055384,
+ 0.053918,
+ 0.055803,
+ 0.055803,
+ 0.09226,
+ 0.056292,
+ 0.055315,
+ 0.055385,
+ 0.054965,
+ 0.054895,
+ 0.056851,
+ 0.055384,
+ 0.055454,
+ 0.055873,
+ 0.054406,
+ 0.056781,
+ 0.054406,
+ 0.053498,
+ 0.054826,
+ 0.054896,
+ 0.053918
+ ],
+ "warm_rows_per_second": 9244463.703743827
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8737053970996589,
+ "character_perplexity": 6.512382711062569,
+ "next_character_accuracy": 0.5038337678270204,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.00029126264351335006,
+ "mean_prediction_kl_nats": 7.012482966510538e-05,
+ "prediction_kl_p99_nats": 0.00021856286480779978,
+ "maximum_prediction_kl_nats": 0.16057564868335505,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.6747455596923828,
+ "maximum_per_target_cross_entropy_increase": 0.704582154750824
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.00029126264351335006,
+ "mean_prediction_kl_nats": 7.012482966510538e-05,
+ "prediction_kl_p99_nats": 0.00021856286480779978,
+ "maximum_prediction_kl_nats": 0.16057564868335505,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.6747455596923828,
+ "maximum_per_target_cross_entropy_increase": 0.704582154750824
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7447623,
+ "unique_rows": 29720,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 929,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 929,
+ "zlib_blocks": 929,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29720
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "fa04c36d71797751dc8dff98932c0d08480d29168cd1219aa9bbd45fa892e823",
+ "table_sha256": "8554325ce8d9cc6b9aaae4e9bd8bf607a17749be533f5e93dfa09fe2d5c5ec3a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6912711260374635,
+ "representatives_sha256": "abcf9fbaa2f2ed444d9b68b184addbd08254ddd1d9be6eef0ebcdf187d0aab12",
+ "structurally_merged_rows": 13432,
+ "backend_artifact_bytes": 7448830,
+ "backend_with_key_registry_bytes": 8184937,
+ "backend_with_datrie_key_index_bytes": 7794990,
+ "full_model_serving_artifact_bytes": 9012935,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9025613601953411,
+ "mean_cosine": 0.9996067258313748,
+ "p05_cosine": 0.9986418040995083,
+ "relative_l2_mean": 0.013220066766763107,
+ "relative_l2_p95": 0.052349069892072136,
+ "relative_l2_max": 0.4405381737648303,
+ "weighted_relative_l2_mean": 0.0023323067816422806,
+ "weighted_relative_frobenius": 0.02297713792438167,
+ "weighted_mean_cosine": 0.9999051293131374,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9025613601953411,
+ "mean_cosine": 0.9996067258313748,
+ "p05_cosine": 0.9986418040995083,
+ "relative_l2_mean": 0.013220066766763107,
+ "relative_l2_p95": 0.052349069892072136,
+ "relative_l2_max": 0.4405381737648303,
+ "weighted_relative_l2_mean": 0.0023323067816422806,
+ "weighted_relative_frobenius": 0.02297713792438167,
+ "weighted_mean_cosine": 0.9999051293131374,
+ "max_relative_norm_change": 0.010097123753310348
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.38882,
+ "warm_p50_ms": 0.054372000000000004,
+ "warm_p95_ms": 0.06370569999999995,
+ "warm_p99_ms": 0.07939648,
+ "samples_ms": [
+ 0.056362,
+ 0.056781,
+ 0.055454,
+ 0.055873,
+ 0.055384,
+ 0.054476,
+ 0.055873,
+ 0.055804,
+ 0.055384,
+ 0.053987,
+ 0.053918,
+ 0.053918,
+ 0.054337,
+ 0.053429,
+ 0.056781,
+ 0.056362,
+ 0.053917,
+ 0.053918,
+ 0.053988,
+ 0.053988,
+ 0.054826,
+ 0.053848,
+ 0.054337,
+ 0.054337,
+ 0.054337,
+ 0.079969,
+ 0.078501,
+ 0.062927,
+ 0.054337,
+ 0.053848,
+ 0.053917,
+ 0.054895,
+ 0.055314,
+ 0.054407,
+ 0.054407,
+ 0.054058,
+ 0.055804,
+ 0.053918,
+ 0.053988,
+ 0.054337
+ ],
+ "warm_rows_per_second": 9416611.49120871
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8741347379235438,
+ "character_perplexity": 6.51517934313239,
+ "next_character_accuracy": 0.504140469253182,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.00013804528254298488,
+ "mean_prediction_kl_nats": 0.00026054510540956394,
+ "prediction_kl_p99_nats": 0.0009175388561113203,
+ "maximum_prediction_kl_nats": 0.4085219028683875,
+ "top1_changed_fraction": 0.0015335071308081583,
+ "accuracy_change": 0.0006134028523232526,
+ "maximum_absolute_logit_change": 1.6836999654769897,
+ "maximum_per_target_cross_entropy_increase": 1.219319224357605
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00013804528254298488,
+ "mean_prediction_kl_nats": 0.00026054510540956394,
+ "prediction_kl_p99_nats": 0.0009175388561113203,
+ "maximum_prediction_kl_nats": 0.4085219028683875,
+ "top1_changed_fraction": 0.0015335071308081583,
+ "accuracy_change": 0.0006134028523232526,
+ "maximum_absolute_logit_change": 1.6836999654769897,
+ "maximum_per_target_cross_entropy_increase": 1.219319224357605
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 128,
+ "storage": {
+ "artifact_bytes": 5523456,
+ "unique_rows": 43152,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "b9e46c364634283ab41e551138f1ada23921e31ee48a04dd3989a3da162f7b1f",
+ "table_sha256": "be124675b47f909e61b0388745ad00b3ed3a74bff02b3b5a333d38787d391165",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.07052363315597177,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5524261,
+ "backend_with_key_registry_bytes": 6260368,
+ "backend_with_datrie_key_index_bytes": 5870421,
+ "full_model_serving_artifact_bytes": 7088366,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999539845378,
+ "mean_cosine": 0.9999999792071412,
+ "p05_cosine": 0.9999999715996549,
+ "relative_l2_mean": 0.00020637351738737785,
+ "relative_l2_p95": 0.00024209629415095646,
+ "relative_l2_max": 0.0003075455058483477,
+ "weighted_relative_l2_mean": 7.822433519936367e-05,
+ "weighted_relative_frobenius": 0.00020811630848376357,
+ "weighted_mean_cosine": 0.9999999921676113,
+ "max_relative_norm_change": 0.00018268239076675474
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.313238,
+ "warm_p50_ms": 0.139613,
+ "warm_p95_ms": 0.14054195,
+ "warm_p99_ms": 0.15278376,
+ "samples_ms": [
+ 0.140102,
+ 0.140521,
+ 0.140032,
+ 0.139613,
+ 0.139543,
+ 0.140521,
+ 0.139543,
+ 0.139124,
+ 0.138704,
+ 0.140032,
+ 0.140032,
+ 0.139962,
+ 0.140032,
+ 0.138635,
+ 0.140032,
+ 0.160356,
+ 0.14094,
+ 0.140032,
+ 0.139124,
+ 0.139124,
+ 0.138216,
+ 0.139613,
+ 0.140032,
+ 0.140032,
+ 0.140032,
+ 0.140102,
+ 0.139613,
+ 0.139543,
+ 0.139473,
+ 0.138635,
+ 0.139543,
+ 0.138146,
+ 0.139613,
+ 0.139543,
+ 0.139613,
+ 0.138705,
+ 0.138705,
+ 0.139124,
+ 0.139612,
+ 0.138635
+ ],
+ "warm_rows_per_second": 3667280.267596857
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8739957233721811,
+ "character_perplexity": 6.514273701349109,
+ "next_character_accuracy": 0.5036037417573992,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -9.787957139825886e-07,
+ "mean_prediction_kl_nats": 4.398342473752485e-09,
+ "prediction_kl_p99_nats": 3.380791472690338e-08,
+ "maximum_prediction_kl_nats": 9.409525687431273e-08,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.001085519790649414,
+ "maximum_per_target_cross_entropy_increase": 0.0008978843688964844
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 5452876,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 675,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 675,
+ "zlib_blocks": 675,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "b9e46c364634283ab41e551138f1ada23921e31ee48a04dd3989a3da162f7b1f",
+ "table_sha256": "47d81729a61e6f79f28618ae7a0f10d49efd41fdc2fae99dbdf19e801e0afc05",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5023319700267166,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5454084,
+ "backend_with_key_registry_bytes": 6190191,
+ "backend_with_datrie_key_index_bytes": 5800244,
+ "full_model_serving_artifact_bytes": 7018189,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999539845378,
+ "mean_cosine": 0.9999999792071412,
+ "p05_cosine": 0.9999999715996549,
+ "relative_l2_mean": 0.00020637351738737785,
+ "relative_l2_p95": 0.00024209629415095646,
+ "relative_l2_max": 0.0003075455058483477,
+ "weighted_relative_l2_mean": 7.822433519936367e-05,
+ "weighted_relative_frobenius": 0.00020811630848376357,
+ "weighted_mean_cosine": 0.9999999921676113,
+ "max_relative_norm_change": 0.00018268239076675474
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.930509,
+ "warm_p50_ms": 0.147156,
+ "warm_p95_ms": 0.15575609999999998,
+ "warm_p99_ms": 0.17258563,
+ "samples_ms": [
+ 0.153232,
+ 0.155117,
+ 0.149111,
+ 0.147156,
+ 0.147645,
+ 0.147085,
+ 0.147156,
+ 0.147156,
+ 0.147155,
+ 0.146248,
+ 0.146666,
+ 0.147156,
+ 0.146666,
+ 0.148064,
+ 0.147575,
+ 0.151346,
+ 0.146667,
+ 0.147156,
+ 0.147156,
+ 0.146666,
+ 0.146597,
+ 0.152813,
+ 0.146178,
+ 0.146667,
+ 0.147156,
+ 0.167899,
+ 0.175582,
+ 0.14995,
+ 0.152394,
+ 0.148064,
+ 0.146178,
+ 0.147644,
+ 0.147086,
+ 0.147645,
+ 0.147644,
+ 0.147156,
+ 0.145759,
+ 0.147645,
+ 0.147156,
+ 0.147644
+ ],
+ "warm_rows_per_second": 3479300.877979831
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8739957233721811,
+ "character_perplexity": 6.514273701349109,
+ "next_character_accuracy": 0.5036037417573992,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -9.787957139825886e-07,
+ "mean_prediction_kl_nats": 4.398342473752485e-09,
+ "prediction_kl_p99_nats": 3.380791472690338e-08,
+ "maximum_prediction_kl_nats": 9.409525687431273e-08,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.001085519790649414,
+ "maximum_per_target_cross_entropy_increase": 0.0008978843688964844
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 5452876,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 675,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 675,
+ "zlib_blocks": 675,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "b9e46c364634283ab41e551138f1ada23921e31ee48a04dd3989a3da162f7b1f",
+ "table_sha256": "47d81729a61e6f79f28618ae7a0f10d49efd41fdc2fae99dbdf19e801e0afc05",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6999258659780025,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5454083,
+ "backend_with_key_registry_bytes": 6190190,
+ "backend_with_datrie_key_index_bytes": 5800243,
+ "full_model_serving_artifact_bytes": 7018188,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999539845378,
+ "mean_cosine": 0.9999999792071412,
+ "p05_cosine": 0.9999999715996549,
+ "relative_l2_mean": 0.00020637351738737785,
+ "relative_l2_p95": 0.00024209629415095646,
+ "relative_l2_max": 0.0003075455058483477,
+ "weighted_relative_l2_mean": 7.822433519936367e-05,
+ "weighted_relative_frobenius": 0.00020811630848376357,
+ "weighted_mean_cosine": 0.9999999921676113,
+ "max_relative_norm_change": 0.00018268239076675474
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.930508,
+ "warm_p50_ms": 0.147574,
+ "warm_p95_ms": 0.1495578,
+ "warm_p99_ms": 0.16496409999999997,
+ "samples_ms": [
+ 0.14946,
+ 0.148063,
+ 0.149041,
+ 0.147645,
+ 0.146667,
+ 0.147575,
+ 0.148134,
+ 0.147156,
+ 0.147574,
+ 0.147644,
+ 0.147575,
+ 0.146178,
+ 0.147226,
+ 0.147645,
+ 0.147574,
+ 0.147645,
+ 0.173626,
+ 0.151416,
+ 0.148063,
+ 0.146178,
+ 0.147225,
+ 0.146667,
+ 0.147155,
+ 0.147645,
+ 0.146667,
+ 0.147644,
+ 0.147086,
+ 0.147156,
+ 0.148064,
+ 0.145759,
+ 0.147225,
+ 0.146178,
+ 0.147644,
+ 0.146178,
+ 0.146178,
+ 0.147086,
+ 0.147575,
+ 0.147644,
+ 0.145829,
+ 0.145759
+ ],
+ "warm_rows_per_second": 3469445.8373426213
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8739957233721811,
+ "character_perplexity": 6.514273701349109,
+ "next_character_accuracy": 0.5036037417573992,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -9.787957139825886e-07,
+ "mean_prediction_kl_nats": 4.398342473752485e-09,
+ "prediction_kl_p99_nats": 3.380791472690338e-08,
+ "maximum_prediction_kl_nats": 9.409525687431273e-08,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.001085519790649414,
+ "maximum_per_target_cross_entropy_increase": 0.0008978843688964844
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 5452876,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 675,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 675,
+ "zlib_blocks": 675,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "b9e46c364634283ab41e551138f1ada23921e31ee48a04dd3989a3da162f7b1f",
+ "table_sha256": "47d81729a61e6f79f28618ae7a0f10d49efd41fdc2fae99dbdf19e801e0afc05",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5259647469501942,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5454083,
+ "backend_with_key_registry_bytes": 6190190,
+ "backend_with_datrie_key_index_bytes": 5800243,
+ "full_model_serving_artifact_bytes": 7018188,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999539845378,
+ "mean_cosine": 0.9999999792071412,
+ "p05_cosine": 0.9999999715996549,
+ "relative_l2_mean": 0.00020637351738737785,
+ "relative_l2_p95": 0.00024209629415095646,
+ "relative_l2_max": 0.0003075455058483477,
+ "weighted_relative_l2_mean": 7.822433519936367e-05,
+ "weighted_relative_frobenius": 0.00020811630848376357,
+ "weighted_mean_cosine": 0.9999999921676113,
+ "max_relative_norm_change": 0.00018268239076675474
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.901175,
+ "warm_p50_ms": 0.146667,
+ "warm_p95_ms": 0.14946705,
+ "warm_p99_ms": 0.16160935,
+ "samples_ms": [
+ 0.149041,
+ 0.164197,
+ 0.145689,
+ 0.146178,
+ 0.147156,
+ 0.148133,
+ 0.147645,
+ 0.146667,
+ 0.147156,
+ 0.148552,
+ 0.148134,
+ 0.14534,
+ 0.145759,
+ 0.147225,
+ 0.147156,
+ 0.146667,
+ 0.147086,
+ 0.147086,
+ 0.146177,
+ 0.146597,
+ 0.145689,
+ 0.147156,
+ 0.145758,
+ 0.146248,
+ 0.147155,
+ 0.146737,
+ 0.147645,
+ 0.157562,
+ 0.147645,
+ 0.145759,
+ 0.145759,
+ 0.1452,
+ 0.146597,
+ 0.145689,
+ 0.146108,
+ 0.14527,
+ 0.1452,
+ 0.146108,
+ 0.146667,
+ 0.146178
+ ],
+ "warm_rows_per_second": 3490901.157042825
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8739957233721811,
+ "character_perplexity": 6.514273701349109,
+ "next_character_accuracy": 0.5036037417573992,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -9.787957139825886e-07,
+ "mean_prediction_kl_nats": 4.398342473752485e-09,
+ "prediction_kl_p99_nats": 3.380791472690338e-08,
+ "maximum_prediction_kl_nats": 9.409525687431273e-08,
+ "top1_changed_fraction": 7.667535654040791e-05,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.001085519790649414,
+ "maximum_per_target_cross_entropy_increase": 0.0008978843688964844
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3877911,
+ "unique_rows": 29798,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29798
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "9d9b25970b5839baa32f46c4eddf5e0afa57cf97b2dc9a28b62711e129a6908c",
+ "table_sha256": "9bbe88814651c53a7c993e9a381fce1e95d41be8f2f7c6bc5c45dd69a37f958d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4008830760139972,
+ "representatives_sha256": "bdaa9b3d678298c3f1664248618576322980c801ca815d8de513407a699cddd1",
+ "structurally_merged_rows": 13354,
+ "backend_artifact_bytes": 3879117,
+ "backend_with_key_registry_bytes": 4615224,
+ "backend_with_datrie_key_index_bytes": 4225277,
+ "full_model_serving_artifact_bytes": 5443222,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.990039451270731,
+ "mean_cosine": 0.9997200119638027,
+ "p05_cosine": 0.9986673654831411,
+ "relative_l2_mean": 0.012667019417472036,
+ "relative_l2_p95": 0.051792523979475126,
+ "relative_l2_max": 0.14112562218626845,
+ "weighted_relative_l2_mean": 0.0020723638537687695,
+ "weighted_relative_frobenius": 0.012524791627337375,
+ "weighted_mean_cosine": 0.9999602331100651,
+ "max_relative_norm_change": 0.010122543655169182
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9900394695708842,
+ "mean_cosine": 0.9997200247157288,
+ "p05_cosine": 0.9986673862818956,
+ "relative_l2_mean": 0.012524519177028527,
+ "relative_l2_p95": 0.05177717868029783,
+ "relative_l2_max": 0.14112538288419477,
+ "weighted_relative_l2_mean": 0.0020059476275550033,
+ "weighted_relative_frobenius": 0.012523978773244526,
+ "weighted_mean_cosine": 0.9999602374746445,
+ "max_relative_norm_change": 0.01009751476207093
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.298026,
+ "warm_p50_ms": 0.1475745,
+ "warm_p95_ms": 0.15098279999999994,
+ "warm_p99_ms": 0.17603133999999998,
+ "samples_ms": [
+ 0.185428,
+ 0.150438,
+ 0.149041,
+ 0.147156,
+ 0.148134,
+ 0.147575,
+ 0.149041,
+ 0.148553,
+ 0.147086,
+ 0.146666,
+ 0.147156,
+ 0.148134,
+ 0.147155,
+ 0.146667,
+ 0.147086,
+ 0.147575,
+ 0.147645,
+ 0.147575,
+ 0.147156,
+ 0.146667,
+ 0.146667,
+ 0.147156,
+ 0.147155,
+ 0.146667,
+ 0.147575,
+ 0.148063,
+ 0.161334,
+ 0.148134,
+ 0.147644,
+ 0.146667,
+ 0.147575,
+ 0.147574,
+ 0.147575,
+ 0.147226,
+ 0.146736,
+ 0.145759,
+ 0.147225,
+ 0.147645,
+ 0.147156,
+ 0.147645
+ ],
+ "warm_rows_per_second": 3469434.082446493
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8740206558292583,
+ "character_perplexity": 6.514436120223299,
+ "next_character_accuracy": 0.5036804171139396,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 2.3953947775293576e-05,
+ "mean_prediction_kl_nats": 7.010273872200154e-06,
+ "prediction_kl_p99_nats": 0.0001103468003908952,
+ "maximum_prediction_kl_nats": 0.006080347595238767,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308086867,
+ "maximum_absolute_logit_change": 0.36091315746307373,
+ "maximum_per_target_cross_entropy_increase": 0.08162784576416016
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 2.4932767893362518e-05,
+ "mean_prediction_kl_nats": 7.003602602614312e-06,
+ "prediction_kl_p99_nats": 0.00011116864650160212,
+ "maximum_prediction_kl_nats": 0.006079277598338604,
+ "top1_changed_fraction": 0.0006134028523232633,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.3609464168548584,
+ "maximum_per_target_cross_entropy_increase": 0.08144903182983398
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3875474,
+ "unique_rows": 29778,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29778
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "782fdd16bf022b85f452a3fa63c38ab86fc43c1e772b1c73e5f0738982fde123",
+ "table_sha256": "79a91ba546e5db38ec196257377e96f0cab6400b051782c78d0d6de8e9240dc5",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.40349130099639297,
+ "representatives_sha256": "daf77d5f2edc82e66988154f46072fa9c85cefe26e5049a12be4d4dff99966b4",
+ "structurally_merged_rows": 13374,
+ "backend_artifact_bytes": 3876680,
+ "backend_with_key_registry_bytes": 4612787,
+ "backend_with_datrie_key_index_bytes": 4222840,
+ "full_model_serving_artifact_bytes": 5440785,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9801798714782947,
+ "mean_cosine": 0.9997099931922108,
+ "p05_cosine": 0.9986591974401371,
+ "relative_l2_mean": 0.012776640825745606,
+ "relative_l2_p95": 0.05196765183434268,
+ "relative_l2_max": 0.19874691425676086,
+ "weighted_relative_l2_mean": 0.0021007940848287313,
+ "weighted_relative_frobenius": 0.013253196770836724,
+ "weighted_mean_cosine": 0.9999578590980495,
+ "max_relative_norm_change": 0.010122543655169182
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9801729436886374,
+ "mean_cosine": 0.9997100055449711,
+ "p05_cosine": 0.9986594034629189,
+ "relative_l2_mean": 0.012634240162831973,
+ "relative_l2_p95": 0.05195050037136643,
+ "relative_l2_max": 0.19878048097632814,
+ "weighted_relative_l2_mean": 0.002034410223639944,
+ "weighted_relative_frobenius": 0.01325250282123261,
+ "weighted_mean_cosine": 0.9999578633213264,
+ "max_relative_norm_change": 0.01009751476207093
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.255353,
+ "warm_p50_ms": 0.14785399999999999,
+ "warm_p95_ms": 0.15053229999999998,
+ "warm_p99_ms": 0.16041869999999997,
+ "samples_ms": [
+ 0.150438,
+ 0.149041,
+ 0.149041,
+ 0.148063,
+ 0.146248,
+ 0.148134,
+ 0.147574,
+ 0.148134,
+ 0.148064,
+ 0.146667,
+ 0.148064,
+ 0.147156,
+ 0.148134,
+ 0.148063,
+ 0.147575,
+ 0.146667,
+ 0.146178,
+ 0.148064,
+ 0.165594,
+ 0.148064,
+ 0.147645,
+ 0.146178,
+ 0.146667,
+ 0.148063,
+ 0.148063,
+ 0.146248,
+ 0.146247,
+ 0.14953,
+ 0.148133,
+ 0.147645,
+ 0.146667,
+ 0.147156,
+ 0.148064,
+ 0.147086,
+ 0.152324,
+ 0.147086,
+ 0.148064,
+ 0.146178,
+ 0.145689,
+ 0.146667
+ ],
+ "warm_rows_per_second": 3462875.539383446
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8740154708462469,
+ "character_perplexity": 6.514402343070254,
+ "next_character_accuracy": 0.50375709247048,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.8768854005388793e-05,
+ "mean_prediction_kl_nats": 1.4386399137375463e-05,
+ "prediction_kl_p99_nats": 0.00012141775829092279,
+ "maximum_prediction_kl_nats": 0.022824656101379563,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 0.36091315746307373,
+ "maximum_per_target_cross_entropy_increase": 0.35405421257019043
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 1.9747669048305204e-05,
+ "mean_prediction_kl_nats": 1.4377007304061832e-05,
+ "prediction_kl_p99_nats": 0.00012129024920230822,
+ "maximum_prediction_kl_nats": 0.0228160493711172,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.3609464168548584,
+ "maximum_per_target_cross_entropy_increase": 0.3538210391998291
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3874277,
+ "unique_rows": 29768,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29768
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "9bd68315e40ae307540c7d256d24998d9d3202ef99e8f5166e7fe238096854d7",
+ "table_sha256": "6aed3044f8176c82215592f19b3fe5e54556a8974232dfbcfa9fc41c56806047",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4015690570231527,
+ "representatives_sha256": "4ed7525d92e17bb55e1ad0327daad3d71d881ca5f358f7438bf7f26113916e33",
+ "structurally_merged_rows": 13384,
+ "backend_artifact_bytes": 3875483,
+ "backend_with_key_registry_bytes": 4611590,
+ "backend_with_datrie_key_index_bytes": 4221643,
+ "full_model_serving_artifact_bytes": 5439588,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9700413472050464,
+ "mean_cosine": 0.9997035559009319,
+ "p05_cosine": 0.998656604137302,
+ "relative_l2_mean": 0.012832814803887876,
+ "relative_l2_p95": 0.052022599300334385,
+ "relative_l2_max": 0.24497778247547566,
+ "weighted_relative_l2_mean": 0.0021434550655841113,
+ "weighted_relative_frobenius": 0.014136114202284165,
+ "weighted_mean_cosine": 0.9999530878739694,
+ "max_relative_norm_change": 0.010122543655169182
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9700313980960931,
+ "mean_cosine": 0.999703568234334,
+ "p05_cosine": 0.9986563464479082,
+ "relative_l2_mean": 0.012690462636334787,
+ "relative_l2_p95": 0.05202736453606879,
+ "relative_l2_max": 0.245019579774079,
+ "weighted_relative_l2_mean": 0.0020771063840673383,
+ "weighted_relative_frobenius": 0.014135485364462215,
+ "weighted_mean_cosine": 0.9999530921948487,
+ "max_relative_norm_change": 0.01009751476207093
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.326869,
+ "warm_p50_ms": 0.14736500000000002,
+ "warm_p95_ms": 0.14915294999999998,
+ "warm_p99_ms": 0.15864066999999998,
+ "samples_ms": [
+ 0.14995,
+ 0.148064,
+ 0.148134,
+ 0.146666,
+ 0.147645,
+ 0.148134,
+ 0.148482,
+ 0.148133,
+ 0.147575,
+ 0.148064,
+ 0.148063,
+ 0.148064,
+ 0.147645,
+ 0.147155,
+ 0.147156,
+ 0.148134,
+ 0.146178,
+ 0.149111,
+ 0.148623,
+ 0.147156,
+ 0.164197,
+ 0.148064,
+ 0.148064,
+ 0.146736,
+ 0.147575,
+ 0.147574,
+ 0.146178,
+ 0.146667,
+ 0.146248,
+ 0.146667,
+ 0.146667,
+ 0.146667,
+ 0.146597,
+ 0.146667,
+ 0.146178,
+ 0.147156,
+ 0.146736,
+ 0.146667,
+ 0.146736,
+ 0.146667
+ ],
+ "warm_rows_per_second": 3474366.369219285
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8738598211558746,
+ "character_perplexity": 6.513388457270146,
+ "next_character_accuracy": 0.5039104431835608,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0001368716447210878,
+ "mean_prediction_kl_nats": 3.6732209213622236e-05,
+ "prediction_kl_p99_nats": 0.00015693520409458672,
+ "maximum_prediction_kl_nats": 0.033190273506776724,
+ "top1_changed_fraction": 0.000996779635025303,
+ "accuracy_change": 0.00038337678270206066,
+ "maximum_absolute_logit_change": 0.40544557571411133,
+ "maximum_per_target_cross_entropy_increase": 0.37369251251220703
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.00013589298201857356,
+ "mean_prediction_kl_nats": 3.673704880259792e-05,
+ "prediction_kl_p99_nats": 0.00015812882003475985,
+ "maximum_prediction_kl_nats": 0.03321187035576082,
+ "top1_changed_fraction": 0.0009201042784848949,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.40515899658203125,
+ "maximum_per_target_cross_entropy_increase": 0.3736705780029297
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3872396,
+ "unique_rows": 29752,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 465,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 465,
+ "zlib_blocks": 465,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29752
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "c1894e6d14d5007517c3fb7175d56704105bc24327207f4d75212cc5133dbfc5",
+ "table_sha256": "0adfe65ef6a0467d9e0a6275605eb7935e2535b649cbbc56d8654e12a6ddc2bc",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4069029809907079,
+ "representatives_sha256": "0c7cc9af822d1589ee21fce800aac5ed8e254f42d9fac525e20b9dd249062fcf",
+ "structurally_merged_rows": 13400,
+ "backend_artifact_bytes": 3873602,
+ "backend_with_key_registry_bytes": 4609709,
+ "backend_with_datrie_key_index_bytes": 4219762,
+ "full_model_serving_artifact_bytes": 5437707,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9511896509087922,
+ "mean_cosine": 0.9996853683765607,
+ "p05_cosine": 0.998652332990108,
+ "relative_l2_mean": 0.012963202589690545,
+ "relative_l2_p95": 0.05210069536939933,
+ "relative_l2_max": 0.3119094889464882,
+ "weighted_relative_l2_mean": 0.0021843990687356023,
+ "weighted_relative_frobenius": 0.01549541807315949,
+ "weighted_mean_cosine": 0.9999475954868816,
+ "max_relative_norm_change": 0.010122543655169182
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9512025302127041,
+ "mean_cosine": 0.9996853806504409,
+ "p05_cosine": 0.9986528086901074,
+ "relative_l2_mean": 0.012820928349103153,
+ "relative_l2_p95": 0.05210575018214038,
+ "relative_l2_max": 0.31186810336214976,
+ "weighted_relative_l2_mean": 0.002118075869077449,
+ "weighted_relative_frobenius": 0.015494646561231546,
+ "weighted_mean_cosine": 0.9999476007700012,
+ "max_relative_norm_change": 0.01009751476207093
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.180622,
+ "warm_p50_ms": 0.1476095,
+ "warm_p95_ms": 0.15099685,
+ "warm_p99_ms": 0.15603036,
+ "samples_ms": [
+ 0.149949,
+ 0.148622,
+ 0.149041,
+ 0.14946,
+ 0.150508,
+ 0.148063,
+ 0.148134,
+ 0.148064,
+ 0.146666,
+ 0.147156,
+ 0.152324,
+ 0.147156,
+ 0.147156,
+ 0.147155,
+ 0.145829,
+ 0.147155,
+ 0.147156,
+ 0.147575,
+ 0.147644,
+ 0.147156,
+ 0.1584,
+ 0.147645,
+ 0.147645,
+ 0.147644,
+ 0.146108,
+ 0.148064,
+ 0.146667,
+ 0.147156,
+ 0.147644,
+ 0.145759,
+ 0.150438,
+ 0.14527,
+ 0.147575,
+ 0.146248,
+ 0.147645,
+ 0.147156,
+ 0.146248,
+ 0.150927,
+ 0.147645,
+ 0.145689
+ ],
+ "warm_rows_per_second": 3468611.437610723
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.873704802799334,
+ "character_perplexity": 6.512378840752558,
+ "next_character_accuracy": 0.5038337678270204,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0002918567705643621,
+ "mean_prediction_kl_nats": 7.006744225829511e-05,
+ "prediction_kl_p99_nats": 0.0002180392407785873,
+ "maximum_prediction_kl_nats": 0.16059932138241623,
+ "top1_changed_fraction": 0.0010734549915657108,
+ "accuracy_change": 0.0003067014261616263,
+ "maximum_absolute_logit_change": 0.6744060516357422,
+ "maximum_per_target_cross_entropy_increase": 0.704634964466095
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.00029087825956077327,
+ "mean_prediction_kl_nats": 7.008232032422754e-05,
+ "prediction_kl_p99_nats": 0.00021922514854452263,
+ "maximum_prediction_kl_nats": 0.16065407099670095,
+ "top1_changed_fraction": 0.000996779635025303,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 0.6747474670410156,
+ "maximum_per_target_cross_entropy_increase": 0.7047379016876221
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3868961,
+ "unique_rows": 29721,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 465,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 465,
+ "zlib_blocks": 465,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29721
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "5f67c5c9174b17f9a730562a640f6edddc1fb2410ce66b1209ed34b965469a37",
+ "table_sha256": "871d86a620030f43e4f7cc52a88633bc1c53230496a9c57a2d84b023ee409808",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.40092847193591297,
+ "representatives_sha256": "a2ea579cb5a89fa4b15bac149075de300a30d73418c9b6b5e8a088103115f9b7",
+ "structurally_merged_rows": 13431,
+ "backend_artifact_bytes": 3870167,
+ "backend_with_key_registry_bytes": 4606274,
+ "backend_with_datrie_key_index_bytes": 4216327,
+ "full_model_serving_artifact_bytes": 5434272,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9025562211609032,
+ "mean_cosine": 0.9996050977484005,
+ "p05_cosine": 0.998641546072824,
+ "relative_l2_mean": 0.013367946057538882,
+ "relative_l2_p95": 0.052355489437242816,
+ "relative_l2_max": 0.44056088115148856,
+ "weighted_relative_l2_mean": 0.0023998581099841584,
+ "weighted_relative_frobenius": 0.02303287968306604,
+ "weighted_mean_cosine": 0.9999046537840203,
+ "max_relative_norm_change": 0.010122543655169182
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9025561051703608,
+ "mean_cosine": 0.9996051090786908,
+ "p05_cosine": 0.9986406985794307,
+ "relative_l2_mean": 0.013225819021885743,
+ "relative_l2_p95": 0.05235133994154282,
+ "relative_l2_max": 0.44059198975890207,
+ "weighted_relative_l2_mean": 0.002333617266413956,
+ "weighted_relative_frobenius": 0.023032520866436014,
+ "weighted_mean_cosine": 0.9999046585494379,
+ "max_relative_norm_change": 0.01009751476207093
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.272882,
+ "warm_p50_ms": 0.146911,
+ "warm_p95_ms": 0.14981974999999997,
+ "warm_p99_ms": 0.1753093,
+ "samples_ms": [
+ 0.164616,
+ 0.149041,
+ 0.148064,
+ 0.148134,
+ 0.147575,
+ 0.147505,
+ 0.147645,
+ 0.147226,
+ 0.146248,
+ 0.145759,
+ 0.147645,
+ 0.146248,
+ 0.146178,
+ 0.146667,
+ 0.146248,
+ 0.146667,
+ 0.146667,
+ 0.148064,
+ 0.145689,
+ 0.146178,
+ 0.147086,
+ 0.1452,
+ 0.147156,
+ 0.147086,
+ 0.146667,
+ 0.147645,
+ 0.146736,
+ 0.182146,
+ 0.147086,
+ 0.146597,
+ 0.146667,
+ 0.146597,
+ 0.147156,
+ 0.146248,
+ 0.146248,
+ 0.148064,
+ 0.147575,
+ 0.146666,
+ 0.146667,
+ 0.147156
+ ],
+ "warm_rows_per_second": 3485103.2257625363
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874134437886126,
+ "character_perplexity": 6.515177388335096,
+ "next_character_accuracy": 0.504140469253182,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0001377452037514349,
+ "mean_prediction_kl_nats": 0.00026050593387185866,
+ "prediction_kl_p99_nats": 0.0009142191542520585,
+ "maximum_prediction_kl_nats": 0.4085835746528589,
+ "top1_changed_fraction": 0.0015335071308081583,
+ "accuracy_change": 0.0006134028523232526,
+ "maximum_absolute_logit_change": 1.6836234331130981,
+ "maximum_per_target_cross_entropy_increase": 1.2192572355270386
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.00013872413524800646,
+ "mean_prediction_kl_nats": 0.000260535456554229,
+ "prediction_kl_p99_nats": 0.0009141126919950614,
+ "maximum_prediction_kl_nats": 0.4085141934177651,
+ "top1_changed_fraction": 0.0014568317742677504,
+ "accuracy_change": 0.0005367274957828183,
+ "maximum_absolute_logit_change": 1.683741807937622,
+ "maximum_per_target_cross_entropy_increase": 1.2193512916564941
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 68,
+ "storage": {
+ "artifact_bytes": 2934336,
+ "unique_rows": 43152,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "bb7617c7380056f8d86ab747df137ba8166ea9c736901a80e4911c8d837e3aa2",
+ "table_sha256": "61f44b3bbe97af40b409ed69d9a05639fccb6f8c969bbfc0894eb7d256edc14c",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06563159497454762,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2935184,
+ "backend_with_key_registry_bytes": 3671291,
+ "backend_with_datrie_key_index_bytes": 3281344,
+ "full_model_serving_artifact_bytes": 4499289,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999336997816471,
+ "mean_cosine": 0.9999885057292354,
+ "p05_cosine": 0.9999772675626297,
+ "relative_l2_mean": 0.004687375390845569,
+ "relative_l2_p95": 0.006792605888808079,
+ "relative_l2_max": 0.011515823593947355,
+ "weighted_relative_l2_mean": 0.00203790374860136,
+ "weighted_relative_frobenius": 0.0053141135041913715,
+ "weighted_mean_cosine": 0.9999943497303216,
+ "max_relative_norm_change": 0.0029012066989757127
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.278667,
+ "warm_p50_ms": 0.095578,
+ "warm_p95_ms": 0.0980534,
+ "warm_p99_ms": 0.11980672999999997,
+ "samples_ms": [
+ 0.09659,
+ 0.097987,
+ 0.094635,
+ 0.095683,
+ 0.096521,
+ 0.095613,
+ 0.095543,
+ 0.096032,
+ 0.094565,
+ 0.094705,
+ 0.09659,
+ 0.096102,
+ 0.096521,
+ 0.094635,
+ 0.095473,
+ 0.095543,
+ 0.095123,
+ 0.095613,
+ 0.132908,
+ 0.099315,
+ 0.095054,
+ 0.096451,
+ 0.095962,
+ 0.094565,
+ 0.094147,
+ 0.095613,
+ 0.095613,
+ 0.095054,
+ 0.094565,
+ 0.094146,
+ 0.094635,
+ 0.09659,
+ 0.095613,
+ 0.094146,
+ 0.094076,
+ 0.096032,
+ 0.095055,
+ 0.094705,
+ 0.093657,
+ 0.097009
+ ],
+ "warm_rows_per_second": 5356881.290673586
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8739410775865617,
+ "character_perplexity": 6.513917733471119,
+ "next_character_accuracy": 0.503373715687778,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -5.5623034793995104e-05,
+ "mean_prediction_kl_nats": 3.7121507843531695e-06,
+ "prediction_kl_p99_nats": 2.7568171146506205e-05,
+ "maximum_prediction_kl_nats": 0.0001025896388855517,
+ "top1_changed_fraction": 0.000843428921944487,
+ "accuracy_change": -0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.02931845188140869,
+ "maximum_per_target_cross_entropy_increase": 0.022075176239013672
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 3288858,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 360,
+ "max_block_bytes": 8192,
+ "raw_blocks": 189,
+ "compressed_blocks": 171,
+ "zlib_blocks": 171,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "bb7617c7380056f8d86ab747df137ba8166ea9c736901a80e4911c8d837e3aa2",
+ "table_sha256": "74f913424babe120d7ec58657fcaa8254d64ec23a6f925f5caa137efd6a1ebba",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.27589925192296505,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3290111,
+ "backend_with_key_registry_bytes": 4026218,
+ "backend_with_datrie_key_index_bytes": 3636271,
+ "full_model_serving_artifact_bytes": 4854216,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999336997816471,
+ "mean_cosine": 0.9999885057292354,
+ "p05_cosine": 0.9999772675626297,
+ "relative_l2_mean": 0.004687375390845569,
+ "relative_l2_p95": 0.006792605888808079,
+ "relative_l2_max": 0.011515823593947355,
+ "weighted_relative_l2_mean": 0.00203790374860136,
+ "weighted_relative_frobenius": 0.0053141135041913715,
+ "weighted_mean_cosine": 0.9999943497303216,
+ "max_relative_norm_change": 0.0029012066989757127
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.276688,
+ "warm_p50_ms": 0.111676,
+ "warm_p95_ms": 0.1168729,
+ "warm_p99_ms": 0.13844921999999998,
+ "samples_ms": [
+ 0.11447,
+ 0.113143,
+ 0.111746,
+ 0.11028,
+ 0.112096,
+ 0.110698,
+ 0.112165,
+ 0.111676,
+ 0.112165,
+ 0.111606,
+ 0.111117,
+ 0.111118,
+ 0.113562,
+ 0.110699,
+ 0.110768,
+ 0.151905,
+ 0.116845,
+ 0.111746,
+ 0.110628,
+ 0.111187,
+ 0.117403,
+ 0.112584,
+ 0.110209,
+ 0.11021,
+ 0.112584,
+ 0.111257,
+ 0.112584,
+ 0.111676,
+ 0.112514,
+ 0.110698,
+ 0.111746,
+ 0.111257,
+ 0.112165,
+ 0.111187,
+ 0.111607,
+ 0.111187,
+ 0.111606,
+ 0.114051,
+ 0.112165,
+ 0.11028
+ ],
+ "warm_rows_per_second": 4584691.428776102
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8739410775865617,
+ "character_perplexity": 6.513917733471119,
+ "next_character_accuracy": 0.503373715687778,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -5.5623034793995104e-05,
+ "mean_prediction_kl_nats": 3.7121507843531695e-06,
+ "prediction_kl_p99_nats": 2.7568171146506205e-05,
+ "maximum_prediction_kl_nats": 0.0001025896388855517,
+ "top1_changed_fraction": 0.000843428921944487,
+ "accuracy_change": -0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.02931845188140869,
+ "maximum_per_target_cross_entropy_increase": 0.022075176239013672
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 3288858,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 360,
+ "max_block_bytes": 8192,
+ "raw_blocks": 189,
+ "compressed_blocks": 171,
+ "zlib_blocks": 171,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "bb7617c7380056f8d86ab747df137ba8166ea9c736901a80e4911c8d837e3aa2",
+ "table_sha256": "74f913424babe120d7ec58657fcaa8254d64ec23a6f925f5caa137efd6a1ebba",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.43378512701019645,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3290110,
+ "backend_with_key_registry_bytes": 4026217,
+ "backend_with_datrie_key_index_bytes": 3636270,
+ "full_model_serving_artifact_bytes": 4854215,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999336997816471,
+ "mean_cosine": 0.9999885057292354,
+ "p05_cosine": 0.9999772675626297,
+ "relative_l2_mean": 0.004687375390845569,
+ "relative_l2_p95": 0.006792605888808079,
+ "relative_l2_max": 0.011515823593947355,
+ "weighted_relative_l2_mean": 0.00203790374860136,
+ "weighted_relative_frobenius": 0.0053141135041913715,
+ "weighted_mean_cosine": 0.9999943497303216,
+ "max_relative_norm_change": 0.0029012066989757127
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.233666,
+ "warm_p50_ms": 0.112095,
+ "warm_p95_ms": 0.11398545,
+ "warm_p99_ms": 0.13829239999999998,
+ "samples_ms": [
+ 0.114051,
+ 0.113981,
+ 0.113981,
+ 0.111606,
+ 0.110698,
+ 0.113562,
+ 0.112165,
+ 0.111677,
+ 0.153791,
+ 0.113562,
+ 0.112235,
+ 0.112096,
+ 0.112515,
+ 0.112026,
+ 0.112096,
+ 0.111257,
+ 0.112654,
+ 0.113073,
+ 0.111746,
+ 0.112095,
+ 0.111606,
+ 0.112095,
+ 0.112095,
+ 0.113982,
+ 0.111606,
+ 0.112584,
+ 0.112654,
+ 0.111257,
+ 0.111187,
+ 0.111606,
+ 0.112165,
+ 0.111676,
+ 0.111676,
+ 0.110279,
+ 0.111606,
+ 0.113073,
+ 0.111606,
+ 0.111676,
+ 0.113562,
+ 0.111258
+ ],
+ "warm_rows_per_second": 4567554.306614925
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8739410775865617,
+ "character_perplexity": 6.513917733471119,
+ "next_character_accuracy": 0.503373715687778,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -5.5623034793995104e-05,
+ "mean_prediction_kl_nats": 3.7121507843531695e-06,
+ "prediction_kl_p99_nats": 2.7568171146506205e-05,
+ "maximum_prediction_kl_nats": 0.0001025896388855517,
+ "top1_changed_fraction": 0.000843428921944487,
+ "accuracy_change": -0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.02931845188140869,
+ "maximum_per_target_cross_entropy_increase": 0.022075176239013672
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 3288858,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 360,
+ "max_block_bytes": 8192,
+ "raw_blocks": 189,
+ "compressed_blocks": 171,
+ "zlib_blocks": 171,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "bb7617c7380056f8d86ab747df137ba8166ea9c736901a80e4911c8d837e3aa2",
+ "table_sha256": "74f913424babe120d7ec58657fcaa8254d64ec23a6f925f5caa137efd6a1ebba",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2810117790941149,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3290110,
+ "backend_with_key_registry_bytes": 4026217,
+ "backend_with_datrie_key_index_bytes": 3636270,
+ "full_model_serving_artifact_bytes": 4854215,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999336997816471,
+ "mean_cosine": 0.9999885057292354,
+ "p05_cosine": 0.9999772675626297,
+ "relative_l2_mean": 0.004687375390845569,
+ "relative_l2_p95": 0.006792605888808079,
+ "relative_l2_max": 0.011515823593947355,
+ "weighted_relative_l2_mean": 0.00203790374860136,
+ "weighted_relative_frobenius": 0.0053141135041913715,
+ "weighted_mean_cosine": 0.9999943497303216,
+ "max_relative_norm_change": 0.0029012066989757127
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.260974,
+ "warm_p50_ms": 0.1111875,
+ "warm_p95_ms": 0.1131673,
+ "warm_p99_ms": 0.11900635,
+ "samples_ms": [
+ 0.114959,
+ 0.111747,
+ 0.11021,
+ 0.11014,
+ 0.110768,
+ 0.113073,
+ 0.112724,
+ 0.112654,
+ 0.110699,
+ 0.110699,
+ 0.111187,
+ 0.110768,
+ 0.112584,
+ 0.110698,
+ 0.110279,
+ 0.121594,
+ 0.113073,
+ 0.111257,
+ 0.111257,
+ 0.111676,
+ 0.110698,
+ 0.111188,
+ 0.111746,
+ 0.111188,
+ 0.111188,
+ 0.110629,
+ 0.111258,
+ 0.110699,
+ 0.110629,
+ 0.110209,
+ 0.111257,
+ 0.110768,
+ 0.111118,
+ 0.111606,
+ 0.110768,
+ 0.110769,
+ 0.111677,
+ 0.111118,
+ 0.111747,
+ 0.11028
+ ],
+ "warm_rows_per_second": 4604834.176503654
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8739410775865617,
+ "character_perplexity": 6.513917733471119,
+ "next_character_accuracy": 0.503373715687778,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -5.5623034793995104e-05,
+ "mean_prediction_kl_nats": 3.7121507843531695e-06,
+ "prediction_kl_p99_nats": 2.7568171146506205e-05,
+ "maximum_prediction_kl_nats": 0.0001025896388855517,
+ "top1_changed_fraction": 0.000843428921944487,
+ "accuracy_change": -0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.02931845188140869,
+ "maximum_per_target_cross_entropy_increase": 0.022075176239013672
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2379211,
+ "unique_rows": 29837,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 73,
+ "compressed_blocks": 176,
+ "zlib_blocks": 176,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29837
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "8172286fbc5317af5b5a2ba6b1ef2e8ddcb94131f7f3161c1742eed051584cae",
+ "table_sha256": "55b9bbcc7c879aec16b4c357810006a72979a7ccdd95a1549ac3022e6d07114f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2349366710986942,
+ "representatives_sha256": "66633aaa9fdad1581dcdb1202f2b849354bcb19e5a8d02743679b259add998bf",
+ "structurally_merged_rows": 13315,
+ "backend_artifact_bytes": 2380461,
+ "backend_with_key_registry_bytes": 3116568,
+ "backend_with_datrie_key_index_bytes": 2726621,
+ "full_model_serving_artifact_bytes": 3944566,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9900799715264293,
+ "mean_cosine": 0.99970954983572,
+ "p05_cosine": 0.9986702491886431,
+ "relative_l2_mean": 0.015874600816270342,
+ "relative_l2_p95": 0.05179789153831649,
+ "relative_l2_max": 0.14050498208265808,
+ "weighted_relative_l2_mean": 0.0037786874534360363,
+ "weighted_relative_frobenius": 0.013529571057738162,
+ "weighted_mean_cosine": 0.9999551225814726,
+ "max_relative_norm_change": 0.011715558658246757
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9901108468632621,
+ "mean_cosine": 0.9997145714319482,
+ "p05_cosine": 0.9986559352308194,
+ "relative_l2_mean": 0.01264815975090175,
+ "relative_l2_p95": 0.052037815749830345,
+ "relative_l2_max": 0.14028726985128304,
+ "weighted_relative_l2_mean": 0.0020385130651573904,
+ "weighted_relative_frobenius": 0.01272182586954238,
+ "weighted_mean_cosine": 0.9999593816082257,
+ "max_relative_norm_change": 0.01009139470261335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.947096,
+ "warm_p50_ms": 0.107905,
+ "warm_p95_ms": 0.1122104,
+ "warm_p99_ms": 0.1185691,
+ "samples_ms": [
+ 0.111606,
+ 0.109371,
+ 0.107416,
+ 0.107416,
+ 0.107835,
+ 0.112165,
+ 0.109791,
+ 0.109232,
+ 0.107904,
+ 0.108394,
+ 0.10979,
+ 0.106857,
+ 0.107416,
+ 0.107416,
+ 0.106857,
+ 0.107416,
+ 0.106997,
+ 0.109302,
+ 0.108813,
+ 0.108813,
+ 0.106438,
+ 0.108813,
+ 0.106438,
+ 0.107835,
+ 0.122083,
+ 0.107416,
+ 0.108324,
+ 0.107905,
+ 0.109791,
+ 0.113073,
+ 0.107835,
+ 0.107416,
+ 0.10595,
+ 0.108813,
+ 0.108813,
+ 0.10553,
+ 0.108323,
+ 0.107905,
+ 0.107416,
+ 0.107905
+ ],
+ "warm_rows_per_second": 4744914.508132153
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8739633755917202,
+ "character_perplexity": 6.514062982461717,
+ "next_character_accuracy": 0.503373715687778,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.332602132978352e-05,
+ "mean_prediction_kl_nats": 1.0863494809786329e-05,
+ "prediction_kl_p99_nats": 0.00011752400598978473,
+ "maximum_prediction_kl_nats": 0.0061883755217517255,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.3648056983947754,
+ "maximum_per_target_cross_entropy_increase": 0.08214378356933594
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 2.2298253760756368e-05,
+ "mean_prediction_kl_nats": 7.166679384055505e-06,
+ "prediction_kl_p99_nats": 0.00011104972881125107,
+ "maximum_prediction_kl_nats": 0.006195155670125636,
+ "top1_changed_fraction": 0.00030670142616163164,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.3624382019042969,
+ "maximum_per_target_cross_entropy_increase": 0.08258438110351562
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2377685,
+ "unique_rows": 29817,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 79,
+ "compressed_blocks": 170,
+ "zlib_blocks": 170,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29817
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "3832fe98629b0bc56baec9e64b59a9982557d483fe1a00beba892840105230fe",
+ "table_sha256": "4435869c4ab0d12beb1785a10160d0b8865ebf9f287588f8469a114e7626355f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24943161406554282,
+ "representatives_sha256": "fad753c23f1e896c85de1704a845834cf5cf67e4a24a73f12602edec0a7162ac",
+ "structurally_merged_rows": 13335,
+ "backend_artifact_bytes": 2378935,
+ "backend_with_key_registry_bytes": 3115042,
+ "backend_with_datrie_key_index_bytes": 2725095,
+ "full_model_serving_artifact_bytes": 3943040,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9801011924241253,
+ "mean_cosine": 0.999700031559149,
+ "p05_cosine": 0.9986624426021717,
+ "relative_l2_mean": 0.015977981942059245,
+ "relative_l2_p95": 0.051972065581319886,
+ "relative_l2_max": 0.1990890449393031,
+ "weighted_relative_l2_mean": 0.00380756158251928,
+ "weighted_relative_frobenius": 0.014230220306088556,
+ "weighted_mean_cosine": 0.999952645760921,
+ "max_relative_norm_change": 0.011715558658246757
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9802193115407607,
+ "mean_cosine": 0.9997050415370483,
+ "p05_cosine": 0.9986480055049344,
+ "relative_l2_mean": 0.012753878617498595,
+ "relative_l2_p95": 0.05220068912360454,
+ "relative_l2_max": 0.19849031765558645,
+ "weighted_relative_l2_mean": 0.0020682031704189092,
+ "weighted_relative_frobenius": 0.01346662504677801,
+ "weighted_mean_cosine": 0.9999568989546193,
+ "max_relative_norm_change": 0.01009139470261335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.821172,
+ "warm_p50_ms": 0.107905,
+ "warm_p95_ms": 0.1115395,
+ "warm_p99_ms": 0.13243932999999997,
+ "samples_ms": [
+ 0.110698,
+ 0.10979,
+ 0.111536,
+ 0.109371,
+ 0.111188,
+ 0.109232,
+ 0.108393,
+ 0.108813,
+ 0.108813,
+ 0.109232,
+ 0.108813,
+ 0.107347,
+ 0.108813,
+ 0.106857,
+ 0.107905,
+ 0.107416,
+ 0.107416,
+ 0.108813,
+ 0.145759,
+ 0.111606,
+ 0.107416,
+ 0.106928,
+ 0.107416,
+ 0.106927,
+ 0.106927,
+ 0.108813,
+ 0.106997,
+ 0.107416,
+ 0.106927,
+ 0.107416,
+ 0.108813,
+ 0.106857,
+ 0.107905,
+ 0.108254,
+ 0.107904,
+ 0.107835,
+ 0.107905,
+ 0.109791,
+ 0.107416,
+ 0.107416
+ ],
+ "warm_rows_per_second": 4744914.508132153
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.873958320818782,
+ "character_perplexity": 6.514030055435655,
+ "next_character_accuracy": 0.5034503910443183,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.838061303750926e-05,
+ "mean_prediction_kl_nats": 1.8468145160849747e-05,
+ "prediction_kl_p99_nats": 0.0001224622658852388,
+ "maximum_prediction_kl_nats": 0.023867958705451445,
+ "top1_changed_fraction": 0.0013034810611869344,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.3648056983947754,
+ "maximum_per_target_cross_entropy_increase": 0.36011672019958496
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 1.724338088566081e-05,
+ "mean_prediction_kl_nats": 1.449889472352411e-05,
+ "prediction_kl_p99_nats": 0.00012051635214573005,
+ "maximum_prediction_kl_nats": 0.02279089290362372,
+ "top1_changed_fraction": 0.00046005213924244745,
+ "accuracy_change": 7.667535654032331e-05,
+ "maximum_absolute_logit_change": 0.3624382019042969,
+ "maximum_per_target_cross_entropy_increase": 0.35428285598754883
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2377065,
+ "unique_rows": 29807,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 84,
+ "compressed_blocks": 165,
+ "zlib_blocks": 165,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29807
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "c47dc06f1ba35107cb854b2515844653fe49cdcd479cc9316b7e836a0aa6a124",
+ "table_sha256": "54bb8be7412ec039fda0c3fd9b57779e0d9384cae1666424f3994f5dd9edcb3f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.234727076953277,
+ "representatives_sha256": "4e9835811d47ac40dae8f84d884a1e10754685c150251171c683590f8842ba5e",
+ "structurally_merged_rows": 13345,
+ "backend_artifact_bytes": 2378315,
+ "backend_with_key_registry_bytes": 3114422,
+ "backend_with_datrie_key_index_bytes": 2724475,
+ "full_model_serving_artifact_bytes": 3942420,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.970227681619455,
+ "mean_cosine": 0.999693591204922,
+ "p05_cosine": 0.9986577694002984,
+ "relative_l2_mean": 0.016032968614897978,
+ "relative_l2_p95": 0.05205299820698988,
+ "relative_l2_max": 0.24421754617070648,
+ "weighted_relative_l2_mean": 0.003849294114434825,
+ "weighted_relative_frobenius": 0.015056704117102472,
+ "weighted_mean_cosine": 0.9999478707056814,
+ "max_relative_norm_change": 0.011715558658246757
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.970287941034826,
+ "mean_cosine": 0.9996985992315968,
+ "p05_cosine": 0.9986450501833666,
+ "relative_l2_mean": 0.012810108164801652,
+ "relative_l2_p95": 0.05227298614602889,
+ "relative_l2_max": 0.24405115715268316,
+ "weighted_relative_l2_mean": 0.0021109493811355957,
+ "weighted_relative_frobenius": 0.014339566466190019,
+ "weighted_mean_cosine": 0.9999521151639343,
+ "max_relative_norm_change": 0.01009139470261335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.606409,
+ "warm_p50_ms": 0.1088125,
+ "warm_p95_ms": 0.1121204,
+ "warm_p99_ms": 0.13622881999999997,
+ "samples_ms": [
+ 0.112584,
+ 0.111187,
+ 0.109301,
+ 0.108813,
+ 0.108324,
+ 0.109651,
+ 0.109861,
+ 0.109791,
+ 0.10979,
+ 0.108813,
+ 0.108883,
+ 0.109302,
+ 0.109302,
+ 0.107905,
+ 0.108464,
+ 0.108324,
+ 0.108394,
+ 0.151346,
+ 0.112096,
+ 0.108813,
+ 0.107835,
+ 0.109301,
+ 0.108324,
+ 0.108324,
+ 0.106438,
+ 0.106509,
+ 0.108324,
+ 0.111676,
+ 0.106997,
+ 0.106439,
+ 0.106927,
+ 0.108813,
+ 0.108324,
+ 0.107905,
+ 0.108883,
+ 0.107905,
+ 0.108812,
+ 0.109302,
+ 0.107905,
+ 0.107905
+ ],
+ "warm_rows_per_second": 4705341.757610568
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8738035119523269,
+ "character_perplexity": 6.513021703879613,
+ "next_character_accuracy": 0.5036037417573992,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.0001931715560190487,
+ "mean_prediction_kl_nats": 4.1061383926831704e-05,
+ "prediction_kl_p99_nats": 0.00017237435872760937,
+ "maximum_prediction_kl_nats": 0.033299796541104196,
+ "top1_changed_fraction": 0.0015335071308081583,
+ "accuracy_change": 7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.4103102684020996,
+ "maximum_per_target_cross_entropy_increase": 0.37328028678894043
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.00013755617251682363,
+ "mean_prediction_kl_nats": 3.6899726929217145e-05,
+ "prediction_kl_p99_nats": 0.00015571677171428625,
+ "maximum_prediction_kl_nats": 0.03323540901279816,
+ "top1_changed_fraction": 0.0006900782088636713,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 0.4097323417663574,
+ "maximum_per_target_cross_entropy_increase": 0.3750145435333252
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2375941,
+ "unique_rows": 29792,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 87,
+ "compressed_blocks": 162,
+ "zlib_blocks": 162,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29792
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "1f17036d03459d64386d4310cb76367eeab48cda2adfef72dc6f36ff130db487",
+ "table_sha256": "05fff4682d17a7232c2308550ce20bd485b1097e49e3abbeb413dc0612e783b1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24294607690535486,
+ "representatives_sha256": "7e5a661f1aaae7116344383645ae9fbe45d1cb4371687e75f4f87c38c48000e3",
+ "structurally_merged_rows": 13360,
+ "backend_artifact_bytes": 2377191,
+ "backend_with_key_registry_bytes": 3113298,
+ "backend_with_datrie_key_index_bytes": 2723351,
+ "full_model_serving_artifact_bytes": 3941296,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9511035339527236,
+ "mean_cosine": 0.9996762788379485,
+ "p05_cosine": 0.9986487496298613,
+ "relative_l2_mean": 0.016155183183444766,
+ "relative_l2_p95": 0.05214904276097739,
+ "relative_l2_max": 0.3121951526683212,
+ "weighted_relative_l2_mean": 0.0038894665587964877,
+ "weighted_relative_frobenius": 0.016348827232659544,
+ "weighted_mean_cosine": 0.999942341395052,
+ "max_relative_norm_change": 0.011715558658246757
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9509326857499247,
+ "mean_cosine": 0.9996812663485738,
+ "p05_cosine": 0.9986380028876795,
+ "relative_l2_mean": 0.012934271859448962,
+ "relative_l2_p95": 0.05237140004079371,
+ "relative_l2_max": 0.31281956387251403,
+ "weighted_relative_l2_mean": 0.002152151021738678,
+ "weighted_relative_frobenius": 0.015697274645901753,
+ "weighted_mean_cosine": 0.9999465589607925,
+ "max_relative_norm_change": 0.01009139470261335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.699578,
+ "warm_p50_ms": 0.108813,
+ "warm_p95_ms": 0.1122069,
+ "warm_p99_ms": 0.12118309999999999,
+ "samples_ms": [
+ 0.111257,
+ 0.112165,
+ 0.111677,
+ 0.108813,
+ 0.108882,
+ 0.126413,
+ 0.109232,
+ 0.108813,
+ 0.107905,
+ 0.110768,
+ 0.109791,
+ 0.109232,
+ 0.10979,
+ 0.108813,
+ 0.108323,
+ 0.113003,
+ 0.109721,
+ 0.107905,
+ 0.109791,
+ 0.108882,
+ 0.107835,
+ 0.108813,
+ 0.108254,
+ 0.107835,
+ 0.108813,
+ 0.108393,
+ 0.108324,
+ 0.108324,
+ 0.109791,
+ 0.108813,
+ 0.108324,
+ 0.108324,
+ 0.108394,
+ 0.109721,
+ 0.108813,
+ 0.108813,
+ 0.106438,
+ 0.108394,
+ 0.109791,
+ 0.108393
+ ],
+ "warm_rows_per_second": 4705320.136380764
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8736482352032853,
+ "character_perplexity": 6.5120104615562395,
+ "next_character_accuracy": 0.5035270664008588,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -0.000348406257527736,
+ "mean_prediction_kl_nats": 7.473230500685238e-05,
+ "prediction_kl_p99_nats": 0.00022106994320018864,
+ "maximum_prediction_kl_nats": 0.16099732540412104,
+ "top1_changed_fraction": 0.0016101824873485662,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6786346435546875,
+ "maximum_per_target_cross_entropy_increase": 0.7059441208839417
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -0.000292799509131025,
+ "mean_prediction_kl_nats": 7.028179700099118e-05,
+ "prediction_kl_p99_nats": 0.00021394409080735188,
+ "maximum_prediction_kl_nats": 0.1600478362486274,
+ "top1_changed_fraction": 0.0007667535654040791,
+ "accuracy_change": 0.00015335071308075765,
+ "maximum_absolute_logit_change": 0.6781201362609863,
+ "maximum_per_target_cross_entropy_increase": 0.7044176459312439
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2373751,
+ "unique_rows": 29760,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 248,
+ "max_block_bytes": 8192,
+ "raw_blocks": 80,
+ "compressed_blocks": 168,
+ "zlib_blocks": 168,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29760
+ },
+ "source_float32_sha256": "b86fa28c9c69888ff34b5f2bbec2e1f5ac90bb3b2115b094e67e2355c0cac2be",
+ "decoded_float32_sha256": "ba3b4a903e6ec9b7a676a4ed226e6f534613730d0e3f36a3486e3382f79b5198",
+ "table_sha256": "4badf2ac68819c3bd931f5d6618717b33d8ed7f9627ac05388755e7d48413d2f",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.23762730904854834,
+ "representatives_sha256": "9159b307f2f34639da3d0241266cb65d89d8e977e6831c6a368ac5d604e21d95",
+ "structurally_merged_rows": 13392,
+ "backend_artifact_bytes": 2375001,
+ "backend_with_key_registry_bytes": 3111108,
+ "backend_with_datrie_key_index_bytes": 2721161,
+ "full_model_serving_artifact_bytes": 3939106,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9001684363688671,
+ "mean_cosine": 0.999592323460612,
+ "p05_cosine": 0.9986353610108466,
+ "relative_l2_mean": 0.01657388790925372,
+ "relative_l2_p95": 0.05242172007517553,
+ "relative_l2_max": 0.4465258112640598,
+ "weighted_relative_l2_mean": 0.004170147652851006,
+ "weighted_relative_frobenius": 0.025142984167262527,
+ "weighted_mean_cosine": 0.9998870304493592,
+ "max_relative_norm_change": 0.011715558658246757
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9000432033719185,
+ "mean_cosine": 0.9995973008579918,
+ "p05_cosine": 0.99861973731563,
+ "relative_l2_mean": 0.013356567722426008,
+ "relative_l2_p95": 0.05272507146664552,
+ "relative_l2_max": 0.4471108671817365,
+ "weighted_relative_l2_mean": 0.00243593633909215,
+ "weighted_relative_frobenius": 0.02472316193697585,
+ "weighted_mean_cosine": 0.999891206520516,
+ "max_relative_norm_change": 0.01009139470261335
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.823127,
+ "warm_p50_ms": 0.108324,
+ "warm_p95_ms": 0.11325474999999999,
+ "warm_p99_ms": 0.12117177999999999,
+ "samples_ms": [
+ 0.111118,
+ 0.109232,
+ 0.113143,
+ 0.108324,
+ 0.109721,
+ 0.108813,
+ 0.108813,
+ 0.108813,
+ 0.107905,
+ 0.108324,
+ 0.108883,
+ 0.109232,
+ 0.108393,
+ 0.108324,
+ 0.108813,
+ 0.107835,
+ 0.108324,
+ 0.109302,
+ 0.111117,
+ 0.108324,
+ 0.115378,
+ 0.106927,
+ 0.106927,
+ 0.106997,
+ 0.107905,
+ 0.106997,
+ 0.108743,
+ 0.106927,
+ 0.108394,
+ 0.111118,
+ 0.107835,
+ 0.124876,
+ 0.107904,
+ 0.108324,
+ 0.107905,
+ 0.108324,
+ 0.108324,
+ 0.108324,
+ 0.108324,
+ 0.108813
+ ],
+ "warm_rows_per_second": 4726561.057568037
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8742551923215474,
+ "character_perplexity": 6.515964172405172,
+ "next_character_accuracy": 0.50375709247048,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0002585235646316773,
+ "mean_prediction_kl_nats": 0.0003787976314879992,
+ "prediction_kl_p99_nats": 0.0019297452549639146,
+ "maximum_prediction_kl_nats": 0.4056340534152349,
+ "top1_changed_fraction": 0.001993559270050606,
+ "accuracy_change": 0.00023002606962119199,
+ "maximum_absolute_logit_change": 1.7124749422073364,
+ "maximum_per_target_cross_entropy_increase": 1.2178491353988647
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0003141640741849018,
+ "mean_prediction_kl_nats": 0.0003752861419167501,
+ "prediction_kl_p99_nats": 0.001977486937159491,
+ "maximum_prediction_kl_nats": 0.406279360229042,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": 0.00038337678270194964,
+ "maximum_absolute_logit_change": 1.7008367776870728,
+ "maximum_per_target_cross_entropy_increase": 1.2212761640548706
+ }
+ }
+ ],
+ "baseline_quality": {
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.873996702168374,
+ "character_perplexity": 6.514280077498528,
+ "next_character_accuracy": 0.5035270664008588,
+ "unknown_targets": 0
+ },
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467822847944068,
+ "character_perplexity": 5.736115762235418,
+ "next_character_accuracy": 0.5234991949704823,
+ "unknown_targets": 0
+ }
+ },
+ "selection": {
+ "fp32": {
+ "feasible": true,
+ "name": "fp32-joint-0.95",
+ "backend_with_key_registry_bytes": 8192404,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7468335463514388,
+ "character_perplexity": 5.736409811997366,
+ "next_character_accuracy": 0.5232691865368397,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 5.126287092804774e-05,
+ "mean_prediction_kl_nats": 4.523398162602008e-05,
+ "prediction_kl_p99_nats": 0.0002614981201286399,
+ "maximum_prediction_kl_nats": 0.031281574788080815,
+ "top1_changed_fraction": 0.0018400674691405351,
+ "accuracy_change": -0.00023000843364251455,
+ "maximum_absolute_logit_change": 0.6498794555664062,
+ "maximum_per_target_cross_entropy_increase": 0.5136287212371826
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00014062225521751923,
+ 0.0002640022597950617
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "fp16": {
+ "feasible": true,
+ "name": "fp16-joint-0.95",
+ "backend_with_key_registry_bytes": 4609709,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.746831640034469,
+ "character_perplexity": 5.7363988765924185,
+ "next_character_accuracy": 0.5232691865368397,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 4.93564580521361e-05,
+ "mean_prediction_kl_nats": 4.522092181157221e-05,
+ "prediction_kl_p99_nats": 0.000260726444150113,
+ "maximum_prediction_kl_nats": 0.03128432153710157,
+ "top1_changed_fraction": 0.0018400674691405351,
+ "accuracy_change": -0.00023000843364251455,
+ "maximum_absolute_logit_change": 0.6495251655578613,
+ "maximum_per_target_cross_entropy_increase": 0.5136415958404541
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00014176603083662905,
+ 0.00026122043476795856
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "int8": {
+ "feasible": true,
+ "name": "int8-joint-0.95",
+ "backend_with_key_registry_bytes": 3113298,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7468335912146888,
+ "character_perplexity": 5.7364100693513596,
+ "next_character_accuracy": 0.5231925170589589,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 5.130773647893945e-05,
+ "mean_prediction_kl_nats": 4.9148167713019224e-05,
+ "prediction_kl_p99_nats": 0.00026751634360011476,
+ "maximum_prediction_kl_nats": 0.030909915141433902,
+ "top1_changed_fraction": 0.0026067622479490913,
+ "accuracy_change": -0.00030667791152338975,
+ "maximum_absolute_logit_change": 0.652705192565918,
+ "maximum_per_target_cross_entropy_increase": 0.5167644023895264
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00014600205117785967,
+ 0.0002798015539844717
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ }
+ },
+ "fixed_quantization_test_controls": {
+ "fp16": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467813805442989,
+ "character_perplexity": 5.736110575354466,
+ "next_character_accuracy": 0.5234991949704823,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -9.042496992371759e-07,
+ "mean_prediction_kl_nats": 3.983294762355042e-09,
+ "prediction_kl_p99_nats": 2.8867819098809675e-08,
+ "maximum_prediction_kl_nats": 7.317560600678412e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001468658447265625,
+ "maximum_per_target_cross_entropy_increase": 0.00078582763671875
+ }
+ },
+ "int8": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467908037670918,
+ "character_perplexity": 5.736164628257058,
+ "next_character_accuracy": 0.5234225254926014,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": 8.519008971568044e-06,
+ "mean_prediction_kl_nats": 3.491065829910206e-06,
+ "prediction_kl_p99_nats": 2.725155572071768e-05,
+ "maximum_prediction_kl_nats": 0.0001554347123477164,
+ "top1_changed_fraction": 0.0009967032124511231,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.03646731376647949,
+ "maximum_per_target_cross_entropy_increase": 0.018882274627685547
+ }
+ }
+ },
+ "checkpoint_sha256": "464242bc5b4367053b82ea7fa20f47536b2311fd0a3f4c4aaf6c0d4ed4a9e0c7",
+ "registry_sha256": "abab407b9c927abc758d3d85a11cec2c2535254fe5af55b6cce10b650a992d4d"
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 11,
+ "rows": 43152,
+ "dim": 64,
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "source_float32_bytes": 11046912,
+ "key_registry_bytes": 736107,
+ "unchanged_backbone_artifact_bytes": 827998,
+ "alternative_datrie_key_index_bytes": 346160,
+ "query_rows": 512,
+ "query_ids_sha256": "43ca78408803e060726877734dd5761b4e85ad9be7de5a6b6d8601f32db85068",
+ "frequency_trace_sha256": "7ba41132ac5296b1ee890e8c97c4f892398df8fb97e4bbb475dab44b5c5af0cc",
+ "variants": [
+ {
+ "name": "fp32-raw",
+ "encoding": "fp32",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 256,
+ "storage": {
+ "artifact_bytes": 11046912,
+ "unique_rows": 43152,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "table_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.0591190280392766,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 11047717,
+ "backend_with_key_registry_bytes": 11783824,
+ "backend_with_datrie_key_index_bytes": 11393877,
+ "full_model_serving_artifact_bytes": 12611822,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.239905,
+ "warm_p50_ms": 0.046934,
+ "warm_p95_ms": 0.05026475,
+ "warm_p99_ms": 0.07390759999999999,
+ "samples_ms": [
+ 0.050146,
+ 0.052521,
+ 0.046933,
+ 0.047283,
+ 0.046864,
+ 0.047283,
+ 0.046863,
+ 0.047771,
+ 0.046863,
+ 0.046934,
+ 0.046864,
+ 0.046374,
+ 0.046933,
+ 0.046934,
+ 0.047353,
+ 0.046933,
+ 0.047352,
+ 0.046444,
+ 0.049168,
+ 0.048679,
+ 0.048261,
+ 0.046864,
+ 0.047283,
+ 0.047282,
+ 0.046374,
+ 0.087581,
+ 0.049657,
+ 0.046305,
+ 0.047352,
+ 0.046375,
+ 0.047283,
+ 0.046863,
+ 0.046863,
+ 0.047771,
+ 0.046305,
+ 0.046864,
+ 0.047353,
+ 0.046863,
+ 0.046793,
+ 0.047352
+ ],
+ "warm_rows_per_second": 10908935.952614311
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746469158554788,
+ "character_perplexity": 6.518517128911684,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-identity",
+ "encoding": "fp32",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 10650963,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1349,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1349,
+ "zlib_blocks": 1349,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "table_sha256": "a545b67631916f8cdacff843e36ed4eeeeafc4c690379bf1e67ffe8dd7e39ce7",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.9225178682245314,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 10652175,
+ "backend_with_key_registry_bytes": 11388282,
+ "backend_with_datrie_key_index_bytes": 10998335,
+ "full_model_serving_artifact_bytes": 12216280,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.246122,
+ "warm_p50_ms": 0.055314,
+ "warm_p95_ms": 0.05684735,
+ "warm_p99_ms": 0.05934325,
+ "samples_ms": [
+ 0.058108,
+ 0.055384,
+ 0.055803,
+ 0.055873,
+ 0.055873,
+ 0.055873,
+ 0.055314,
+ 0.055734,
+ 0.055804,
+ 0.055803,
+ 0.056222,
+ 0.053917,
+ 0.060133,
+ 0.055384,
+ 0.055314,
+ 0.055384,
+ 0.055873,
+ 0.054896,
+ 0.054896,
+ 0.053988,
+ 0.055385,
+ 0.054477,
+ 0.054477,
+ 0.053848,
+ 0.053848,
+ 0.053987,
+ 0.054057,
+ 0.054476,
+ 0.053428,
+ 0.055314,
+ 0.056781,
+ 0.056362,
+ 0.056362,
+ 0.055873,
+ 0.053987,
+ 0.053917,
+ 0.053429,
+ 0.054407,
+ 0.053499,
+ 0.053568
+ ],
+ "warm_rows_per_second": 9256246.158296272
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746469158554788,
+ "character_perplexity": 6.518517128911684,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-layout",
+ "encoding": "fp32",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 10650963,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1349,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1349,
+ "zlib_blocks": 1349,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "table_sha256": "a545b67631916f8cdacff843e36ed4eeeeafc4c690379bf1e67ffe8dd7e39ce7",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 1.0841774109285325,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 10652174,
+ "backend_with_key_registry_bytes": 11388281,
+ "backend_with_datrie_key_index_bytes": 10998334,
+ "full_model_serving_artifact_bytes": 12216279,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.219652,
+ "warm_p50_ms": 0.055245,
+ "warm_p95_ms": 0.05580745,
+ "warm_p99_ms": 0.05698076,
+ "samples_ms": [
+ 0.057689,
+ 0.055803,
+ 0.055314,
+ 0.055245,
+ 0.055315,
+ 0.055385,
+ 0.055314,
+ 0.053359,
+ 0.055385,
+ 0.055803,
+ 0.054825,
+ 0.054895,
+ 0.053987,
+ 0.054476,
+ 0.054895,
+ 0.055384,
+ 0.055873,
+ 0.053988,
+ 0.055454,
+ 0.053499,
+ 0.054826,
+ 0.053918,
+ 0.055384,
+ 0.055245,
+ 0.055315,
+ 0.055804,
+ 0.055733,
+ 0.055314,
+ 0.055384,
+ 0.055314,
+ 0.054826,
+ 0.054336,
+ 0.054476,
+ 0.053429,
+ 0.054895,
+ 0.055454,
+ 0.053987,
+ 0.053987,
+ 0.053917,
+ 0.054336
+ ],
+ "warm_rows_per_second": 9267807.041361209
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746469158554788,
+ "character_perplexity": 6.518517128911684,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-dedup",
+ "encoding": "fp32",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 10650963,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 1349,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 1349,
+ "zlib_blocks": 1349,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "table_sha256": "a545b67631916f8cdacff843e36ed4eeeeafc4c690379bf1e67ffe8dd7e39ce7",
+ "source_bytes_preserved": true,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.9682164608966559,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 10652174,
+ "backend_with_key_registry_bytes": 11388281,
+ "backend_with_datrie_key_index_bytes": 10998334,
+ "full_model_serving_artifact_bytes": 12216279,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 6.187035,
+ "warm_p50_ms": 0.0548255,
+ "warm_p95_ms": 0.0563409,
+ "warm_p99_ms": 0.06302108,
+ "samples_ms": [
+ 0.05727,
+ 0.056292,
+ 0.054406,
+ 0.055803,
+ 0.055384,
+ 0.054965,
+ 0.054896,
+ 0.054896,
+ 0.056292,
+ 0.055384,
+ 0.055873,
+ 0.066698,
+ 0.055873,
+ 0.054896,
+ 0.055385,
+ 0.054477,
+ 0.054826,
+ 0.055803,
+ 0.055384,
+ 0.054895,
+ 0.054476,
+ 0.054406,
+ 0.054476,
+ 0.054406,
+ 0.054406,
+ 0.054895,
+ 0.053847,
+ 0.054825,
+ 0.054406,
+ 0.054406,
+ 0.054336,
+ 0.054895,
+ 0.053498,
+ 0.054406,
+ 0.054406,
+ 0.054825,
+ 0.054825,
+ 0.054825,
+ 0.054337,
+ 0.053987
+ ],
+ "warm_rows_per_second": 9338720.121111527
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746469158554788,
+ "character_perplexity": 6.518517128911684,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp32-joint-0.99",
+ "encoding": "fp32",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7468265,
+ "unique_rows": 29808,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 932,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 932,
+ "zlib_blocks": 932,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29808
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "c65447b4d409ccbfedd7c270e0e49fbbf695ab4a358cadd55e85d13246041d80",
+ "table_sha256": "d79fe9c030582a1ba7f7dfabb5207324fab6d87659bdb3230479ed8523553c9a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6926265370566398,
+ "representatives_sha256": "a838bcae111e98748e3dce6abedefe77b5bc3acf70e32f76d5b98e57f3752c6b",
+ "structurally_merged_rows": 13344,
+ "backend_artifact_bytes": 7469472,
+ "backend_with_key_registry_bytes": 8205579,
+ "backend_with_datrie_key_index_bytes": 7815632,
+ "full_model_serving_artifact_bytes": 9033577,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9902459060645252,
+ "mean_cosine": 0.9997189716982479,
+ "p05_cosine": 0.9986685109158533,
+ "relative_l2_mean": 0.012565436873781678,
+ "relative_l2_p95": 0.05185636339077146,
+ "relative_l2_max": 0.1400040347803519,
+ "weighted_relative_l2_mean": 0.0020033097036198408,
+ "weighted_relative_frobenius": 0.013052669069275194,
+ "weighted_mean_cosine": 0.999959890607113,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9902459060645252,
+ "mean_cosine": 0.9997189716982479,
+ "p05_cosine": 0.9986685109158533,
+ "relative_l2_mean": 0.012565436873781678,
+ "relative_l2_p95": 0.05185636339077146,
+ "relative_l2_max": 0.1400040347803519,
+ "weighted_relative_l2_mean": 0.0020033097036198408,
+ "weighted_relative_frobenius": 0.013052669069275194,
+ "weighted_mean_cosine": 0.999959890607113,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.532552,
+ "warm_p50_ms": 0.054337,
+ "warm_p95_ms": 0.05822685,
+ "warm_p99_ms": 0.06524767999999999,
+ "samples_ms": [
+ 0.059155,
+ 0.056711,
+ 0.055803,
+ 0.058178,
+ 0.055314,
+ 0.056781,
+ 0.056292,
+ 0.054337,
+ 0.053988,
+ 0.069143,
+ 0.055873,
+ 0.053987,
+ 0.054895,
+ 0.054336,
+ 0.054895,
+ 0.056292,
+ 0.053358,
+ 0.053987,
+ 0.053987,
+ 0.054057,
+ 0.053988,
+ 0.053918,
+ 0.053918,
+ 0.054337,
+ 0.053987,
+ 0.054057,
+ 0.053987,
+ 0.053917,
+ 0.053917,
+ 0.053917,
+ 0.053918,
+ 0.053499,
+ 0.053918,
+ 0.054476,
+ 0.054895,
+ 0.054337,
+ 0.054337,
+ 0.054337,
+ 0.054407,
+ 0.054896
+ ],
+ "warm_rows_per_second": 9422676.997257853
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874576445459657,
+ "character_perplexity": 6.518057782614786,
+ "next_character_accuracy": 0.49271584112866124,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -7.046791284201621e-05,
+ "mean_prediction_kl_nats": 9.676395478281967e-06,
+ "prediction_kl_p99_nats": 0.00013294663297309573,
+ "maximum_prediction_kl_nats": 0.0047893837222148065,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.3017038106918335,
+ "maximum_per_target_cross_entropy_increase": 0.1393296718597412
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -7.046791284201621e-05,
+ "mean_prediction_kl_nats": 9.676395478281967e-06,
+ "prediction_kl_p99_nats": 0.00013294663297309573,
+ "maximum_prediction_kl_nats": 0.0047893837222148065,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.3017038106918335,
+ "maximum_per_target_cross_entropy_increase": 0.1393296718597412
+ }
+ },
+ {
+ "name": "fp32-joint-0.98",
+ "encoding": "fp32",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7466244,
+ "unique_rows": 29799,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 932,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 932,
+ "zlib_blocks": 932,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29799
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "b4d54e46e87aeb4711779bcdd2b7d067c8bdf1fa2f4a50188867c57189041436",
+ "table_sha256": "e3868b17869f77825297b2756b988244fbd79ac04d525fb90e8612fa6b5f4944",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6810663971118629,
+ "representatives_sha256": "46bc18844cb26a93245354eb5174585bd3eb9b80d4ba7afbbe5f1ce113b4053f",
+ "structurally_merged_rows": 13353,
+ "backend_artifact_bytes": 7467451,
+ "backend_with_key_registry_bytes": 8203558,
+ "backend_with_datrie_key_index_bytes": 7813611,
+ "full_model_serving_artifact_bytes": 9031556,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9812375062946396,
+ "mean_cosine": 0.9997153374071882,
+ "p05_cosine": 0.9986655857379028,
+ "relative_l2_mean": 0.01260639967491202,
+ "relative_l2_p95": 0.051916390301729676,
+ "relative_l2_max": 0.1932779931725137,
+ "weighted_relative_l2_mean": 0.0020400091222027047,
+ "weighted_relative_frobenius": 0.013916796380774301,
+ "weighted_mean_cosine": 0.9999569248762874,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9812375062946396,
+ "mean_cosine": 0.9997153374071882,
+ "p05_cosine": 0.9986655857379028,
+ "relative_l2_mean": 0.01260639967491202,
+ "relative_l2_p95": 0.051916390301729676,
+ "relative_l2_max": 0.1932779931725137,
+ "weighted_relative_l2_mean": 0.0020400091222027047,
+ "weighted_relative_frobenius": 0.013916796380774301,
+ "weighted_mean_cosine": 0.9999569248762874,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.538349,
+ "warm_p50_ms": 0.055385,
+ "warm_p95_ms": 0.05769245,
+ "warm_p99_ms": 0.0580142,
+ "samples_ms": [
+ 0.058178,
+ 0.057689,
+ 0.057758,
+ 0.05727,
+ 0.055803,
+ 0.056293,
+ 0.055873,
+ 0.054965,
+ 0.054895,
+ 0.055873,
+ 0.055384,
+ 0.055385,
+ 0.055454,
+ 0.055384,
+ 0.055384,
+ 0.054895,
+ 0.056292,
+ 0.055384,
+ 0.055454,
+ 0.055873,
+ 0.055803,
+ 0.055384,
+ 0.054476,
+ 0.054476,
+ 0.054895,
+ 0.055873,
+ 0.055384,
+ 0.055384,
+ 0.054476,
+ 0.053917,
+ 0.055734,
+ 0.055385,
+ 0.055803,
+ 0.055803,
+ 0.053429,
+ 0.056292,
+ 0.054825,
+ 0.055314,
+ 0.054895,
+ 0.055734
+ ],
+ "warm_rows_per_second": 9244380.247359393
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874600747195864,
+ "character_perplexity": 6.518216184660315,
+ "next_character_accuracy": 0.49271584112866124,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.6167593858877154e-05,
+ "mean_prediction_kl_nats": 1.0040152880568824e-05,
+ "prediction_kl_p99_nats": 0.00013764166758583975,
+ "maximum_prediction_kl_nats": 0.0047893837222148065,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.4481656551361084,
+ "maximum_per_target_cross_entropy_increase": 0.25843191146850586
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.6167593858877154e-05,
+ "mean_prediction_kl_nats": 1.0040152880568824e-05,
+ "prediction_kl_p99_nats": 0.00013764166758583975,
+ "maximum_prediction_kl_nats": 0.0047893837222148065,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.4481656551361084,
+ "maximum_per_target_cross_entropy_increase": 0.25843191146850586
+ }
+ },
+ {
+ "name": "fp32-joint-0.97",
+ "encoding": "fp32",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7463655,
+ "unique_rows": 29788,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 931,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 931,
+ "zlib_blocks": 931,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29788
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "5776d13fa5d3348bc52ab527c768fac0cac741869b88dda7573cf47b782d1cd1",
+ "table_sha256": "e20eabc67f867ce0e0e2bde225055766d5b283149197712e6b8a4acc85e7a90e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6838157009333372,
+ "representatives_sha256": "8c586ce381d5e20d375f5b61a8eadc7698d7d40f09ad5e057d56c55e8c045ce5",
+ "structurally_merged_rows": 13364,
+ "backend_artifact_bytes": 7464862,
+ "backend_with_key_registry_bytes": 8200969,
+ "backend_with_datrie_key_index_bytes": 7811022,
+ "full_model_serving_artifact_bytes": 9028967,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9703576562286211,
+ "mean_cosine": 0.9997070682357191,
+ "p05_cosine": 0.9986618868574548,
+ "relative_l2_mean": 0.01267762616506722,
+ "relative_l2_p95": 0.05197669763443726,
+ "relative_l2_max": 0.2449152666071064,
+ "weighted_relative_l2_mean": 0.002082663876929235,
+ "weighted_relative_frobenius": 0.01511647551684654,
+ "weighted_mean_cosine": 0.9999520616546923,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9703576562286211,
+ "mean_cosine": 0.9997070682357191,
+ "p05_cosine": 0.9986618868574548,
+ "relative_l2_mean": 0.01267762616506722,
+ "relative_l2_p95": 0.05197669763443726,
+ "relative_l2_max": 0.2449152666071064,
+ "weighted_relative_l2_mean": 0.002082663876929235,
+ "weighted_relative_frobenius": 0.01511647551684654,
+ "weighted_mean_cosine": 0.9999520616546923,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.497213,
+ "warm_p50_ms": 0.054965,
+ "warm_p95_ms": 0.056292,
+ "warm_p99_ms": 0.05744185,
+ "samples_ms": [
+ 0.058177,
+ 0.055873,
+ 0.055384,
+ 0.055803,
+ 0.055315,
+ 0.055315,
+ 0.055804,
+ 0.056222,
+ 0.055314,
+ 0.056292,
+ 0.055315,
+ 0.055734,
+ 0.054825,
+ 0.055314,
+ 0.054755,
+ 0.053917,
+ 0.055314,
+ 0.054825,
+ 0.053009,
+ 0.054406,
+ 0.054476,
+ 0.054965,
+ 0.053917,
+ 0.054895,
+ 0.056292,
+ 0.054825,
+ 0.054476,
+ 0.053359,
+ 0.054825,
+ 0.053498,
+ 0.054057,
+ 0.055384,
+ 0.055873,
+ 0.054895,
+ 0.054965,
+ 0.055384,
+ 0.054965,
+ 0.054895,
+ 0.054407,
+ 0.054826
+ ],
+ "warm_rows_per_second": 9315018.648230692
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8745769143085478,
+ "character_perplexity": 6.518060838599665,
+ "next_character_accuracy": 0.49271584112866124,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -6.999909687999551e-05,
+ "mean_prediction_kl_nats": 2.5804656921662418e-05,
+ "prediction_kl_p99_nats": 0.00016574862158967506,
+ "maximum_prediction_kl_nats": 0.02597693295804385,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.45133912563323975,
+ "maximum_per_target_cross_entropy_increase": 0.47270774841308594
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -6.999909687999551e-05,
+ "mean_prediction_kl_nats": 2.5804656921662418e-05,
+ "prediction_kl_p99_nats": 0.00016574862158967506,
+ "maximum_prediction_kl_nats": 0.02597693295804385,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.45133912563323975,
+ "maximum_per_target_cross_entropy_increase": 0.47270774841308594
+ }
+ },
+ {
+ "name": "fp32-joint-0.95",
+ "encoding": "fp32",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7456459,
+ "unique_rows": 29759,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 930,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 930,
+ "zlib_blocks": 930,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29759
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "9cb75af0bd848f19ad7daafbb360313a37dbf95c8fb63ae038e7987ab7253c9c",
+ "table_sha256": "eda8529fa1e43701cd5e9eaed0248811c40c128008ffa16a21a0b1f1d098f885",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6939489128999412,
+ "representatives_sha256": "a9a02de62e856fddf6769221690688799c98284c2059352332170a432fb31fce",
+ "structurally_merged_rows": 13393,
+ "backend_artifact_bytes": 7457666,
+ "backend_with_key_registry_bytes": 8193773,
+ "backend_with_datrie_key_index_bytes": 7803826,
+ "full_model_serving_artifact_bytes": 9021771,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9500023397388492,
+ "mean_cosine": 0.9996656139836287,
+ "p05_cosine": 0.9986486740297048,
+ "relative_l2_mean": 0.012960582104716335,
+ "relative_l2_p95": 0.052241887765272006,
+ "relative_l2_max": 0.31645540845328646,
+ "weighted_relative_l2_mean": 0.002131533865437752,
+ "weighted_relative_frobenius": 0.016184835758130237,
+ "weighted_mean_cosine": 0.9999450578328678,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9500023397388492,
+ "mean_cosine": 0.9996656139836287,
+ "p05_cosine": 0.9986486740297048,
+ "relative_l2_mean": 0.012960582104716335,
+ "relative_l2_p95": 0.052241887765272006,
+ "relative_l2_max": 0.31645540845328646,
+ "weighted_relative_l2_mean": 0.002131533865437752,
+ "weighted_relative_frobenius": 0.016184835758130237,
+ "weighted_mean_cosine": 0.9999450578328678,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.667905,
+ "warm_p50_ms": 0.055384,
+ "warm_p95_ms": 0.056407399999999996,
+ "warm_p99_ms": 0.05752559,
+ "samples_ms": [
+ 0.057689,
+ 0.056292,
+ 0.055384,
+ 0.055733,
+ 0.055314,
+ 0.055314,
+ 0.055733,
+ 0.056223,
+ 0.056292,
+ 0.055384,
+ 0.056292,
+ 0.055873,
+ 0.055385,
+ 0.055873,
+ 0.05727,
+ 0.056292,
+ 0.055873,
+ 0.056362,
+ 0.054825,
+ 0.055314,
+ 0.055314,
+ 0.055803,
+ 0.055314,
+ 0.055314,
+ 0.054896,
+ 0.055315,
+ 0.055384,
+ 0.055873,
+ 0.055733,
+ 0.054825,
+ 0.055314,
+ 0.055734,
+ 0.055804,
+ 0.055314,
+ 0.055314,
+ 0.055244,
+ 0.055314,
+ 0.055314,
+ 0.055245,
+ 0.055315
+ ],
+ "warm_rows_per_second": 9244547.161635129
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746374539850317,
+ "character_perplexity": 6.518455451838894,
+ "next_character_accuracy": 0.49279251648520167,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -9.46182568385008e-06,
+ "mean_prediction_kl_nats": 4.180147943176077e-05,
+ "prediction_kl_p99_nats": 0.00018860474570523585,
+ "maximum_prediction_kl_nats": 0.03543225249138276,
+ "top1_changed_fraction": 0.0015335071308081583,
+ "accuracy_change": -0.00015335071308081316,
+ "maximum_absolute_logit_change": 1.071291446685791,
+ "maximum_per_target_cross_entropy_increase": 0.9157590866088867
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -9.46182568385008e-06,
+ "mean_prediction_kl_nats": 4.180147943176077e-05,
+ "prediction_kl_p99_nats": 0.00018860474570523585,
+ "maximum_prediction_kl_nats": 0.03543225249138276,
+ "top1_changed_fraction": 0.0015335071308081583,
+ "accuracy_change": -0.00015335071308081316,
+ "maximum_absolute_logit_change": 1.071291446685791,
+ "maximum_per_target_cross_entropy_increase": 0.9157590866088867
+ }
+ },
+ {
+ "name": "fp32-joint-0.9",
+ "encoding": "fp32",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp32",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 256,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 11046912,
+ "artifact_bytes": 7448881,
+ "unique_rows": 29727,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 256,
+ "block_rows": 32,
+ "blocks": 929,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 929,
+ "zlib_blocks": 929,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29727
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "a35551fb6a4063ba410cb6107ef0242d15045847d7dae499f148ea97ee1f44e8",
+ "table_sha256": "db10121b72a08f8330d37d84a6fcb79373ebd4e2aead34b39941ecc4e8a6821d",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp32",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6733195239212364,
+ "representatives_sha256": "e7cd9c89499c63c4f9f30e50a1d393ff3d1fe880471b0e1f4594656192965952",
+ "structurally_merged_rows": 13425,
+ "backend_artifact_bytes": 7450088,
+ "backend_with_key_registry_bytes": 8186195,
+ "backend_with_datrie_key_index_bytes": 7796248,
+ "full_model_serving_artifact_bytes": 9014193,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9001956866909555,
+ "mean_cosine": 0.999587579106024,
+ "p05_cosine": 0.9986360287973155,
+ "relative_l2_mean": 0.013349663575628954,
+ "relative_l2_p95": 0.0524339650002821,
+ "relative_l2_max": 0.44829107892167214,
+ "weighted_relative_l2_mean": 0.0024165394792690195,
+ "weighted_relative_frobenius": 0.029884694087567335,
+ "weighted_mean_cosine": 0.9998840290166183,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9001956866909555,
+ "mean_cosine": 0.999587579106024,
+ "p05_cosine": 0.9986360287973155,
+ "relative_l2_mean": 0.013349663575628954,
+ "relative_l2_p95": 0.0524339650002821,
+ "relative_l2_max": 0.44829107892167214,
+ "weighted_relative_l2_mean": 0.0024165394792690195,
+ "weighted_relative_frobenius": 0.029884694087567335,
+ "weighted_mean_cosine": 0.9998840290166183,
+ "max_relative_norm_change": 0.010087672046877508
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.82407,
+ "warm_p50_ms": 0.05514,
+ "warm_p95_ms": 0.057737899999999995,
+ "warm_p99_ms": 0.06701729,
+ "samples_ms": [
+ 0.058667,
+ 0.072356,
+ 0.057689,
+ 0.055803,
+ 0.056781,
+ 0.055873,
+ 0.054896,
+ 0.055873,
+ 0.054406,
+ 0.055384,
+ 0.054965,
+ 0.054477,
+ 0.055454,
+ 0.054966,
+ 0.054477,
+ 0.053918,
+ 0.055315,
+ 0.055315,
+ 0.055314,
+ 0.055314,
+ 0.054337,
+ 0.056222,
+ 0.054336,
+ 0.054406,
+ 0.055314,
+ 0.055384,
+ 0.053987,
+ 0.055803,
+ 0.056781,
+ 0.053918,
+ 0.054896,
+ 0.055385,
+ 0.054407,
+ 0.054476,
+ 0.054406,
+ 0.053498,
+ 0.054965,
+ 0.055384,
+ 0.054965,
+ 0.054965
+ ],
+ "warm_rows_per_second": 9285455.204932898
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874662206129923,
+ "character_perplexity": 6.518616799589548,
+ "next_character_accuracy": 0.49263916577212086,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.529039134093668e-05,
+ "mean_prediction_kl_nats": 9.758053428754134e-05,
+ "prediction_kl_p99_nats": 0.00036003335280749833,
+ "maximum_prediction_kl_nats": 0.12535552186813684,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -0.0003067014261616263,
+ "maximum_absolute_logit_change": 1.5011210441589355,
+ "maximum_per_target_cross_entropy_increase": 0.9157590866088867
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 1.529039134093668e-05,
+ "mean_prediction_kl_nats": 9.758053428754134e-05,
+ "prediction_kl_p99_nats": 0.00036003335280749833,
+ "maximum_prediction_kl_nats": 0.12535552186813684,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -0.0003067014261616263,
+ "maximum_absolute_logit_change": 1.5011210441589355,
+ "maximum_per_target_cross_entropy_increase": 0.9157590866088867
+ }
+ },
+ {
+ "name": "fp16-raw",
+ "encoding": "fp16",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 128,
+ "storage": {
+ "artifact_bytes": 5523456,
+ "unique_rows": 43152,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "bdea909d0738ce1a9112cfdf55e058286ffdf48616c95623b85c5865be33ab4e",
+ "table_sha256": "0676720f8d94a19e6237b1d04171a8f31c1f922cb1452938976a65b4f39f5ef7",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.06894284300506115,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5524261,
+ "backend_with_key_registry_bytes": 6260368,
+ "backend_with_datrie_key_index_bytes": 5870421,
+ "full_model_serving_artifact_bytes": 7088366,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999513463311,
+ "mean_cosine": 0.9999999792195698,
+ "p05_cosine": 0.9999999716268251,
+ "relative_l2_mean": 0.00020639671283615942,
+ "relative_l2_p95": 0.00024186731774094268,
+ "relative_l2_max": 0.00031522546967662694,
+ "weighted_relative_l2_mean": 7.847420784280686e-05,
+ "weighted_relative_frobenius": 0.00020874166338934465,
+ "weighted_mean_cosine": 0.9999999921182615,
+ "max_relative_norm_change": 0.0001740708570888487
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.313657,
+ "warm_p50_ms": 0.139124,
+ "warm_p95_ms": 0.14103795,
+ "warm_p99_ms": 0.16517111999999998,
+ "samples_ms": [
+ 0.141569,
+ 0.139612,
+ 0.139054,
+ 0.138635,
+ 0.138146,
+ 0.138565,
+ 0.139054,
+ 0.138565,
+ 0.138566,
+ 0.139124,
+ 0.139613,
+ 0.140032,
+ 0.180261,
+ 0.14101,
+ 0.139612,
+ 0.138216,
+ 0.140032,
+ 0.138147,
+ 0.139124,
+ 0.138565,
+ 0.138565,
+ 0.138566,
+ 0.140102,
+ 0.139124,
+ 0.139543,
+ 0.139124,
+ 0.140102,
+ 0.138146,
+ 0.139124,
+ 0.138635,
+ 0.139124,
+ 0.139613,
+ 0.139543,
+ 0.139543,
+ 0.139543,
+ 0.138705,
+ 0.140102,
+ 0.139613,
+ 0.138147,
+ 0.139543
+ ],
+ "warm_rows_per_second": 3680170.207872114
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746435529593024,
+ "character_perplexity": 6.518495207852214,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.362890522061065e-06,
+ "mean_prediction_kl_nats": 4.166460967471791e-09,
+ "prediction_kl_p99_nats": 2.713723610475145e-08,
+ "maximum_prediction_kl_nats": 9.861947594446357e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0012116432189941406,
+ "maximum_per_target_cross_entropy_increase": 0.00064849853515625
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-identity",
+ "encoding": "fp16",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 5451718,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 675,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 675,
+ "zlib_blocks": 675,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "bdea909d0738ce1a9112cfdf55e058286ffdf48616c95623b85c5865be33ab4e",
+ "table_sha256": "35e05ce93fa67f0c19d85448fa6f58f92850761c9ce2bb07cc778719d17ded8e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5560433219652623,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5452926,
+ "backend_with_key_registry_bytes": 6189033,
+ "backend_with_datrie_key_index_bytes": 5799086,
+ "full_model_serving_artifact_bytes": 7017031,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999513463311,
+ "mean_cosine": 0.9999999792195698,
+ "p05_cosine": 0.9999999716268251,
+ "relative_l2_mean": 0.00020639671283615942,
+ "relative_l2_p95": 0.00024186731774094268,
+ "relative_l2_max": 0.00031522546967662694,
+ "weighted_relative_l2_mean": 7.847420784280686e-05,
+ "weighted_relative_frobenius": 0.00020874166338934465,
+ "weighted_mean_cosine": 0.9999999921182615,
+ "max_relative_norm_change": 0.0001740708570888487
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.896426,
+ "warm_p50_ms": 0.1469115,
+ "warm_p95_ms": 0.15310974999999993,
+ "warm_p99_ms": 0.18639561999999998,
+ "samples_ms": [
+ 0.193042,
+ 0.151905,
+ 0.147574,
+ 0.147156,
+ 0.147645,
+ 0.147156,
+ 0.148972,
+ 0.146597,
+ 0.147086,
+ 0.147645,
+ 0.148064,
+ 0.146178,
+ 0.147645,
+ 0.14569,
+ 0.147156,
+ 0.146178,
+ 0.146178,
+ 0.147994,
+ 0.146178,
+ 0.146178,
+ 0.146667,
+ 0.14527,
+ 0.146178,
+ 0.146597,
+ 0.146737,
+ 0.146666,
+ 0.176,
+ 0.14953,
+ 0.146248,
+ 0.146737,
+ 0.147155,
+ 0.148134,
+ 0.146667,
+ 0.147155,
+ 0.146178,
+ 0.147156,
+ 0.14527,
+ 0.146666,
+ 0.147156,
+ 0.146736
+ ],
+ "warm_rows_per_second": 3485091.3645289848
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746435529593024,
+ "character_perplexity": 6.518495207852214,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.362890522061065e-06,
+ "mean_prediction_kl_nats": 4.166460967471791e-09,
+ "prediction_kl_p99_nats": 2.713723610475145e-08,
+ "maximum_prediction_kl_nats": 9.861947594446357e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0012116432189941406,
+ "maximum_per_target_cross_entropy_increase": 0.00064849853515625
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-layout",
+ "encoding": "fp16",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 5451718,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 675,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 675,
+ "zlib_blocks": 675,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "bdea909d0738ce1a9112cfdf55e058286ffdf48616c95623b85c5865be33ab4e",
+ "table_sha256": "35e05ce93fa67f0c19d85448fa6f58f92850761c9ce2bb07cc778719d17ded8e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.6727040119003505,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5452925,
+ "backend_with_key_registry_bytes": 6189032,
+ "backend_with_datrie_key_index_bytes": 5799085,
+ "full_model_serving_artifact_bytes": 7017030,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999513463311,
+ "mean_cosine": 0.9999999792195698,
+ "p05_cosine": 0.9999999716268251,
+ "relative_l2_mean": 0.00020639671283615942,
+ "relative_l2_p95": 0.00024186731774094268,
+ "relative_l2_max": 0.00031522546967662694,
+ "weighted_relative_l2_mean": 7.847420784280686e-05,
+ "weighted_relative_frobenius": 0.00020874166338934465,
+ "weighted_mean_cosine": 0.9999999921182615,
+ "max_relative_norm_change": 0.0001740708570888487
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.984566,
+ "warm_p50_ms": 0.1474,
+ "warm_p95_ms": 0.1676687,
+ "warm_p99_ms": 0.17912646,
+ "samples_ms": [
+ 0.149531,
+ 0.148134,
+ 0.148133,
+ 0.147086,
+ 0.148064,
+ 0.147645,
+ 0.146248,
+ 0.146737,
+ 0.147155,
+ 0.147575,
+ 0.145201,
+ 0.169924,
+ 0.148553,
+ 0.16755,
+ 0.147156,
+ 0.147644,
+ 0.146178,
+ 0.146178,
+ 0.146178,
+ 0.147155,
+ 0.147156,
+ 0.147016,
+ 0.146247,
+ 0.146178,
+ 0.147644,
+ 0.146248,
+ 0.145758,
+ 0.147226,
+ 0.147574,
+ 0.146667,
+ 0.145759,
+ 0.147645,
+ 0.147575,
+ 0.147575,
+ 0.146667,
+ 0.147645,
+ 0.148552,
+ 0.16706,
+ 0.18501,
+ 0.155188
+ ],
+ "warm_rows_per_second": 3473541.383989145
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746435529593024,
+ "character_perplexity": 6.518495207852214,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.362890522061065e-06,
+ "mean_prediction_kl_nats": 4.166460967471791e-09,
+ "prediction_kl_p99_nats": 2.713723610475145e-08,
+ "maximum_prediction_kl_nats": 9.861947594446357e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0012116432189941406,
+ "maximum_per_target_cross_entropy_increase": 0.00064849853515625
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-dedup",
+ "encoding": "fp16",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 5451718,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 675,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 675,
+ "zlib_blocks": 675,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "bdea909d0738ce1a9112cfdf55e058286ffdf48616c95623b85c5865be33ab4e",
+ "table_sha256": "35e05ce93fa67f0c19d85448fa6f58f92850761c9ce2bb07cc778719d17ded8e",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5485050650313497,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 5452925,
+ "backend_with_key_registry_bytes": 6189032,
+ "backend_with_datrie_key_index_bytes": 5799085,
+ "full_model_serving_artifact_bytes": 7017030,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999513463311,
+ "mean_cosine": 0.9999999792195698,
+ "p05_cosine": 0.9999999716268251,
+ "relative_l2_mean": 0.00020639671283615942,
+ "relative_l2_p95": 0.00024186731774094268,
+ "relative_l2_max": 0.00031522546967662694,
+ "weighted_relative_l2_mean": 7.847420784280686e-05,
+ "weighted_relative_frobenius": 0.00020874166338934465,
+ "weighted_mean_cosine": 0.9999999921182615,
+ "max_relative_norm_change": 0.0001740708570888487
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.941404,
+ "warm_p50_ms": 0.147156,
+ "warm_p95_ms": 0.15012095,
+ "warm_p99_ms": 0.15759589000000002,
+ "samples_ms": [
+ 0.154699,
+ 0.14988,
+ 0.148972,
+ 0.147086,
+ 0.148553,
+ 0.147644,
+ 0.147575,
+ 0.148553,
+ 0.148483,
+ 0.146247,
+ 0.146737,
+ 0.146248,
+ 0.147156,
+ 0.146666,
+ 0.147645,
+ 0.145689,
+ 0.146667,
+ 0.148063,
+ 0.146597,
+ 0.146597,
+ 0.147155,
+ 0.147645,
+ 0.146248,
+ 0.148064,
+ 0.147156,
+ 0.147574,
+ 0.159448,
+ 0.149042,
+ 0.147575,
+ 0.147155,
+ 0.147645,
+ 0.148553,
+ 0.146177,
+ 0.146667,
+ 0.146178,
+ 0.146527,
+ 0.147156,
+ 0.146247,
+ 0.147156,
+ 0.148064
+ ],
+ "warm_rows_per_second": 3479300.877979831
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746435529593024,
+ "character_perplexity": 6.518495207852214,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -3.362890522061065e-06,
+ "mean_prediction_kl_nats": 4.166460967471791e-09,
+ "prediction_kl_p99_nats": 2.713723610475145e-08,
+ "maximum_prediction_kl_nats": 9.861947594446357e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0012116432189941406,
+ "maximum_per_target_cross_entropy_increase": 0.00064849853515625
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "fp16-joint-0.99",
+ "encoding": "fp16",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3878921,
+ "unique_rows": 29810,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29810
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "125ad803adef474db5d65aac757c7d4d595c2a90dcc5604c8c3e61f7accd1a1e",
+ "table_sha256": "332c03fd3b8aadc1f5dc436b256bf6089fed708c3e9e7889b7151dade6e19075",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3961187079548836,
+ "representatives_sha256": "d4955c317d6d856716a740f4120db129490651f55043fd12797b6d4edf8a4b52",
+ "structurally_merged_rows": 13342,
+ "backend_artifact_bytes": 3880127,
+ "backend_with_key_registry_bytes": 4616234,
+ "backend_with_datrie_key_index_bytes": 4226287,
+ "full_model_serving_artifact_bytes": 5444232,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9902543419830719,
+ "mean_cosine": 0.9997190776427124,
+ "p05_cosine": 0.9986690949302727,
+ "relative_l2_mean": 0.01270466779811439,
+ "relative_l2_p95": 0.0518538727095818,
+ "relative_l2_max": 0.13994770767903386,
+ "weighted_relative_l2_mean": 0.002069047250165433,
+ "weighted_relative_frobenius": 0.013052639637527957,
+ "weighted_mean_cosine": 0.9999599095257184,
+ "max_relative_norm_change": 0.010116184706844977
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9902580478946749,
+ "mean_cosine": 0.9997190850797039,
+ "p05_cosine": 0.998668959958706,
+ "relative_l2_mean": 0.012562250183851622,
+ "relative_l2_p95": 0.051859147381795756,
+ "relative_l2_max": 0.13992493091393668,
+ "weighted_relative_l2_mean": 0.0020023304476201316,
+ "weighted_relative_frobenius": 0.01305117528543782,
+ "weighted_mean_cosine": 0.9999599160050102,
+ "max_relative_norm_change": 0.010087895669284265
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 4.907193,
+ "warm_p50_ms": 0.147156,
+ "warm_p95_ms": 0.15095829999999996,
+ "warm_p99_ms": 0.16749055999999998,
+ "samples_ms": [
+ 0.150438,
+ 0.148063,
+ 0.147575,
+ 0.147505,
+ 0.160844,
+ 0.149041,
+ 0.148552,
+ 0.148552,
+ 0.147644,
+ 0.148134,
+ 0.147575,
+ 0.146178,
+ 0.146667,
+ 0.14527,
+ 0.147156,
+ 0.147156,
+ 0.146597,
+ 0.146597,
+ 0.145689,
+ 0.147575,
+ 0.146667,
+ 0.146667,
+ 0.146248,
+ 0.146736,
+ 0.148064,
+ 0.147645,
+ 0.147085,
+ 0.145619,
+ 0.1452,
+ 0.146597,
+ 0.17174,
+ 0.149949,
+ 0.147156,
+ 0.147086,
+ 0.147226,
+ 0.147156,
+ 0.147226,
+ 0.146736,
+ 0.145759,
+ 0.147155
+ ],
+ "warm_rows_per_second": 3479300.877979831
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874574246673585,
+ "character_perplexity": 6.518043450815874,
+ "next_character_accuracy": 0.49271584112866124,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -7.266654155279356e-05,
+ "mean_prediction_kl_nats": 9.671910930345072e-06,
+ "prediction_kl_p99_nats": 0.00013471611525336594,
+ "maximum_prediction_kl_nats": 0.004780077474893736,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.30183470249176025,
+ "maximum_per_target_cross_entropy_increase": 0.1393224000930786
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -6.930388409210745e-05,
+ "mean_prediction_kl_nats": 9.668194206757754e-06,
+ "prediction_kl_p99_nats": 0.00013446976339783222,
+ "maximum_prediction_kl_nats": 0.004787369487992659,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.3018685579299927,
+ "maximum_per_target_cross_entropy_increase": 0.13929986953735352
+ }
+ },
+ {
+ "name": "fp16-joint-0.98",
+ "encoding": "fp16",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3877703,
+ "unique_rows": 29801,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29801
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "38f219bbaa111886abe433803c907c9f4c011988547dfa7b03c85fa6107fba5b",
+ "table_sha256": "110e332f9528f9210380f65ff9a010e579809aa4556b0b844b86a4179311e3eb",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.5946081799920648,
+ "representatives_sha256": "d56552783e04aae85d9de6ff8093b0136dd12a38b578852754d4949075a9a3d3",
+ "structurally_merged_rows": 13351,
+ "backend_artifact_bytes": 3878909,
+ "backend_with_key_registry_bytes": 4615016,
+ "backend_with_datrie_key_index_bytes": 4225069,
+ "full_model_serving_artifact_bytes": 5443014,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9812282968421467,
+ "mean_cosine": 0.9997154431587459,
+ "p05_cosine": 0.998665514808482,
+ "relative_l2_mean": 0.012745591071179944,
+ "relative_l2_p95": 0.05191842837465708,
+ "relative_l2_max": 0.19332654279600592,
+ "weighted_relative_l2_mean": 0.002105698474020462,
+ "weighted_relative_frobenius": 0.013916586005065044,
+ "weighted_mean_cosine": 0.9999569443360461,
+ "max_relative_norm_change": 0.010116184706844977
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9812273533691351,
+ "mean_cosine": 0.9997154506283183,
+ "p05_cosine": 0.9986658133165788,
+ "relative_l2_mean": 0.012603213295592029,
+ "relative_l2_p95": 0.051907282774863284,
+ "relative_l2_max": 0.1933325230358195,
+ "weighted_relative_l2_mean": 0.0020390277848896113,
+ "weighted_relative_frobenius": 0.01391532039964873,
+ "weighted_mean_cosine": 0.999956950429156,
+ "max_relative_norm_change": 0.010087895669284265
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.083193,
+ "warm_p50_ms": 0.147086,
+ "warm_p95_ms": 0.14997695,
+ "warm_p99_ms": 0.16716588,
+ "samples_ms": [
+ 0.150508,
+ 0.149461,
+ 0.149042,
+ 0.147575,
+ 0.148552,
+ 0.148064,
+ 0.147156,
+ 0.146178,
+ 0.147645,
+ 0.147086,
+ 0.146109,
+ 0.1452,
+ 0.145689,
+ 0.147505,
+ 0.147086,
+ 0.145689,
+ 0.1452,
+ 0.145689,
+ 0.177816,
+ 0.149949,
+ 0.147644,
+ 0.147086,
+ 0.147645,
+ 0.147086,
+ 0.147086,
+ 0.147085,
+ 0.147156,
+ 0.146597,
+ 0.147156,
+ 0.145759,
+ 0.147644,
+ 0.146248,
+ 0.145759,
+ 0.14534,
+ 0.14534,
+ 0.14527,
+ 0.145759,
+ 0.1452,
+ 0.146597,
+ 0.146667
+ ],
+ "warm_rows_per_second": 3480956.719198292
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8745985378953693,
+ "character_perplexity": 6.5182017839779824,
+ "next_character_accuracy": 0.49271584112866124,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -4.837678991483754e-05,
+ "mean_prediction_kl_nats": 1.0035555638620316e-05,
+ "prediction_kl_p99_nats": 0.0001374315140857644,
+ "maximum_prediction_kl_nats": 0.004780077474893736,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.4480018615722656,
+ "maximum_per_target_cross_entropy_increase": 0.25829124450683594
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.501405077010117e-05,
+ "mean_prediction_kl_nats": 1.0031980910063025e-05,
+ "prediction_kl_p99_nats": 0.00013765976797404305,
+ "maximum_prediction_kl_nats": 0.004787369487992659,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.44802331924438477,
+ "maximum_per_target_cross_entropy_increase": 0.2583656311035156
+ }
+ },
+ {
+ "name": "fp16-joint-0.97",
+ "encoding": "fp16",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3876367,
+ "unique_rows": 29790,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 466,
+ "zlib_blocks": 466,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29790
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "c9bceb8ddbb8626afafcac0cbfaa184b5b0ab2ac3dd6a6a9f17803dfd44b23db",
+ "table_sha256": "3795c24d94caca2a7c11c065db26125df7043fda442d1c4d6b155ddcf96b7f82",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.47510257293470204,
+ "representatives_sha256": "0442ab13267b0a1637ef197a94cc19d4fad9d4368204426ed4180dd6e9c0dcf4",
+ "structurally_merged_rows": 13362,
+ "backend_artifact_bytes": 3877573,
+ "backend_with_key_registry_bytes": 4613680,
+ "backend_with_datrie_key_index_bytes": 4223733,
+ "full_model_serving_artifact_bytes": 5441678,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9703580930035629,
+ "mean_cosine": 0.9997071747644565,
+ "p05_cosine": 0.9986617977957264,
+ "relative_l2_mean": 0.01281676131336959,
+ "relative_l2_p95": 0.05197299549482736,
+ "relative_l2_max": 0.24491417991277825,
+ "weighted_relative_l2_mean": 0.002148307878191639,
+ "weighted_relative_frobenius": 0.015115978536123264,
+ "weighted_mean_cosine": 0.9999520822397424,
+ "max_relative_norm_change": 0.010116184706844977
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9703560166990202,
+ "mean_cosine": 0.9997071825853483,
+ "p05_cosine": 0.998661316509401,
+ "relative_l2_mean": 0.012674435105461485,
+ "relative_l2_p95": 0.05196615509578719,
+ "relative_l2_max": 0.2449187343259896,
+ "weighted_relative_l2_mean": 0.0020816757058458495,
+ "weighted_relative_frobenius": 0.01511475371180603,
+ "weighted_mean_cosine": 0.9999520885920036,
+ "max_relative_norm_change": 0.010087895669284265
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.373174,
+ "warm_p50_ms": 0.147156,
+ "warm_p95_ms": 0.17379024999999998,
+ "warm_p99_ms": 0.17988394,
+ "samples_ms": [
+ 0.149042,
+ 0.148064,
+ 0.147645,
+ 0.146177,
+ 0.146248,
+ 0.182635,
+ 0.173696,
+ 0.14995,
+ 0.147575,
+ 0.148063,
+ 0.146597,
+ 0.147156,
+ 0.147645,
+ 0.146737,
+ 0.148552,
+ 0.146178,
+ 0.146247,
+ 0.146667,
+ 0.146178,
+ 0.147086,
+ 0.147086,
+ 0.148552,
+ 0.147575,
+ 0.147645,
+ 0.147645,
+ 0.147156,
+ 0.147645,
+ 0.147155,
+ 0.146737,
+ 0.145828,
+ 0.147645,
+ 0.175581,
+ 0.151835,
+ 0.147156,
+ 0.147156,
+ 0.146597,
+ 0.147086,
+ 0.146736,
+ 0.147226,
+ 0.146736
+ ],
+ "warm_rows_per_second": 3479300.877979831
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874574723909101,
+ "character_perplexity": 6.518046561458446,
+ "next_character_accuracy": 0.49271584112866124,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -7.218934060193374e-05,
+ "mean_prediction_kl_nats": 2.579058573039068e-05,
+ "prediction_kl_p99_nats": 0.0001654147878750057,
+ "maximum_prediction_kl_nats": 0.02596963456959056,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.4513537883758545,
+ "maximum_per_target_cross_entropy_increase": 0.4725003242492676
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -6.882668153646766e-05,
+ "mean_prediction_kl_nats": 2.579188971355897e-05,
+ "prediction_kl_p99_nats": 0.00016557841705998745,
+ "maximum_prediction_kl_nats": 0.025976165433402816,
+ "top1_changed_fraction": 0.0011501303481061187,
+ "accuracy_change": -0.0002300260696212475,
+ "maximum_absolute_logit_change": 0.4514756202697754,
+ "maximum_per_target_cross_entropy_increase": 0.47253847122192383
+ }
+ },
+ {
+ "name": "fp16-joint-0.95",
+ "encoding": "fp16",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3873085,
+ "unique_rows": 29761,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 466,
+ "max_block_bytes": 8192,
+ "raw_blocks": 1,
+ "compressed_blocks": 465,
+ "zlib_blocks": 465,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29761
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "8c3e008b6398ecab23f6d1ea0550e6451038531f84bbc9c43f87616bd38ddd08",
+ "table_sha256": "54139fd1a85fcaf991b1b9b6e136649f8f72af9954d8448bc5c18327703a6744",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.49241547496058047,
+ "representatives_sha256": "981bcaf7dfaf2fd938d900cf558ed6aa9e8064d37250c908052ada39bec8c7e4",
+ "structurally_merged_rows": 13391,
+ "backend_artifact_bytes": 3874291,
+ "backend_with_key_registry_bytes": 4610398,
+ "backend_with_datrie_key_index_bytes": 4220451,
+ "full_model_serving_artifact_bytes": 5438396,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9500031851017586,
+ "mean_cosine": 0.9996657232299763,
+ "p05_cosine": 0.9986488635607634,
+ "relative_l2_mean": 0.013099563214621524,
+ "relative_l2_p95": 0.05224116177040271,
+ "relative_l2_max": 0.31642273328501797,
+ "weighted_relative_l2_mean": 0.0021971459755752215,
+ "weighted_relative_frobenius": 0.01618432619570471,
+ "weighted_mean_cosine": 0.9999450786415176,
+ "max_relative_norm_change": 0.010116184706844977
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9500094261912371,
+ "mean_cosine": 0.9996657303396614,
+ "p05_cosine": 0.9986487273419825,
+ "relative_l2_mean": 0.012957383484852904,
+ "relative_l2_p95": 0.052251951384778984,
+ "relative_l2_max": 0.31644699684229655,
+ "weighted_relative_l2_mean": 0.0021305421334455736,
+ "weighted_relative_frobenius": 0.01618312405999657,
+ "weighted_mean_cosine": 0.9999450856022387,
+ "max_relative_norm_change": 0.010087895669284265
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.21659,
+ "warm_p50_ms": 0.1462125,
+ "warm_p95_ms": 0.14910364999999998,
+ "warm_p99_ms": 0.16986126,
+ "samples_ms": [
+ 0.148482,
+ 0.147645,
+ 0.147994,
+ 0.146109,
+ 0.147156,
+ 0.147575,
+ 0.146248,
+ 0.145759,
+ 0.146667,
+ 0.146248,
+ 0.146178,
+ 0.160915,
+ 0.146666,
+ 0.145759,
+ 0.146247,
+ 0.145759,
+ 0.146667,
+ 0.146178,
+ 0.146737,
+ 0.145829,
+ 0.147645,
+ 0.14527,
+ 0.146248,
+ 0.146247,
+ 0.14534,
+ 0.146178,
+ 0.1452,
+ 0.1452,
+ 0.146597,
+ 0.145689,
+ 0.145689,
+ 0.145619,
+ 0.145689,
+ 0.14562,
+ 0.145689,
+ 0.146667,
+ 0.146178,
+ 0.145689,
+ 0.175581,
+ 0.146248
+ ],
+ "warm_rows_per_second": 3501752.586133197
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746352945099112,
+ "character_perplexity": 6.5184413754117205,
+ "next_character_accuracy": 0.49279251648520167,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": -1.162127804015254e-05,
+ "mean_prediction_kl_nats": 4.178353212325416e-05,
+ "prediction_kl_p99_nats": 0.00018900797759844118,
+ "maximum_prediction_kl_nats": 0.035442818935089986,
+ "top1_changed_fraction": 0.0015335071308081583,
+ "accuracy_change": -0.00015335071308081316,
+ "maximum_absolute_logit_change": 1.071639060974121,
+ "maximum_per_target_cross_entropy_increase": 0.915989875793457
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -8.258415290237982e-06,
+ "mean_prediction_kl_nats": 4.178929646424802e-05,
+ "prediction_kl_p99_nats": 0.00018906349936431463,
+ "maximum_prediction_kl_nats": 0.035441682833820196,
+ "top1_changed_fraction": 0.0015335071308081583,
+ "accuracy_change": -0.00015335071308081316,
+ "maximum_absolute_logit_change": 1.071523666381836,
+ "maximum_per_target_cross_entropy_increase": 0.9160799980163574
+ }
+ },
+ {
+ "name": "fp16-joint-0.9",
+ "encoding": "fp16",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "fp16",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 128,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 5523456,
+ "artifact_bytes": 3869182,
+ "unique_rows": 29729,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 128,
+ "block_rows": 64,
+ "blocks": 465,
+ "max_block_bytes": 8192,
+ "raw_blocks": 0,
+ "compressed_blocks": 465,
+ "zlib_blocks": 465,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29729
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "cad838d877ad94f85d1fa2c1318b41d0abb0dc96b9c459156207a621e6d8e16b",
+ "table_sha256": "f80003ec63e3a75bc7dea0b14f1547a0a54d9ac47f2d642014af67fcadde072a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 0,
+ "rounding": "fp16",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.3968648931477219,
+ "representatives_sha256": "ffc008e3152f0a165b2e393c95f5fdcd9c8a18a2c4ea01894a088bfa6295f872",
+ "structurally_merged_rows": 13423,
+ "backend_artifact_bytes": 3870388,
+ "backend_with_key_registry_bytes": 4606495,
+ "backend_with_datrie_key_index_bytes": 4216548,
+ "full_model_serving_artifact_bytes": 5434493,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9002018188155697,
+ "mean_cosine": 0.9995876893226137,
+ "p05_cosine": 0.998635669954663,
+ "relative_l2_mean": 0.013488494525890795,
+ "relative_l2_p95": 0.05243002897845533,
+ "relative_l2_max": 0.448270940806544,
+ "weighted_relative_l2_mean": 0.0024820834648912075,
+ "weighted_relative_frobenius": 0.029883640139702614,
+ "weighted_mean_cosine": 0.9998840531650607,
+ "max_relative_norm_change": 0.010116184706844977
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9001934321830203,
+ "mean_cosine": 0.9995876933680093,
+ "p05_cosine": 0.9986359124618528,
+ "relative_l2_mean": 0.013346473807101505,
+ "relative_l2_p95": 0.05244098744596039,
+ "relative_l2_max": 0.44828597129981507,
+ "weighted_relative_l2_mean": 0.0024155443655948516,
+ "weighted_relative_frobenius": 0.02988267517300625,
+ "weighted_mean_cosine": 0.999884059693872,
+ "max_relative_norm_change": 0.010087895669284265
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 5.44141,
+ "warm_p50_ms": 0.147575,
+ "warm_p95_ms": 0.15954564999999998,
+ "warm_p99_ms": 0.16362526,
+ "samples_ms": [
+ 0.150368,
+ 0.14946,
+ 0.147575,
+ 0.162731,
+ 0.150369,
+ 0.147645,
+ 0.148064,
+ 0.148622,
+ 0.148063,
+ 0.147575,
+ 0.148064,
+ 0.147574,
+ 0.147645,
+ 0.146667,
+ 0.147644,
+ 0.147645,
+ 0.146667,
+ 0.147644,
+ 0.147086,
+ 0.147226,
+ 0.14534,
+ 0.14527,
+ 0.146737,
+ 0.146248,
+ 0.146248,
+ 0.147156,
+ 0.146667,
+ 0.147575,
+ 0.146178,
+ 0.164197,
+ 0.159378,
+ 0.147645,
+ 0.146667,
+ 0.146178,
+ 0.146597,
+ 0.147156,
+ 0.14569,
+ 0.147575,
+ 0.147156,
+ 0.147644
+ ],
+ "warm_rows_per_second": 3469422.327630018
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746600056504465,
+ "character_perplexity": 6.518602455522848,
+ "next_character_accuracy": 0.49263916577212086,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 1.3089880639377128e-05,
+ "mean_prediction_kl_nats": 9.756990594690155e-05,
+ "prediction_kl_p99_nats": 0.00036106261136570783,
+ "maximum_prediction_kl_nats": 0.12539222471158687,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -0.0003067014261616263,
+ "maximum_absolute_logit_change": 1.5011892318725586,
+ "maximum_per_target_cross_entropy_increase": 0.915989875793457
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 1.6452826490492457e-05,
+ "mean_prediction_kl_nats": 9.757705936839794e-05,
+ "prediction_kl_p99_nats": 0.0003615714725105157,
+ "maximum_prediction_kl_nats": 0.12539424050629688,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -0.0003067014261616263,
+ "maximum_absolute_logit_change": 1.501293659210205,
+ "maximum_per_target_cross_entropy_increase": 0.9160799980163574
+ }
+ },
+ {
+ "name": "int8-raw",
+ "encoding": "int8",
+ "policy": "raw",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "raw_mmap",
+ "table_file": "table.rows",
+ "row_bytes": 68,
+ "storage": {
+ "artifact_bytes": 2934336,
+ "unique_rows": 43152,
+ "row_map_bytes": 0,
+ "deduplicate": false
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "126915e8d6c63c5a39ba052c391baeecb89382e4804054d3eecec1f79d5ff7a1",
+ "table_sha256": "f84441c1beaa17ea1a7fb7964f5ef9138678520636b2201e52c2ff5a42fa6216",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.07151377317495644,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 2935184,
+ "backend_with_key_registry_bytes": 3671291,
+ "backend_with_datrie_key_index_bytes": 3281344,
+ "full_model_serving_artifact_bytes": 4499289,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999330222579732,
+ "mean_cosine": 0.9999885624214073,
+ "p05_cosine": 0.9999775423662155,
+ "relative_l2_mean": 0.004676661233739485,
+ "relative_l2_p95": 0.006748828249821454,
+ "relative_l2_max": 0.011753130328384171,
+ "weighted_relative_l2_mean": 0.00202791963996097,
+ "weighted_relative_frobenius": 0.005295016649703832,
+ "weighted_mean_cosine": 0.9999944235324602,
+ "max_relative_norm_change": 0.003025896302998395
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 0.276781,
+ "warm_p50_ms": 0.095543,
+ "warm_p95_ms": 0.1018009,
+ "warm_p99_ms": 0.10668704999999999,
+ "samples_ms": [
+ 0.097499,
+ 0.09701,
+ 0.098895,
+ 0.097359,
+ 0.095543,
+ 0.095962,
+ 0.095543,
+ 0.095543,
+ 0.095613,
+ 0.095193,
+ 0.095124,
+ 0.095962,
+ 0.095543,
+ 0.095543,
+ 0.095613,
+ 0.095194,
+ 0.095613,
+ 0.095124,
+ 0.095054,
+ 0.094076,
+ 0.094565,
+ 0.109302,
+ 0.095473,
+ 0.102597,
+ 0.09652,
+ 0.096102,
+ 0.094705,
+ 0.095124,
+ 0.094146,
+ 0.095543,
+ 0.093727,
+ 0.095683,
+ 0.095124,
+ 0.094635,
+ 0.095124,
+ 0.095613,
+ 0.095683,
+ 0.095683,
+ 0.095124,
+ 0.101759
+ ],
+ "warm_rows_per_second": 5358843.662016056
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874735750473572,
+ "character_perplexity": 6.5190962246128565,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 8.883856400462751e-05,
+ "mean_prediction_kl_nats": 3.7871945751883092e-06,
+ "prediction_kl_p99_nats": 2.8762457161950016e-05,
+ "maximum_prediction_kl_nats": 0.00014814864790065312,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.035512447357177734,
+ "maximum_per_target_cross_entropy_increase": 0.02411508560180664
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-identity",
+ "encoding": "int8",
+ "policy": "identity",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 3289029,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 360,
+ "max_block_bytes": 8192,
+ "raw_blocks": 201,
+ "compressed_blocks": 159,
+ "zlib_blocks": 159,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "126915e8d6c63c5a39ba052c391baeecb89382e4804054d3eecec1f79d5ff7a1",
+ "table_sha256": "4c8101acae791d2cee4243e5df028d24488afb64d8213676581d77214f367fe1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.29085787292569876,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3290282,
+ "backend_with_key_registry_bytes": 4026389,
+ "backend_with_datrie_key_index_bytes": 3636442,
+ "full_model_serving_artifact_bytes": 4854387,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999330222579732,
+ "mean_cosine": 0.9999885624214073,
+ "p05_cosine": 0.9999775423662155,
+ "relative_l2_mean": 0.004676661233739485,
+ "relative_l2_p95": 0.006748828249821454,
+ "relative_l2_max": 0.011753130328384171,
+ "weighted_relative_l2_mean": 0.00202791963996097,
+ "weighted_relative_frobenius": 0.005295016649703832,
+ "weighted_mean_cosine": 0.9999944235324602,
+ "max_relative_norm_change": 0.003025896302998395
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.178212,
+ "warm_p50_ms": 0.112165,
+ "warm_p95_ms": 0.11696625,
+ "warm_p99_ms": 0.13743771999999999,
+ "samples_ms": [
+ 0.116844,
+ 0.119289,
+ 0.112095,
+ 0.111606,
+ 0.113003,
+ 0.112165,
+ 0.112654,
+ 0.112096,
+ 0.112165,
+ 0.112165,
+ 0.113562,
+ 0.111746,
+ 0.112584,
+ 0.112165,
+ 0.112096,
+ 0.112096,
+ 0.113073,
+ 0.112584,
+ 0.113981,
+ 0.113493,
+ 0.111606,
+ 0.114051,
+ 0.111537,
+ 0.110698,
+ 0.110768,
+ 0.114959,
+ 0.11028,
+ 0.114959,
+ 0.113562,
+ 0.111257,
+ 0.112095,
+ 0.112095,
+ 0.111606,
+ 0.112584,
+ 0.149041,
+ 0.113004,
+ 0.110699,
+ 0.112096,
+ 0.109861,
+ 0.111676
+ ],
+ "warm_rows_per_second": 4564703.784603041
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874735750473572,
+ "character_perplexity": 6.5190962246128565,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 8.883856400462751e-05,
+ "mean_prediction_kl_nats": 3.7871945751883092e-06,
+ "prediction_kl_p99_nats": 2.8762457161950016e-05,
+ "maximum_prediction_kl_nats": 0.00014814864790065312,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.035512447357177734,
+ "maximum_per_target_cross_entropy_increase": 0.02411508560180664
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-layout",
+ "encoding": "int8",
+ "policy": "layout",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 3289029,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 360,
+ "max_block_bytes": 8192,
+ "raw_blocks": 201,
+ "compressed_blocks": 159,
+ "zlib_blocks": 159,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "126915e8d6c63c5a39ba052c391baeecb89382e4804054d3eecec1f79d5ff7a1",
+ "table_sha256": "4c8101acae791d2cee4243e5df028d24488afb64d8213676581d77214f367fe1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.4411535288672894,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3290281,
+ "backend_with_key_registry_bytes": 4026388,
+ "backend_with_datrie_key_index_bytes": 3636441,
+ "full_model_serving_artifact_bytes": 4854386,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999330222579732,
+ "mean_cosine": 0.9999885624214073,
+ "p05_cosine": 0.9999775423662155,
+ "relative_l2_mean": 0.004676661233739485,
+ "relative_l2_p95": 0.006748828249821454,
+ "relative_l2_max": 0.011753130328384171,
+ "weighted_relative_l2_mean": 0.00202791963996097,
+ "weighted_relative_frobenius": 0.005295016649703832,
+ "weighted_mean_cosine": 0.9999944235324602,
+ "max_relative_norm_change": 0.003025896302998395
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.243095,
+ "warm_p50_ms": 0.111676,
+ "warm_p95_ms": 0.1140754,
+ "warm_p99_ms": 0.13328490999999998,
+ "samples_ms": [
+ 0.114051,
+ 0.113981,
+ 0.112585,
+ 0.112095,
+ 0.113562,
+ 0.113074,
+ 0.112096,
+ 0.113492,
+ 0.112025,
+ 0.112096,
+ 0.112585,
+ 0.112515,
+ 0.111257,
+ 0.14527,
+ 0.113143,
+ 0.111187,
+ 0.111187,
+ 0.111118,
+ 0.112096,
+ 0.114539,
+ 0.112514,
+ 0.110768,
+ 0.111187,
+ 0.111187,
+ 0.111187,
+ 0.110209,
+ 0.110209,
+ 0.111676,
+ 0.111188,
+ 0.112096,
+ 0.111606,
+ 0.112165,
+ 0.111187,
+ 0.110209,
+ 0.111676,
+ 0.111606,
+ 0.110209,
+ 0.111117,
+ 0.110349,
+ 0.109791
+ ],
+ "warm_rows_per_second": 4584691.428776102
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874735750473572,
+ "character_perplexity": 6.5190962246128565,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 8.883856400462751e-05,
+ "mean_prediction_kl_nats": 3.7871945751883092e-06,
+ "prediction_kl_p99_nats": 2.8762457161950016e-05,
+ "maximum_prediction_kl_nats": 0.00014814864790065312,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.035512447357177734,
+ "maximum_per_target_cross_entropy_increase": 0.02411508560180664
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-dedup",
+ "encoding": "int8",
+ "policy": "dedup",
+ "minimum_cosine": null,
+ "minimum_norm_ratio": null,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 3289029,
+ "unique_rows": 43152,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 360,
+ "max_block_bytes": 8192,
+ "raw_blocks": 201,
+ "compressed_blocks": 159,
+ "zlib_blocks": 159,
+ "xor_blocks": 0,
+ "deduplicate": true,
+ "predictor": "xor",
+ "reordered": false,
+ "external_representatives": false,
+ "marked_rows": 43152
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "126915e8d6c63c5a39ba052c391baeecb89382e4804054d3eecec1f79d5ff7a1",
+ "table_sha256": "4c8101acae791d2cee4243e5df028d24488afb64d8213676581d77214f367fe1",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.2992117360699922,
+ "representatives_sha256": "f48fb38b644244dc0fdf4b7babaea40f0028c483dc013ba186ad13dc9606139b",
+ "structurally_merged_rows": 0,
+ "backend_artifact_bytes": 3290281,
+ "backend_with_key_registry_bytes": 4026388,
+ "backend_with_datrie_key_index_bytes": 3636441,
+ "full_model_serving_artifact_bytes": 4854386,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999330222579732,
+ "mean_cosine": 0.9999885624214073,
+ "p05_cosine": 0.9999775423662155,
+ "relative_l2_mean": 0.004676661233739485,
+ "relative_l2_p95": 0.006748828249821454,
+ "relative_l2_max": 0.011753130328384171,
+ "weighted_relative_l2_mean": 0.00202791963996097,
+ "weighted_relative_frobenius": 0.005295016649703832,
+ "weighted_mean_cosine": 0.9999944235324602,
+ "max_relative_norm_change": 0.003025896302998395
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9999999999999998,
+ "mean_cosine": 1.0,
+ "p05_cosine": 0.9999999999999998,
+ "relative_l2_mean": 0.0,
+ "relative_l2_p95": 0.0,
+ "relative_l2_max": 0.0,
+ "weighted_relative_l2_mean": 0.0,
+ "weighted_relative_frobenius": 0.0,
+ "weighted_mean_cosine": 1.0,
+ "max_relative_norm_change": 0.0
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.186244,
+ "warm_p50_ms": 0.112096,
+ "warm_p95_ms": 0.11689035,
+ "warm_p99_ms": 0.12299251,
+ "samples_ms": [
+ 0.116845,
+ 0.114889,
+ 0.114051,
+ 0.112165,
+ 0.112095,
+ 0.112096,
+ 0.113422,
+ 0.113073,
+ 0.126343,
+ 0.113562,
+ 0.112026,
+ 0.112096,
+ 0.112026,
+ 0.112096,
+ 0.111676,
+ 0.113073,
+ 0.11447,
+ 0.111607,
+ 0.111188,
+ 0.111677,
+ 0.110699,
+ 0.110629,
+ 0.112096,
+ 0.113981,
+ 0.117752,
+ 0.111677,
+ 0.111257,
+ 0.111257,
+ 0.111677,
+ 0.111607,
+ 0.112096,
+ 0.111607,
+ 0.112584,
+ 0.111676,
+ 0.113562,
+ 0.112165,
+ 0.112654,
+ 0.112235,
+ 0.111118,
+ 0.11021
+ ],
+ "warm_rows_per_second": 4567513.55980588
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.874735750473572,
+ "character_perplexity": 6.5190962246128565,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 8.883856400462751e-05,
+ "mean_prediction_kl_nats": 3.7871945751883092e-06,
+ "prediction_kl_p99_nats": 2.8762457161950016e-05,
+ "maximum_prediction_kl_nats": 0.00014814864790065312,
+ "top1_changed_fraction": 0.00038337678270203957,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.035512447357177734,
+ "maximum_per_target_cross_entropy_increase": 0.02411508560180664
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 0.0,
+ "mean_prediction_kl_nats": 0.0,
+ "prediction_kl_p99_nats": 0.0,
+ "maximum_prediction_kl_nats": 0.0,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.0,
+ "maximum_per_target_cross_entropy_increase": 0.0
+ }
+ },
+ {
+ "name": "int8-joint-0.99",
+ "encoding": "int8",
+ "policy": "joint-0.99",
+ "minimum_cosine": 0.99,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2379853,
+ "unique_rows": 29846,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 72,
+ "compressed_blocks": 177,
+ "zlib_blocks": 177,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29846
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "de7eaa6d2a5be1f706c23e4f930aaf1e0e9d4a29ccbc573380e2eb0974d75a71",
+ "table_sha256": "949503f9997f9cb539b3a3a543c888998333f8fbcb7ae4dd3264d1b0dcd90b5a",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24713083100505173,
+ "representatives_sha256": "c080d9c9e31eecbbf986a597348e8008148ab5134dfba58849528f43e1f093e3",
+ "structurally_merged_rows": 13306,
+ "backend_artifact_bytes": 2381103,
+ "backend_with_key_registry_bytes": 3117210,
+ "backend_with_datrie_key_index_bytes": 2727263,
+ "full_model_serving_artifact_bytes": 3945208,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9900045008372362,
+ "mean_cosine": 0.9997080968464004,
+ "p05_cosine": 0.9986578003297385,
+ "relative_l2_mean": 0.015918658014219725,
+ "relative_l2_p95": 0.05205936305802378,
+ "relative_l2_max": 0.1410713069702977,
+ "weighted_relative_l2_mean": 0.0037717178974230347,
+ "weighted_relative_frobenius": 0.014090331843047801,
+ "weighted_mean_cosine": 0.9999544167140528,
+ "max_relative_norm_change": 0.011492262128833867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9900075876747326,
+ "mean_cosine": 0.9997131100119235,
+ "p05_cosine": 0.9986422640608953,
+ "relative_l2_mean": 0.012692939117199618,
+ "relative_l2_p95": 0.05229383057532564,
+ "relative_l2_max": 0.14107486043401196,
+ "weighted_relative_l2_mean": 0.0020412890536581946,
+ "weighted_relative_frobenius": 0.013331484320787465,
+ "weighted_mean_cosine": 0.9999585629111997,
+ "max_relative_norm_change": 0.010096391613368234
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.952263,
+ "warm_p50_ms": 0.109546,
+ "warm_p95_ms": 0.11169695,
+ "warm_p99_ms": 0.12223441999999998,
+ "samples_ms": [
+ 0.112095,
+ 0.11021,
+ 0.109721,
+ 0.109721,
+ 0.10972,
+ 0.111676,
+ 0.109791,
+ 0.107346,
+ 0.109791,
+ 0.110629,
+ 0.109372,
+ 0.128717,
+ 0.11014,
+ 0.109861,
+ 0.10979,
+ 0.109232,
+ 0.109791,
+ 0.110209,
+ 0.108882,
+ 0.110629,
+ 0.108324,
+ 0.110628,
+ 0.109232,
+ 0.109302,
+ 0.107905,
+ 0.108882,
+ 0.109791,
+ 0.108813,
+ 0.107905,
+ 0.108394,
+ 0.109791,
+ 0.108883,
+ 0.107835,
+ 0.109302,
+ 0.108254,
+ 0.108324,
+ 0.11021,
+ 0.108324,
+ 0.108324,
+ 0.107346
+ ],
+ "warm_rows_per_second": 4673835.648951125
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746908137499918,
+ "character_perplexity": 6.518803284369755,
+ "next_character_accuracy": 0.49286919184174205,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 4.389885803966729e-05,
+ "mean_prediction_kl_nats": 1.3655938868588348e-05,
+ "prediction_kl_p99_nats": 0.00014786152414243808,
+ "maximum_prediction_kl_nats": 0.004587678278745104,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.3009474277496338,
+ "maximum_per_target_cross_entropy_increase": 0.13836276531219482
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.493571394066121e-05,
+ "mean_prediction_kl_nats": 9.852603724910931e-06,
+ "prediction_kl_p99_nats": 0.00014666116345027564,
+ "maximum_prediction_kl_nats": 0.0047143691122537665,
+ "top1_changed_fraction": 0.0013034810611869344,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.30302488803863525,
+ "maximum_per_target_cross_entropy_increase": 0.14164996147155762
+ }
+ },
+ {
+ "name": "int8-joint-0.98",
+ "encoding": "int8",
+ "policy": "joint-0.98",
+ "minimum_cosine": 0.98,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2379385,
+ "unique_rows": 29839,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 81,
+ "compressed_blocks": 168,
+ "zlib_blocks": 168,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29839
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "0580c7d3bf518fa0aca5cc3195b31e950e0bf8e54ba66c43e01683cd0702d4dd",
+ "table_sha256": "b4e55e897271198a89e1c24459913e6592921aa13018d6f55162faf28943cdeb",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.24981343699619174,
+ "representatives_sha256": "eeed1d7f988a6a71aa0885d278dcc6a8e00f7dee19de702ddfbc3ac989bc0426",
+ "structurally_merged_rows": 13313,
+ "backend_artifact_bytes": 2380635,
+ "backend_with_key_registry_bytes": 3116742,
+ "backend_with_datrie_key_index_bytes": 2726795,
+ "full_model_serving_artifact_bytes": 3944740,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.981201790935401,
+ "mean_cosine": 0.9997044925472083,
+ "p05_cosine": 0.9986558993875337,
+ "relative_l2_mean": 0.015957505284212074,
+ "relative_l2_p95": 0.05211572202221419,
+ "relative_l2_max": 0.1934387575369315,
+ "weighted_relative_l2_mean": 0.00380829590926195,
+ "weighted_relative_frobenius": 0.014932757752551326,
+ "weighted_mean_cosine": 0.9999513883877973,
+ "max_relative_norm_change": 0.011492262128833867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9812446464931801,
+ "mean_cosine": 0.9997094996992814,
+ "p05_cosine": 0.9986393816828784,
+ "relative_l2_mean": 0.01273253761773353,
+ "relative_l2_p95": 0.052356677148328194,
+ "relative_l2_max": 0.1932106405627494,
+ "weighted_relative_l2_mean": 0.0020788796147398767,
+ "weighted_relative_frobenius": 0.014224178888892137,
+ "weighted_mean_cosine": 0.9999555206537855,
+ "max_relative_norm_change": 0.010096391613368234
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.915877,
+ "warm_p50_ms": 0.107835,
+ "warm_p95_ms": 0.11030079999999999,
+ "warm_p99_ms": 0.11845905999999999,
+ "samples_ms": [
+ 0.112026,
+ 0.11021,
+ 0.109302,
+ 0.107835,
+ 0.107346,
+ 0.109721,
+ 0.107835,
+ 0.109301,
+ 0.107905,
+ 0.107905,
+ 0.107905,
+ 0.107346,
+ 0.10986,
+ 0.108813,
+ 0.109791,
+ 0.107835,
+ 0.107835,
+ 0.107346,
+ 0.106438,
+ 0.106858,
+ 0.107416,
+ 0.108743,
+ 0.107835,
+ 0.107835,
+ 0.106508,
+ 0.107905,
+ 0.107905,
+ 0.109721,
+ 0.106508,
+ 0.107835,
+ 0.122572,
+ 0.108813,
+ 0.108743,
+ 0.107346,
+ 0.106927,
+ 0.107835,
+ 0.107905,
+ 0.10553,
+ 0.105391,
+ 0.10595
+ ],
+ "warm_rows_per_second": 4747994.621412343
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8747143007178286,
+ "character_perplexity": 6.5189563930908525,
+ "next_character_accuracy": 0.49286919184174205,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 6.738713276056816e-05,
+ "mean_prediction_kl_nats": 1.4004362616397858e-05,
+ "prediction_kl_p99_nats": 0.00015551443616178517,
+ "maximum_prediction_kl_nats": 0.004587678278745104,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.44987034797668457,
+ "maximum_per_target_cross_entropy_increase": 0.24598169326782227
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -2.1449525698997674e-05,
+ "mean_prediction_kl_nats": 1.0217633862415313e-05,
+ "prediction_kl_p99_nats": 0.0001517913261544031,
+ "maximum_prediction_kl_nats": 0.0047143691122537665,
+ "top1_changed_fraction": 0.0013034810611869344,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.44704198837280273,
+ "maximum_per_target_cross_entropy_increase": 0.2515549659729004
+ }
+ },
+ {
+ "name": "int8-joint-0.97",
+ "encoding": "int8",
+ "policy": "joint-0.97",
+ "minimum_cosine": 0.97,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2378706,
+ "unique_rows": 29829,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 81,
+ "compressed_blocks": 168,
+ "zlib_blocks": 168,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29829
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "1e96b6053daa54587ab5ea83b49d3588199e5feda611a7ae86ed5b1eb9754722",
+ "table_sha256": "fd07970de04a13707396af5641daf99a845a5db43183794a0002ffc5509be158",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25833526882342994,
+ "representatives_sha256": "6acf75d3e2ba801464e220ea404b08e05e3fc42548c067684b11cc074397c082",
+ "structurally_merged_rows": 13323,
+ "backend_artifact_bytes": 2379956,
+ "backend_with_key_registry_bytes": 3116063,
+ "backend_with_datrie_key_index_bytes": 2726116,
+ "full_model_serving_artifact_bytes": 3944061,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.97040245654302,
+ "mean_cosine": 0.9996967602717127,
+ "p05_cosine": 0.9986501049778329,
+ "relative_l2_mean": 0.016022564144458906,
+ "relative_l2_p95": 0.05215003659856219,
+ "relative_l2_max": 0.24480937559793678,
+ "weighted_relative_l2_mean": 0.003849943713934918,
+ "weighted_relative_frobenius": 0.016057651017675024,
+ "weighted_mean_cosine": 0.9999465217457014,
+ "max_relative_norm_change": 0.011492262128833867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9702562647044194,
+ "mean_cosine": 0.999701746958545,
+ "p05_cosine": 0.9986360562965715,
+ "relative_l2_mean": 0.012798844304167282,
+ "relative_l2_p95": 0.0524129035256841,
+ "relative_l2_max": 0.24442440666938092,
+ "weighted_relative_l2_mean": 0.0021216697628386925,
+ "weighted_relative_frobenius": 0.0154071093387881,
+ "weighted_mean_cosine": 0.9999506274423351,
+ "max_relative_norm_change": 0.010096391613368234
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 3.1386,
+ "warm_p50_ms": 0.1088125,
+ "warm_p95_ms": 0.11180254999999999,
+ "warm_p99_ms": 0.13681981,
+ "samples_ms": [
+ 0.110768,
+ 0.111607,
+ 0.109721,
+ 0.108813,
+ 0.108324,
+ 0.110279,
+ 0.108394,
+ 0.109791,
+ 0.11028,
+ 0.107836,
+ 0.108813,
+ 0.107904,
+ 0.108813,
+ 0.108812,
+ 0.108813,
+ 0.107974,
+ 0.115518,
+ 0.109301,
+ 0.107416,
+ 0.107836,
+ 0.108394,
+ 0.108254,
+ 0.107416,
+ 0.10972,
+ 0.107346,
+ 0.150439,
+ 0.11028,
+ 0.107835,
+ 0.108883,
+ 0.108254,
+ 0.108883,
+ 0.108324,
+ 0.106438,
+ 0.107346,
+ 0.108813,
+ 0.106927,
+ 0.107905,
+ 0.108813,
+ 0.108464,
+ 0.108813
+ ],
+ "warm_rows_per_second": 4705341.757610568
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746904356031506,
+ "character_perplexity": 6.518800819305351,
+ "next_character_accuracy": 0.49286919184174205,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 4.352069466971398e-05,
+ "mean_prediction_kl_nats": 2.979705460222902e-05,
+ "prediction_kl_p99_nats": 0.00017673648535457696,
+ "maximum_prediction_kl_nats": 0.025972837018001935,
+ "top1_changed_fraction": 0.001686857843888974,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.45468056201934814,
+ "maximum_per_target_cross_entropy_increase": 0.4720644950866699
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": -4.5313843718108086e-05,
+ "mean_prediction_kl_nats": 2.6014357149968005e-05,
+ "prediction_kl_p99_nats": 0.00016585546912271107,
+ "maximum_prediction_kl_nats": 0.02601218763702007,
+ "top1_changed_fraction": 0.0013034810611869344,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 0.4569981098175049,
+ "maximum_per_target_cross_entropy_increase": 0.47402477264404297
+ }
+ },
+ {
+ "name": "int8-joint-0.95",
+ "encoding": "int8",
+ "policy": "joint-0.95",
+ "minimum_cosine": 0.95,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2376894,
+ "unique_rows": 29803,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 83,
+ "compressed_blocks": 166,
+ "zlib_blocks": 166,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29803
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "95b6ba72d207d2380587f8a30590e0fb2ad1961be90b01c11eb12dab3b61921c",
+ "table_sha256": "ee72798e78e294518eb409acad29fc4680c5689b3a67d4326d1ee3a6739d5483",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25348946708254516,
+ "representatives_sha256": "af5093f8bd42c3455666516db33046941def1ac1cb4630fe53dd09be3454ef09",
+ "structurally_merged_rows": 13349,
+ "backend_artifact_bytes": 2378144,
+ "backend_with_key_registry_bytes": 3114251,
+ "backend_with_datrie_key_index_bytes": 2724304,
+ "full_model_serving_artifact_bytes": 3942249,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9500310490484243,
+ "mean_cosine": 0.9996589611534579,
+ "p05_cosine": 0.9986374972004458,
+ "relative_l2_mean": 0.01627759091524889,
+ "relative_l2_p95": 0.0523743928211285,
+ "relative_l2_max": 0.3169463056015357,
+ "weighted_relative_l2_mean": 0.003892891244357256,
+ "weighted_relative_frobenius": 0.017019670055103134,
+ "weighted_mean_cosine": 0.9999402561558993,
+ "max_relative_norm_change": 0.011492262128833867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9501599222200492,
+ "mean_cosine": 0.9996639282393538,
+ "p05_cosine": 0.9986254662412041,
+ "relative_l2_mean": 0.013056697783505997,
+ "relative_l2_p95": 0.05263862802663197,
+ "relative_l2_max": 0.31691284212741533,
+ "weighted_relative_l2_mean": 0.0021653347140213747,
+ "weighted_relative_frobenius": 0.01640990629562992,
+ "weighted_mean_cosine": 0.9999443442918322,
+ "max_relative_norm_change": 0.010096391613368234
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.946118,
+ "warm_p50_ms": 0.10982600000000001,
+ "warm_p95_ms": 0.11733729999999999,
+ "warm_p99_ms": 0.13381163999999998,
+ "samples_ms": [
+ 0.110768,
+ 0.112235,
+ 0.109791,
+ 0.114959,
+ 0.108813,
+ 0.111607,
+ 0.109372,
+ 0.10979,
+ 0.11873,
+ 0.112165,
+ 0.112165,
+ 0.109231,
+ 0.108813,
+ 0.108813,
+ 0.108882,
+ 0.109302,
+ 0.11021,
+ 0.110279,
+ 0.10979,
+ 0.11014,
+ 0.112166,
+ 0.110699,
+ 0.111677,
+ 0.109721,
+ 0.110699,
+ 0.10979,
+ 0.108254,
+ 0.109791,
+ 0.109651,
+ 0.108813,
+ 0.108813,
+ 0.143454,
+ 0.114051,
+ 0.111188,
+ 0.108813,
+ 0.109232,
+ 0.111257,
+ 0.117264,
+ 0.109861,
+ 0.109791
+ ],
+ "warm_rows_per_second": 4661919.763990312
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8747468378917542,
+ "character_perplexity": 6.51916850495957,
+ "next_character_accuracy": 0.49302254255482286,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 9.992702864821589e-05,
+ "mean_prediction_kl_nats": 4.551468955017082e-05,
+ "prediction_kl_p99_nats": 0.00020112406909674518,
+ "maximum_prediction_kl_nats": 0.0364496174654843,
+ "top1_changed_fraction": 0.001993559270050606,
+ "accuracy_change": 7.667535654037883e-05,
+ "maximum_absolute_logit_change": 1.067946434020996,
+ "maximum_per_target_cross_entropy_increase": 0.9218840599060059
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 1.1087479647818034e-05,
+ "mean_prediction_kl_nats": 4.1565697000894676e-05,
+ "prediction_kl_p99_nats": 0.00018631396767026089,
+ "maximum_prediction_kl_nats": 0.03575100198331784,
+ "top1_changed_fraction": 0.0016101824873485662,
+ "accuracy_change": 7.667535654037883e-05,
+ "maximum_absolute_logit_change": 1.0640273094177246,
+ "maximum_per_target_cross_entropy_increase": 0.9170961380004883
+ }
+ },
+ {
+ "name": "int8-joint-0.9",
+ "encoding": "int8",
+ "policy": "joint-0.9",
+ "minimum_cosine": 0.9,
+ "minimum_norm_ratio": 0.99,
+ "metadata": {
+ "schema_version": 1,
+ "shape": [
+ 43152,
+ 64
+ ],
+ "encoding": "int8",
+ "byteorder": "little",
+ "backend": "native",
+ "table_file": "table.men",
+ "row_bytes": 68,
+ "storage": {
+ "format": "MEFTNAT1",
+ "format_version": 2,
+ "input_bytes": 2934336,
+ "artifact_bytes": 2374644,
+ "unique_rows": 29770,
+ "row_map_bytes": 345216,
+ "n_rows": 43152,
+ "row_bytes": 68,
+ "block_rows": 120,
+ "blocks": 249,
+ "max_block_bytes": 8192,
+ "raw_blocks": 81,
+ "compressed_blocks": 168,
+ "zlib_blocks": 168,
+ "xor_blocks": 0,
+ "deduplicate": false,
+ "predictor": "xor",
+ "reordered": true,
+ "external_representatives": true,
+ "marked_rows": 29770
+ },
+ "source_float32_sha256": "8b300423d3217435173b444876d6a4b31be10e2baf79c32d3766439aebd52b75",
+ "decoded_float32_sha256": "d0e4d71cb12e446a9b61ad1cc09eca808fe045b0c6c503b2f75e691c4d6ac502",
+ "table_sha256": "ee3ff9282668949db928a8682861d85d27e2e3715872e58448a8660fdc2c5855",
+ "source_bytes_preserved": false,
+ "int8_scale_bytes_per_physical_row": 4,
+ "rounding": "nearest-even; symmetric [-127,127] per-row scale",
+ "contract": "frozen original IDs; exact encoded representative bytes; quantization/sharing may change float32 rows"
+ },
+ "export_seconds": 0.25261233001947403,
+ "representatives_sha256": "9e964f3458307b77a92f7c4e6d55fb96d66dbe9a75ab57c4663e6881e11279c8",
+ "structurally_merged_rows": 13382,
+ "backend_artifact_bytes": 2375894,
+ "backend_with_key_registry_bytes": 3112001,
+ "backend_with_datrie_key_index_bytes": 2722054,
+ "full_model_serving_artifact_bytes": 3939999,
+ "fidelity_vs_float32": {
+ "rows": 43152,
+ "minimum_cosine": 0.9002505497681637,
+ "mean_cosine": 0.9995771039885059,
+ "p05_cosine": 0.9986266003056047,
+ "relative_l2_mean": 0.01668135380304124,
+ "relative_l2_p95": 0.05261329348892294,
+ "relative_l2_max": 0.4476735226377543,
+ "weighted_relative_l2_mean": 0.004175748016733891,
+ "weighted_relative_frobenius": 0.030339692323942288,
+ "weighted_mean_cosine": 0.9998793091128563,
+ "max_relative_norm_change": 0.011492262128833867
+ },
+ "added_fidelity_vs_quantization_only": {
+ "rows": 43152,
+ "minimum_cosine": 0.9001671583065514,
+ "mean_cosine": 0.9995820648045777,
+ "p05_cosine": 0.9986114016095644,
+ "relative_l2_mean": 0.013464057944793923,
+ "relative_l2_p95": 0.052907295257222174,
+ "relative_l2_max": 0.4478462401559686,
+ "weighted_relative_l2_mean": 0.002449430186867565,
+ "weighted_relative_frobenius": 0.029967218368968207,
+ "weighted_mean_cosine": 0.9998835445533141,
+ "max_relative_norm_change": 0.010096391613368234
+ },
+ "all_encoded_rows_and_original_ids_verified": true,
+ "queries": {
+ "first_empty_application_cache_ms": 2.929007,
+ "warm_p50_ms": 0.109302,
+ "warm_p95_ms": 0.11542689999999999,
+ "warm_p99_ms": 0.13654944,
+ "samples_ms": [
+ 0.113073,
+ 0.10986,
+ 0.109302,
+ 0.107486,
+ 0.108813,
+ 0.113492,
+ 0.109721,
+ 0.11447,
+ 0.111257,
+ 0.111187,
+ 0.109302,
+ 0.108883,
+ 0.14946,
+ 0.113562,
+ 0.116356,
+ 0.108813,
+ 0.108324,
+ 0.110699,
+ 0.110629,
+ 0.111676,
+ 0.108812,
+ 0.115378,
+ 0.108813,
+ 0.108324,
+ 0.107905,
+ 0.108324,
+ 0.108743,
+ 0.108813,
+ 0.109232,
+ 0.10979,
+ 0.11021,
+ 0.11028,
+ 0.108394,
+ 0.108323,
+ 0.111746,
+ 0.108813,
+ 0.108813,
+ 0.110698,
+ 0.108743,
+ 0.109301
+ ],
+ "warm_rows_per_second": 4684269.27229145
+ },
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8747692168674108,
+ "character_perplexity": 6.519314398905316,
+ "next_character_accuracy": 0.49286919184174205,
+ "unknown_targets": 0
+ },
+ "validation_delta": {
+ "perplexity_relative_change": 0.0001223084910055302,
+ "mean_prediction_kl_nats": 0.00010083192637675476,
+ "prediction_kl_p99_nats": 0.00035157782737860616,
+ "maximum_prediction_kl_nats": 0.1234834254535015,
+ "top1_changed_fraction": 0.0022235853396718294,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.488051414489746,
+ "maximum_per_target_cross_entropy_increase": 0.9218840599060059
+ },
+ "added_validation_delta_vs_quantization_only": {
+ "perplexity_relative_change": 3.346695384478151e-05,
+ "mean_prediction_kl_nats": 9.71627967356333e-05,
+ "prediction_kl_p99_nats": 0.00034940811032745464,
+ "maximum_prediction_kl_nats": 0.12394890936422365,
+ "top1_changed_fraction": 0.0018402085569697898,
+ "accuracy_change": -7.667535654043434e-05,
+ "maximum_absolute_logit_change": 1.493788242340088,
+ "maximum_per_target_cross_entropy_increase": 0.9170961380004883
+ }
+ }
+ ],
+ "baseline_quality": {
+ "validation": {
+ "targets": 13042,
+ "cross_entropy_nats": 1.8746469158554788,
+ "character_perplexity": 6.518517128911684,
+ "next_character_accuracy": 0.4929458671982825,
+ "unknown_targets": 0
+ },
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7494370523487461,
+ "character_perplexity": 5.751364047620025,
+ "next_character_accuracy": 0.5255692708732653,
+ "unknown_targets": 0
+ }
+ },
+ "selection": {
+ "fp32": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 8186195,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7491682414699585,
+ "character_perplexity": 5.749818226172276,
+ "next_character_accuracy": 0.5252625929617419,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -0.0002687747523804419,
+ "mean_prediction_kl_nats": 0.0001433063275818475,
+ "prediction_kl_p99_nats": 0.00047574067175172846,
+ "maximum_prediction_kl_nats": 0.14949254966897688,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": -0.00030667791152338975,
+ "maximum_absolute_logit_change": 1.5382030010223389,
+ "maximum_per_target_cross_entropy_increase": 0.7367382049560547
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0005444561415285015,
+ 1.527656112117525e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "fp16": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 4606495,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7491683695294322,
+ "character_perplexity": 5.749818962491019,
+ "next_character_accuracy": 0.5252625929617419,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -0.0002686467273177534,
+ "mean_prediction_kl_nats": 0.00014329286792632542,
+ "prediction_kl_p99_nats": 0.0004777807634440265,
+ "maximum_prediction_kl_nats": 0.1495226025186514,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": -0.00030667791152338975,
+ "maximum_absolute_logit_change": 1.538262128829956,
+ "maximum_per_target_cross_entropy_increase": 0.7369163036346436
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0005467265885629537,
+ 1.8822879602251045e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "int8": {
+ "feasible": true,
+ "name": "int8-joint-0.95",
+ "backend_with_key_registry_bytes": 3114251,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7496451964622277,
+ "character_perplexity": 5.752561284785639,
+ "next_character_accuracy": 0.5254159319175037,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00020816577697060302,
+ "mean_prediction_kl_nats": 6.48036425053271e-05,
+ "prediction_kl_p99_nats": 0.00021414928574367274,
+ "maximum_prediction_kl_nats": 0.15112539268236266,
+ "top1_changed_fraction": 0.0026067622479490913,
+ "accuracy_change": -0.00015333895576163936,
+ "maximum_absolute_logit_change": 1.0701985359191895,
+ "maximum_per_target_cross_entropy_increase": 0.7534070014953613
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 1.1390863152666506e-05,
+ 0.0004288786464272205
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ }
+ },
+ "fixed_quantization_test_controls": {
+ "fp16": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7494358820052505,
+ "character_perplexity": 5.751357316552459,
+ "next_character_accuracy": 0.5255692708732653,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": -1.1703428108266528e-06,
+ "mean_prediction_kl_nats": 3.8673117189102365e-09,
+ "prediction_kl_p99_nats": 2.5890530409318675e-08,
+ "maximum_prediction_kl_nats": 7.611372185147914e-08,
+ "top1_changed_fraction": 0.0,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.001432180404663086,
+ "maximum_per_target_cross_entropy_increase": 0.0007014274597167969
+ }
+ },
+ "int8": {
+ "metrics": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.749516485595348,
+ "character_perplexity": 5.751820915283718,
+ "next_character_accuracy": 0.5255692708732653,
+ "unknown_targets": 0
+ },
+ "delta": {
+ "perplexity_relative_change": 7.943640150563587e-05,
+ "mean_prediction_kl_nats": 3.5398041673480987e-06,
+ "prediction_kl_p99_nats": 2.894596014564925e-05,
+ "maximum_prediction_kl_nats": 0.00010490708635250501,
+ "top1_changed_fraction": 0.0013033811239745456,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.03159666061401367,
+ "maximum_per_target_cross_entropy_increase": 0.022692203521728516
+ }
+ }
+ },
+ "checkpoint_sha256": "d2b8acfabccdc0f727220d8529bbe5b60c75d87bdefe10184232ad1310d868aa",
+ "registry_sha256": "abab407b9c927abc758d3d85a11cec2c2535254fe5af55b6cce10b650a992d4d"
+ }
+ ],
+ "failures": []
+}
diff --git a/docs/quantized_structural_evidence_20261031/resources-launch-failure.txt b/docs/quantized_structural_evidence_20261031/resources-launch-failure.txt
new file mode 100644
index 0000000..09726bf
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/resources-launch-failure.txt
@@ -0,0 +1 @@
+/llm-align/duyimin/ngram-quantized-20261031/run.sh: 行 5: /usr/bin/time: 没有那个文件或目录
diff --git a/docs/quantized_structural_evidence_20261031/resources.txt b/docs/quantized_structural_evidence_20261031/resources.txt
new file mode 100644
index 0000000..e69de29
diff --git a/docs/quantized_structural_evidence_20261031/run.log b/docs/quantized_structural_evidence_20261031/run.log
new file mode 100644
index 0000000..0cdc0e3
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/run.log
@@ -0,0 +1,58 @@
+{"category": "sgns", "order": 4, "seed": 3, "encoding": "fp32", "completed_variants": 9}
+{"category": "sgns", "order": 4, "seed": 3, "encoding": "fp16", "completed_variants": 18}
+{"category": "sgns", "order": 4, "seed": 3, "encoding": "int8", "completed_variants": 27}
+{"category": "sgns", "order": 4, "seed": 7, "encoding": "fp32", "completed_variants": 9}
+{"category": "sgns", "order": 4, "seed": 7, "encoding": "fp16", "completed_variants": 18}
+{"category": "sgns", "order": 4, "seed": 7, "encoding": "int8", "completed_variants": 27}
+{"category": "sgns", "order": 4, "seed": 11, "encoding": "fp32", "completed_variants": 9}
+{"category": "sgns", "order": 4, "seed": 11, "encoding": "fp16", "completed_variants": 18}
+{"category": "sgns", "order": 4, "seed": 11, "encoding": "int8", "completed_variants": 27}
+{"category": "sgns", "order": 8, "seed": 3, "encoding": "fp32", "completed_variants": 9}
+{"category": "sgns", "order": 8, "seed": 3, "encoding": "fp16", "completed_variants": 18}
+{"category": "sgns", "order": 8, "seed": 3, "encoding": "int8", "completed_variants": 27}
+{"category": "sgns", "order": 8, "seed": 7, "encoding": "fp32", "completed_variants": 9}
+{"category": "sgns", "order": 8, "seed": 7, "encoding": "fp16", "completed_variants": 18}
+{"category": "sgns", "order": 8, "seed": 7, "encoding": "int8", "completed_variants": 27}
+{"category": "sgns", "order": 8, "seed": 11, "encoding": "fp32", "completed_variants": 9}
+{"category": "sgns", "order": 8, "seed": 11, "encoding": "fp16", "completed_variants": 18}
+{"category": "sgns", "order": 8, "seed": 11, "encoding": "int8", "completed_variants": 27}
+{"category": "sgns", "order": 12, "seed": 3, "encoding": "fp32", "completed_variants": 9}
+{"category": "sgns", "order": 12, "seed": 3, "encoding": "fp16", "completed_variants": 18}
+{"category": "sgns", "order": 12, "seed": 3, "encoding": "int8", "completed_variants": 27}
+{"category": "sgns", "order": 12, "seed": 7, "encoding": "fp32", "completed_variants": 9}
+{"category": "sgns", "order": 12, "seed": 7, "encoding": "fp16", "completed_variants": 18}
+{"category": "sgns", "order": 12, "seed": 7, "encoding": "int8", "completed_variants": 27}
+{"category": "sgns", "order": 12, "seed": 11, "encoding": "fp32", "completed_variants": 9}
+{"category": "sgns", "order": 12, "seed": 11, "encoding": "fp16", "completed_variants": 18}
+{"category": "sgns", "order": 12, "seed": 11, "encoding": "int8", "completed_variants": 27}
+{"category": "gru", "order": 4, "seed": 3, "encoding": "fp32", "completed_variants": 9}
+{"category": "gru", "order": 4, "seed": 3, "encoding": "fp16", "completed_variants": 18}
+{"category": "gru", "order": 4, "seed": 3, "encoding": "int8", "completed_variants": 27}
+{"category": "gru", "order": 4, "seed": 7, "encoding": "fp32", "completed_variants": 9}
+{"category": "gru", "order": 4, "seed": 7, "encoding": "fp16", "completed_variants": 18}
+{"category": "gru", "order": 4, "seed": 7, "encoding": "int8", "completed_variants": 27}
+{"category": "gru", "order": 4, "seed": 11, "encoding": "fp32", "completed_variants": 9}
+{"category": "gru", "order": 4, "seed": 11, "encoding": "fp16", "completed_variants": 18}
+{"category": "gru", "order": 4, "seed": 11, "encoding": "int8", "completed_variants": 27}
+{"category": "gru", "order": 8, "seed": 3, "encoding": "fp32", "completed_variants": 9}
+{"category": "gru", "order": 8, "seed": 3, "encoding": "fp16", "completed_variants": 18}
+{"category": "gru", "order": 8, "seed": 3, "encoding": "int8", "completed_variants": 27}
+{"category": "gru", "order": 8, "seed": 7, "encoding": "fp32", "completed_variants": 9}
+{"category": "gru", "order": 8, "seed": 7, "encoding": "fp16", "completed_variants": 18}
+{"category": "gru", "order": 8, "seed": 7, "encoding": "int8", "completed_variants": 27}
+{"category": "gru", "order": 8, "seed": 11, "encoding": "fp32", "completed_variants": 9}
+{"category": "gru", "order": 8, "seed": 11, "encoding": "fp16", "completed_variants": 18}
+{"category": "gru", "order": 8, "seed": 11, "encoding": "int8", "completed_variants": 27}
+{"category": "gru", "order": 12, "seed": 3, "encoding": "fp32", "completed_variants": 9}
+{"category": "gru", "order": 12, "seed": 3, "encoding": "fp16", "completed_variants": 18}
+{"category": "gru", "order": 12, "seed": 3, "encoding": "int8", "completed_variants": 27}
+{"category": "gru", "order": 12, "seed": 7, "encoding": "fp32", "completed_variants": 9}
+{"category": "gru", "order": 12, "seed": 7, "encoding": "fp16", "completed_variants": 18}
+{"category": "gru", "order": 12, "seed": 7, "encoding": "int8", "completed_variants": 27}
+{"category": "gru", "order": 12, "seed": 11, "encoding": "fp32", "completed_variants": 9}
+{"category": "gru", "order": 12, "seed": 11, "encoding": "fp16", "completed_variants": 18}
+{"category": "gru", "order": 12, "seed": 11, "encoding": "int8", "completed_variants": 27}
+
+real 8m48.349s
+user 8m7.516s
+sys 0m27.774s
diff --git a/docs/quantized_structural_evidence_20261031/run.sh b/docs/quantized_structural_evidence_20261031/run.sh
new file mode 100644
index 0000000..a819348
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/run.sh
@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+set -uo pipefail
+cd /llm-align/duyimin/ngram-quantized-20261031/source
+export CUDA_VISIBLE_DEVICES="" OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1
+time nice -n 19 /llm-align/duyimin/meft-cuda-20261030/compat/bin/python -m tools.benchmark_quantized_structural \
+ --sgns-dir artifacts/native-ngram-20261018/trained-reference \
+ --gru-dir artifacts/causal-lm-compression-20261024 \
+ --corpus artifacts/native-ngram-20261018/corpus/tinyshakespeare.txt \
+ --native-library build/libmeft_rows.so --structural-library build/libmeft_structural.so \
+ --planner native --output-dir /llm-align/duyimin/ngram-quantized-20261031/results \
+ --orders 4,8,12 --seeds 3,7,11 --thresholds .99,.98,.97,.95,.90 \
+ --norm-ratio .99 --repeat 40 2> /llm-align/duyimin/ngram-quantized-20261031/resources.txt
+result=$?
+printf '%s\n' "$result" > /llm-align/duyimin/ngram-quantized-20261031/exit.txt
+exit "$result"
diff --git a/docs/quantized_structural_evidence_20261031/storage-controls.png b/docs/quantized_structural_evidence_20261031/storage-controls.png
new file mode 100644
index 0000000..85b4603
Binary files /dev/null and b/docs/quantized_structural_evidence_20261031/storage-controls.png differ
diff --git a/docs/quantized_structural_evidence_20261031/storage-controls.svg b/docs/quantized_structural_evidence_20261031/storage-controls.svg
new file mode 100644
index 0000000..fda7909
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/storage-controls.svg
@@ -0,0 +1,1526 @@
+
+
+
diff --git a/docs/quantized_structural_evidence_20261031/summary.json b/docs/quantized_structural_evidence_20261031/summary.json
new file mode 100644
index 0000000..121b28f
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/summary.json
@@ -0,0 +1,3586 @@
+{
+ "source_commit": "3a54de2d8a7b97ee1432980c41b4edb0f64566d3",
+ "planner": "native",
+ "trained_models": 18,
+ "variant_count": 486,
+ "storage_denominator": "actual encoded backend + codec metadata + common keys JSON; whole GRU serving artifacts also include unchanged serialized backbone/registry",
+ "baseline_rule": "best of raw mmap / native identity / native layout / exact encoded-byte dedup, matched encoding; structural candidates compared against this lower byte count",
+ "error_rule": "final decoded vectors vs original FP32; additional error separately vs quantization-only; frequency weights from actual query traces",
+ "uncertainty": "three-seed min/mean/max are descriptive ranges, not confidence intervals or cross-corpus evidence",
+ "test_selection": "retain per-encoding validation choices; no test-based replacement",
+ "test_budget_exceedances": [
+ {
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.95",
+ "delta": {
+ "perplexity_relative_change": 0.00041829374079927037,
+ "mean_prediction_kl_nats": 0.0001224659868776353,
+ "prediction_kl_p99_nats": 6.492612482710993e-05,
+ "maximum_prediction_kl_nats": 0.049823547914828695,
+ "top1_changed_fraction": 0.0002300084336425669,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6587400436401367,
+ "maximum_per_target_cross_entropy_increase": 0.3685579299926758
+ }
+ },
+ {
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.95",
+ "delta": {
+ "perplexity_relative_change": 0.0004171065973414183,
+ "mean_prediction_kl_nats": 0.0001224485145529409,
+ "prediction_kl_p99_nats": 6.468627058898175e-05,
+ "maximum_prediction_kl_nats": 0.049807480276882234,
+ "top1_changed_fraction": 0.0002300084336425669,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6586532592773438,
+ "maximum_per_target_cross_entropy_increase": 0.3685433864593506
+ }
+ },
+ {
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.95",
+ "delta": {
+ "perplexity_relative_change": 0.00010423505069745625,
+ "mean_prediction_kl_nats": 0.00013875969726630407,
+ "prediction_kl_p99_nats": 0.00038196038107028794,
+ "maximum_prediction_kl_nats": 0.09296217712975313,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": 7.666947788076417e-05,
+ "maximum_absolute_logit_change": 1.504372239112854,
+ "maximum_per_target_cross_entropy_increase": 0.5252523422241211
+ }
+ },
+ {
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.95",
+ "delta": {
+ "perplexity_relative_change": 0.00010519237107762973,
+ "mean_prediction_kl_nats": 0.0001388013511152666,
+ "prediction_kl_p99_nats": 0.0003827823135934586,
+ "maximum_prediction_kl_nats": 0.09302745751300914,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": 7.666947788076417e-05,
+ "maximum_absolute_logit_change": 1.5042415857315063,
+ "maximum_per_target_cross_entropy_increase": 0.5253055095672607
+ }
+ },
+ {
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "int8",
+ "name": "int8-joint-0.95",
+ "delta": {
+ "perplexity_relative_change": 6.299487197111428e-05,
+ "mean_prediction_kl_nats": 0.00013896683904643248,
+ "prediction_kl_p99_nats": 0.00035191132917633707,
+ "maximum_prediction_kl_nats": 0.09188349093556392,
+ "top1_changed_fraction": 0.0017633979912596795,
+ "accuracy_change": 0.00015333895576163936,
+ "maximum_absolute_logit_change": 1.4925585985183716,
+ "maximum_per_target_cross_entropy_increase": 0.5185873508453369
+ }
+ },
+ {
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp32",
+ "name": "fp32-joint-0.9",
+ "delta": {
+ "perplexity_relative_change": -0.0002687747523804419,
+ "mean_prediction_kl_nats": 0.0001433063275818475,
+ "prediction_kl_p99_nats": 0.00047574067175172846,
+ "maximum_prediction_kl_nats": 0.14949254966897688,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": -0.00030667791152338975,
+ "maximum_absolute_logit_change": 1.5382030010223389,
+ "maximum_per_target_cross_entropy_increase": 0.7367382049560547
+ }
+ },
+ {
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp16",
+ "name": "fp16-joint-0.9",
+ "delta": {
+ "perplexity_relative_change": -0.0002686467273177534,
+ "mean_prediction_kl_nats": 0.00014329286792632542,
+ "prediction_kl_p99_nats": 0.0004777807634440265,
+ "maximum_prediction_kl_nats": 0.1495226025186514,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": -0.00030667791152338975,
+ "maximum_absolute_logit_change": 1.538262128829956,
+ "maximum_per_target_cross_entropy_increase": 0.7369163036346436
+ }
+ }
+ ],
+ "models": [
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 2802233,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 2679981,
+ "joint_099_backend_with_keys_bytes": 2039832,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2388632605977431,
+ "joint_099_saving_vs_raw_encoded": 0.2720690963242528,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.373756760360657,
+ "joint_099_weighted_relative_frobenius": 0.00684819310974004,
+ "joint_099_added_weighted_relative_frobenius": 0.00684819310974004,
+ "joint_099_max_relative_l2": 0.14212765851757603,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1351826966148333,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.24137410952742033,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.3789356019073995
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 1436986,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1401408,
+ "joint_099_backend_with_keys_bytes": 1081785,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2280727668173722,
+ "joint_099_saving_vs_raw_encoded": 0.2471847324886951,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.590378864561812,
+ "joint_099_weighted_relative_frobenius": 0.0068512795300496,
+ "joint_099_added_weighted_relative_frobenius": 0.006848195096817759,
+ "joint_099_max_relative_l2": 0.14214298103013276,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0375786909548252,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.23270187036322465,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.6324476448111644
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 775737,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 775737,
+ "joint_099_backend_with_keys_bytes": 621162,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.19926212105391394,
+ "joint_099_saving_vs_raw_encoded": 0.19926212105391394,
+ "joint_099_factor_vs_raw_float32_with_keys": 4.511275641459072,
+ "joint_099_weighted_relative_frobenius": 0.009423937975726222,
+ "joint_099_added_weighted_relative_frobenius": 0.006974880559954164,
+ "joint_099_max_relative_l2": 0.14188875915362323,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0675390597903482,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.20669003114223405,
+ "joint_099_factor_vs_raw_float32_with_datrie": 4.676268026779755
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 2802233,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 2679584,
+ "joint_099_backend_with_keys_bytes": 2029029,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.24278208856300088,
+ "joint_099_saving_vs_raw_encoded": 0.2759242361359673,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.3810709457577985,
+ "joint_099_weighted_relative_frobenius": 0.006743679010149183,
+ "joint_099_added_weighted_relative_frobenius": 0.006743679010149183,
+ "joint_099_max_relative_l2": 0.14087181586808722,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1301373730529407,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.24533451295128494,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.386379638518033
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 1436986,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1401405,
+ "joint_099_backend_with_keys_bytes": 1075994,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.23220339587770844,
+ "joint_099_saving_vs_raw_encoded": 0.2512146952023193,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.604320284313853,
+ "joint_099_weighted_relative_frobenius": 0.006748195908759052,
+ "joint_099_added_weighted_relative_frobenius": 0.0067452159501694255,
+ "joint_099_max_relative_l2": 0.1408746714806218,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.039525,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.2369163474762418,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.64699231764423
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 775737,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 775737,
+ "joint_099_backend_with_keys_bytes": 618504,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.20268854005932424,
+ "joint_099_saving_vs_raw_encoded": 0.20268854005932424,
+ "joint_099_factor_vs_raw_float32_with_keys": 4.530662695795015,
+ "joint_099_weighted_relative_frobenius": 0.009414340895132815,
+ "joint_099_added_weighted_relative_frobenius": 0.006861641269675814,
+ "joint_099_max_relative_l2": 0.1411979477508241,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0735416564627516,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.2102441770440685,
+ "joint_099_factor_vs_raw_float32_with_datrie": 4.697312681798634
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 2802233,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 2679520,
+ "joint_099_backend_with_keys_bytes": 2040209,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2385916134232997,
+ "joint_099_saving_vs_raw_encoded": 0.27193456075922307,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.3735029107312045,
+ "joint_099_weighted_relative_frobenius": 0.006788218213522045,
+ "joint_099_added_weighted_relative_frobenius": 0.006788218213522045,
+ "joint_099_max_relative_l2": 0.14137685298673708,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1413276576532998,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.24110004291680398,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.3786772653206654
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 1436986,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1401091,
+ "joint_099_backend_with_keys_bytes": 1081734,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2279345167444513,
+ "joint_099_saving_vs_raw_encoded": 0.24722022343989436,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.5905009919259263,
+ "joint_099_weighted_relative_frobenius": 0.006792836185212061,
+ "joint_099_added_weighted_relative_frobenius": 0.0067896740613937155,
+ "joint_099_max_relative_l2": 0.14139286698093315,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.037615823600401,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.23256188224259455,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.632575038714967
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 775737,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 775737,
+ "joint_099_backend_with_keys_bytes": 621331,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.19904426371308837,
+ "joint_099_saving_vs_raw_encoded": 0.19904426371308837,
+ "joint_099_factor_vs_raw_float32_with_keys": 4.510048589238265,
+ "joint_099_weighted_relative_frobenius": 0.00940051238979808,
+ "joint_099_added_weighted_relative_frobenius": 0.006914346925094502,
+ "joint_099_max_relative_l2": 0.1417056444563921,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0778523599944267,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.20646405271581947,
+ "joint_099_factor_vs_raw_float32_with_datrie": 4.674936347107521
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 7077841,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 6685716,
+ "joint_099_backend_with_keys_bytes": 4301281,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.356646169236025,
+ "joint_099_saving_vs_raw_encoded": 0.3922891175430474,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.6455193231969731,
+ "joint_099_weighted_relative_frobenius": 0.009587863324407826,
+ "joint_099_added_weighted_relative_frobenius": 0.009587863324407826,
+ "joint_099_max_relative_l2": 0.14211742856778714,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.149934767889287,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.3618064870507832,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.6601544250075786
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 3641298,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 3452287,
+ "joint_099_backend_with_keys_bytes": 2308296,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.33137192823192274,
+ "joint_099_saving_vs_raw_encoded": 0.36607879937319054,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.0662622991158845,
+ "joint_099_weighted_relative_frobenius": 0.009591286480164959,
+ "joint_099_added_weighted_relative_frobenius": 0.009589247498575619,
+ "joint_099_max_relative_l2": 0.14211255515914975,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.036509090909091,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.3407847822907293,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.155297929451318
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 1976766,
+ "best_encoded_baseline": "int8-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1860550,
+ "joint_099_backend_with_keys_bytes": 1334408,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2827884227782107,
+ "joint_099_saving_vs_raw_encoded": 0.32495399050772833,
+ "joint_099_factor_vs_raw_float32_with_keys": 5.304105640853472,
+ "joint_099_weighted_relative_frobenius": 0.011525639114851247,
+ "joint_099_added_weighted_relative_frobenius": 0.009713811108145775,
+ "joint_099_max_relative_l2": 0.14306266325996833,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0692774268035872,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.29806468864045543,
+ "joint_099_factor_vs_raw_float32_with_datrie": 5.63534460216359
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 7077841,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 6687782,
+ "joint_099_backend_with_keys_bytes": 4285923,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.3591413416286595,
+ "joint_099_saving_vs_raw_encoded": 0.3944589882705757,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.6514158093834164,
+ "joint_099_weighted_relative_frobenius": 0.009429837428214146,
+ "joint_099_added_weighted_relative_frobenius": 0.009429837428214146,
+ "joint_099_max_relative_l2": 0.14236738964640794,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1489232585983502,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.3643361336175788,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.6662387213949807
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 3641298,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 3453902,
+ "joint_099_backend_with_keys_bytes": 2300988,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.3338004378815612,
+ "joint_099_saving_vs_raw_encoded": 0.36808577600624837,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.0760008309474016,
+ "joint_099_weighted_relative_frobenius": 0.009429329011061569,
+ "joint_099_added_weighted_relative_frobenius": 0.009427344865478818,
+ "joint_099_max_relative_l2": 0.1423848947835388,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0368272727272727,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.34327771601163126,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.165752491802803
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 1976766,
+ "best_encoded_baseline": "int8-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1856058,
+ "joint_099_backend_with_keys_bytes": 1326283,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.28543019668566394,
+ "joint_099_saving_vs_raw_encoded": 0.32906423926757133,
+ "joint_099_factor_vs_raw_float32_with_keys": 5.336599353230042,
+ "joint_099_weighted_relative_frobenius": 0.011437149201513133,
+ "joint_099_added_weighted_relative_frobenius": 0.00957913732749358,
+ "joint_099_max_relative_l2": 0.14274862236861094,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.078121273428569,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.30088850924233634,
+ "joint_099_factor_vs_raw_float32_with_datrie": 5.6725419135334585
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 7077841,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 6688172,
+ "joint_099_backend_with_keys_bytes": 4305324,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.3562779186898901,
+ "joint_099_saving_vs_raw_encoded": 0.3917178981556664,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.6439740655987796,
+ "joint_099_weighted_relative_frobenius": 0.009517378602255747,
+ "joint_099_added_weighted_relative_frobenius": 0.009517378602255747,
+ "joint_099_max_relative_l2": 0.14199628030247896,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1293635966132687,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.3614309879116906,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.6585601125709268
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 3641298,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 3454732,
+ "joint_099_backend_with_keys_bytes": 2310183,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.33129892564748875,
+ "joint_099_saving_vs_raw_encoded": 0.36556057757426064,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.0637577196265404,
+ "joint_099_weighted_relative_frobenius": 0.009520497179207372,
+ "joint_099_added_weighted_relative_frobenius": 0.009518618838387841,
+ "joint_099_max_relative_l2": 0.14199922735932338,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0364142927742688,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.3407028567210101,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.1526096620638993
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 1976766,
+ "best_encoded_baseline": "int8-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1849681,
+ "joint_099_backend_with_keys_bytes": 1330314,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2807873357622206,
+ "joint_099_saving_vs_raw_encoded": 0.32702505000591875,
+ "joint_099_factor_vs_raw_float32_with_keys": 5.320428861156088,
+ "joint_099_weighted_relative_frobenius": 0.012055645188533535,
+ "joint_099_added_weighted_relative_frobenius": 0.00963143524074153,
+ "joint_099_max_relative_l2": 0.14283478277746048,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0735416564627516,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.29604947771935075,
+ "joint_099_factor_vs_raw_float32_with_datrie": 5.6540262907726415
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 7454867,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 7029964,
+ "joint_099_backend_with_keys_bytes": 4527126,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.35602429827521165,
+ "joint_099_saving_vs_raw_encoded": 0.39272880388073994,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.6467107387777589,
+ "joint_099_weighted_relative_frobenius": 0.009597568466419546,
+ "joint_099_added_weighted_relative_frobenius": 0.009597568466419546,
+ "joint_099_max_relative_l2": 0.14213808847272216,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1747773128392094,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.36141031635485366,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.6620313588219864
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 3837076,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 3626449,
+ "joint_099_backend_with_keys_bytes": 2428020,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.3304690070093361,
+ "joint_099_saving_vs_raw_encoded": 0.3672212903784027,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.0703482673124602,
+ "joint_099_weighted_relative_frobenius": 0.009600233959965716,
+ "joint_099_added_weighted_relative_frobenius": 0.009598270767519452,
+ "joint_099_max_relative_l2": 0.14217809684457094,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0363361657654384,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.3403000781160599,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.1637096073007944
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 2084752,
+ "best_encoded_baseline": "int8-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1953224,
+ "joint_099_backend_with_keys_bytes": 1405422,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2804604080228381,
+ "joint_099_saving_vs_raw_encoded": 0.3258565047545223,
+ "joint_099_factor_vs_raw_float32_with_keys": 5.304361963879888,
+ "joint_099_weighted_relative_frobenius": 0.011273324565210031,
+ "joint_099_added_weighted_relative_frobenius": 0.009708818886474999,
+ "joint_099_max_relative_l2": 0.14267755652829675,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0778555181250178,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.29635620609178026,
+ "joint_099_factor_vs_raw_float32_with_datrie": 5.65107222816794
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 7454867,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 7030864,
+ "joint_099_backend_with_keys_bytes": 4511864,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.35827744641341375,
+ "joint_099_saving_vs_raw_encoded": 0.3947760570376373,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.6522809641425362,
+ "joint_099_weighted_relative_frobenius": 0.009467831744684692,
+ "joint_099_added_weighted_relative_frobenius": 0.009467831744684692,
+ "joint_099_max_relative_l2": 0.1422066988799605,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1662823511783362,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.3636968463339676,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.6677870562442678
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 3837076,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 3626389,
+ "joint_099_backend_with_keys_bytes": 2420451,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.3325451296041324,
+ "joint_099_saving_vs_raw_encoded": 0.3691938861779126,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.0799495631186087,
+ "joint_099_weighted_relative_frobenius": 0.009470324713106664,
+ "joint_099_added_weighted_relative_frobenius": 0.009468034050420173,
+ "joint_099_max_relative_l2": 0.1422308437423335,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0361909090909092,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.3424381315092502,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.1740504429574834
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 2084752,
+ "best_encoded_baseline": "int8-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1946680,
+ "joint_099_backend_with_keys_bytes": 1395550,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2831127868987199,
+ "joint_099_saving_vs_raw_encoded": 0.3305918401805107,
+ "joint_099_factor_vs_raw_float32_with_keys": 5.341884561642363,
+ "joint_099_weighted_relative_frobenius": 0.011592827136221024,
+ "joint_099_added_weighted_relative_frobenius": 0.009600591901907743,
+ "joint_099_max_relative_l2": 0.14200820684716933,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0753869892808483,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.29921592430482635,
+ "joint_099_factor_vs_raw_float32_with_datrie": 5.694291996182165
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 7454867,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 7032071,
+ "joint_099_backend_with_keys_bytes": 4532540,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.3554473497210139,
+ "joint_099_saving_vs_raw_encoded": 0.39200256691366864,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.6447437860449108,
+ "joint_099_weighted_relative_frobenius": 0.009554490322035257,
+ "joint_099_added_weighted_relative_frobenius": 0.009554490322035257,
+ "joint_099_max_relative_l2": 0.1419349417581233,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1280072991532342,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.3608230040398106,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.6599991327470642
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 3837076,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 3628003,
+ "joint_099_backend_with_keys_bytes": 2430195,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.33015628708135025,
+ "joint_099_saving_vs_raw_encoded": 0.3666544525049803,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.0676003365985034,
+ "joint_099_weighted_relative_frobenius": 0.009560715856289807,
+ "joint_099_added_weighted_relative_frobenius": 0.009558478461585218,
+ "joint_099_max_relative_l2": 0.1419611647297449,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.036754915059693,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.33997372302799955,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.1607505539838026
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 2084752,
+ "best_encoded_baseline": "int8-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1944152,
+ "joint_099_backend_with_keys_bytes": 1402010,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2788578259313058,
+ "joint_099_saving_vs_raw_encoded": 0.32749315026439596,
+ "joint_099_factor_vs_raw_float32_with_keys": 5.3172709181817535,
+ "joint_099_weighted_relative_frobenius": 0.011588159396950687,
+ "joint_099_added_weighted_relative_frobenius": 0.00966408415736297,
+ "joint_099_max_relative_l2": 0.142252491783661,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0752103506868385,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.294740745009476,
+ "joint_099_factor_vs_raw_float32_with_datrie": 5.665935629688787
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 3173089,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 3071670,
+ "joint_099_backend_with_keys_bytes": 2895339,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.05740558067761181,
+ "joint_099_saving_vs_raw_encoded": 0.0875330001774296,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.0959300448064977,
+ "joint_099_weighted_relative_frobenius": 0.0052940877733579465,
+ "joint_099_added_weighted_relative_frobenius": 0.0052940877733579465,
+ "joint_099_max_relative_l2": 0.14017569822550982,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.174781279807998,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.05867480804265901,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.0981832475872195,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 2893447,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.610478337983415,
+ "character_perplexity": 5.005204834898665,
+ "next_character_accuracy": 0.5303227785018784,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00012570950190301332,
+ "mean_prediction_kl_nats": 3.522476674748964e-05,
+ "prediction_kl_p99_nats": 3.073411681134878e-05,
+ "maximum_prediction_kl_nats": 0.069073305132374,
+ "top1_changed_fraction": 0.0007666947788085563,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6207461357116699,
+ "maximum_per_target_cross_entropy_increase": 0.810072660446167
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -4.4151146362623543e-05,
+ 0.00031408959841323583
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.058021532260952524,
+ "selected_factor_vs_raw_float32_with_keys": 1.0966466639962646,
+ "selected_full_model_factor_vs_raw_float32": 1.087713404679165
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 1667810,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1654552,
+ "joint_099_backend_with_keys_bytes": 1567320,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.05272242879039157,
+ "joint_099_saving_vs_raw_encoded": 0.06025266667066398,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.0245316846591632,
+ "joint_099_weighted_relative_frobenius": 0.005302751473250229,
+ "joint_099_added_weighted_relative_frobenius": 0.005299381939658822,
+ "joint_099_max_relative_l2": 0.14016502249394194,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0618789300797746,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.054928288836961214,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.069888565003748,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 1566355,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6104803438537196,
+ "character_perplexity": 5.0052148747004805,
+ "next_character_accuracy": 0.5303227785018784,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00012771562637668983,
+ "mean_prediction_kl_nats": 3.5253858489440136e-05,
+ "prediction_kl_p99_nats": 3.0714668360973565e-05,
+ "maximum_prediction_kl_nats": 0.0691435144903208,
+ "top1_changed_fraction": 0.0007666947788085563,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6211075782775879,
+ "maximum_per_target_cross_entropy_increase": 0.8102507591247559
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -4.18414171096396e-05,
+ 0.00031688913937407293
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.05330566824131244,
+ "selected_factor_vs_raw_float32_with_keys": 2.0257789581544414,
+ "selected_full_model_factor_vs_raw_float32": 1.8633522850920534
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 3,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 962252,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 962252,
+ "joint_099_backend_with_keys_bytes": 1006310,
+ "joint_099_added_saving_vs_best_encoded_baseline": -0.04578634287068262,
+ "joint_099_saving_vs_raw_encoded": -0.04578634287068262,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.153192356232175,
+ "joint_099_weighted_relative_frobenius": 0.007636228340675911,
+ "joint_099_added_weighted_relative_frobenius": 0.005520060633454694,
+ "joint_099_max_relative_l2": 0.12213938394952205,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1206418242604694,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": -0.04918246899164669,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.305415139408319,
+ "validation_selection": {
+ "feasible": true,
+ "name": "int8-raw",
+ "backend_with_key_registry_bytes": 962252,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6103997491236468,
+ "character_perplexity": 5.004811497013974,
+ "next_character_accuracy": 0.5307061258912826,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 4.711385118005514e-05,
+ "mean_prediction_kl_nats": 3.3553609696799543e-06,
+ "prediction_kl_p99_nats": 2.355900870385665e-05,
+ "maximum_prediction_kl_nats": 7.72629051648873e-05,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": 0.00038334738940426494,
+ "maximum_absolute_logit_change": 0.029384613037109375,
+ "maximum_per_target_cross_entropy_increase": 0.021229267120361328
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 8.000077854203593e-07,
+ 9.973376735015557e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.0,
+ "selected_factor_vs_raw_float32_with_keys": 3.2975655025918367,
+ "selected_full_model_factor_vs_raw_float32": 2.7589069627937097
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 3173089,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 3071649,
+ "joint_099_backend_with_keys_bytes": 2895906,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.057214545021257335,
+ "joint_099_saving_vs_raw_encoded": 0.08735430994844451,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.095715468665074,
+ "joint_099_weighted_relative_frobenius": 0.005299696904874133,
+ "joint_099_added_weighted_relative_frobenius": 0.005299696904874133,
+ "joint_099_max_relative_l2": 0.11643283329987608,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1615017998270556,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.05847955746099098,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.0979631809733372,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp32-joint-0.95",
+ "backend_with_key_registry_bytes": 2894980,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6159370315575023,
+ "character_perplexity": 5.0326014211237675,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00041829374079927037,
+ "mean_prediction_kl_nats": 0.0001224659868776353,
+ "prediction_kl_p99_nats": 6.492612482710993e-05,
+ "maximum_prediction_kl_nats": 0.049823547914828695,
+ "top1_changed_fraction": 0.0002300084336425669,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6587400436401367,
+ "maximum_per_target_cross_entropy_increase": 0.3685579299926758
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 9.471540831570154e-05,
+ 0.0007101388272276771
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.05751601175785381,
+ "selected_factor_vs_raw_float32_with_keys": 1.0960659486421322,
+ "selected_full_model_factor_vs_raw_float32": 1.0871906337528756
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 1667810,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1654548,
+ "joint_099_backend_with_keys_bytes": 1567759,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.05245480941018332,
+ "joint_099_saving_vs_raw_encoded": 0.05998944723919386,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.023964780301054,
+ "joint_099_weighted_relative_frobenius": 0.0053040528702744776,
+ "joint_099_added_weighted_relative_frobenius": 0.005301355563665608,
+ "joint_099_max_relative_l2": 0.1164297624166434,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0363339928189348,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.054649478025039944,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.0692833078223476,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp16-joint-0.95",
+ "backend_with_key_registry_bytes": 1567336,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6159358449097074,
+ "character_perplexity": 5.032595449201931,
+ "next_character_accuracy": 0.5292494058115464,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.0004171065973414183,
+ "mean_prediction_kl_nats": 0.0001224485145529409,
+ "prediction_kl_p99_nats": 6.468627058898175e-05,
+ "maximum_prediction_kl_nats": 0.049807480276882234,
+ "top1_changed_fraction": 0.0002300084336425669,
+ "accuracy_change": 0.0,
+ "maximum_absolute_logit_change": 0.6586532592773438,
+ "maximum_per_target_cross_entropy_increase": 0.3685433864593506
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 9.510922932824986e-05,
+ 0.0007089831271132332
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.052710468357521245,
+ "selected_factor_vs_raw_float32_with_keys": 2.024511017420642,
+ "selected_full_model_factor_vs_raw_float32": 1.862370584235847
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 7,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 962252,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 962252,
+ "joint_099_backend_with_keys_bytes": 1005997,
+ "joint_099_added_saving_vs_best_encoded_baseline": -0.045461064253438765,
+ "joint_099_saving_vs_raw_encoded": -0.045461064253438765,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.154173421988336,
+ "joint_099_weighted_relative_frobenius": 0.007637971226019644,
+ "joint_099_added_weighted_relative_frobenius": 0.005479980336186081,
+ "joint_099_max_relative_l2": 0.11602073490844073,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1232070236978005,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": -0.0488330633719094,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.3065162971288444,
+ "validation_selection": {
+ "feasible": true,
+ "name": "int8-raw",
+ "backend_with_key_registry_bytes": 962252,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.615536524896122,
+ "character_perplexity": 5.0305862343053445,
+ "next_character_accuracy": 0.5290960668557847,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 1.7699775619881114e-05,
+ "mean_prediction_kl_nats": 3.5440430999465437e-06,
+ "prediction_kl_p99_nats": 2.424724186036613e-05,
+ "maximum_prediction_kl_nats": 0.00011723431737023591,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": -0.00015333895576175038,
+ "maximum_absolute_logit_change": 0.02951216697692871,
+ "maximum_per_target_cross_entropy_increase": 0.0291440486907959
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -3.6594241267184036e-05,
+ 7.115336931680388e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.0,
+ "selected_factor_vs_raw_float32_with_keys": 3.2975655025918367,
+ "selected_full_model_factor_vs_raw_float32": 2.7589069627937097
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 3173089,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 3071342,
+ "joint_099_backend_with_keys_bytes": 2894750,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.057496690371830916,
+ "joint_099_saving_vs_raw_encoded": 0.08771862371335948,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.0961530356680196,
+ "joint_099_weighted_relative_frobenius": 0.005649047524990139,
+ "joint_099_added_weighted_relative_frobenius": 0.005649047524990139,
+ "joint_099_max_relative_l2": 0.11261831137802518,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.2134813021805575,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.05876807091890335,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.0984119463777775,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 2891408,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.605444559273432,
+ "character_perplexity": 4.98007304835536,
+ "next_character_accuracy": 0.5310128038028061,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -5.5385669269262635e-05,
+ "mean_prediction_kl_nats": 2.0394254485204406e-05,
+ "prediction_kl_p99_nats": 6.292058917393448e-05,
+ "maximum_prediction_kl_nats": 0.028937972454724202,
+ "top1_changed_fraction": 0.00038334738940427816,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.6746394634246826,
+ "maximum_per_target_cross_entropy_increase": 0.259132981300354
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00013110323318399718,
+ 2.8910119316644106e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.05858481406499183,
+ "selected_factor_vs_raw_float32_with_keys": 1.0974200112886179,
+ "selected_full_model_factor_vs_raw_float32": 1.0884095070641355
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 1667810,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 1654633,
+ "joint_099_backend_with_keys_bytes": 1567446,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.05269265148223201,
+ "joint_099_saving_vs_raw_encoded": 0.06017711849671126,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.0243689415775727,
+ "joint_099_weighted_relative_frobenius": 0.005608434419236457,
+ "joint_099_added_weighted_relative_frobenius": 0.005604080214656733,
+ "joint_099_max_relative_l2": 0.11263148326372621,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0635481057016951,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.05489715323374811,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.069714810316582,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 1565742,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6054448741033316,
+ "character_perplexity": 4.980074616231503,
+ "next_character_accuracy": 0.5310128038028061,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -5.5070856757365275e-05,
+ "mean_prediction_kl_nats": 2.0374639076691894e-05,
+ "prediction_kl_p99_nats": 6.243811826060971e-05,
+ "maximum_prediction_kl_nats": 0.028911209971021787,
+ "top1_changed_fraction": 0.00038334738940427816,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.67431640625,
+ "maximum_per_target_cross_entropy_increase": 0.2590094804763794
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00012999504614585823,
+ 2.8694272122474748e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.05372248710136929,
+ "selected_factor_vs_raw_float32_with_keys": 2.0265720661513837,
+ "selected_full_model_factor_vs_raw_float32": 1.8639662486266602
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "seed": 11,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 962252,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 962252,
+ "joint_099_backend_with_keys_bytes": 1005999,
+ "joint_099_added_saving_vs_best_encoded_baseline": -0.045463142711057,
+ "joint_099_saving_vs_raw_encoded": -0.045463142711057,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.154167151259594,
+ "joint_099_weighted_relative_frobenius": 0.007915395391086759,
+ "joint_099_added_weighted_relative_frobenius": 0.005829303558166024,
+ "joint_099_max_relative_l2": 0.11380766731945188,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1052328226907326,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": -0.04883529599567771,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.3065092586482523,
+ "validation_selection": {
+ "feasible": true,
+ "name": "int8-raw",
+ "backend_with_key_registry_bytes": 962252,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.6055288730005035,
+ "character_perplexity": 4.980492954576834,
+ "next_character_accuracy": 0.5311661427585678,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 2.8926942335284213e-05,
+ "mean_prediction_kl_nats": 3.594473232078083e-06,
+ "prediction_kl_p99_nats": 2.5250991242415894e-05,
+ "maximum_prediction_kl_nats": 9.39906069725258e-05,
+ "top1_changed_fraction": 0.0011500421682128345,
+ "accuracy_change": 7.666947788076417e-05,
+ "maximum_absolute_logit_change": 0.03656506538391113,
+ "maximum_per_target_cross_entropy_increase": 0.019799232482910156
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -2.5258228333548203e-05,
+ 8.626173798269781e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.0,
+ "selected_factor_vs_raw_float32_with_keys": 3.2975655025918367,
+ "selected_full_model_factor_vs_raw_float32": 2.7589069627937097
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 10013493,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 9681520,
+ "joint_099_backend_with_keys_bytes": 7440315,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.23149309199381918,
+ "joint_099_saving_vs_raw_encoded": 0.25697106893668376,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.345842615534423,
+ "joint_099_weighted_relative_frobenius": 0.011433905190859801,
+ "joint_099_added_weighted_relative_frobenius": 0.011433905190859801,
+ "joint_099_max_relative_l2": 0.1407557896887402,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1877255184964386,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.23896095504842274,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.3605024773899466,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp32-joint-0.95",
+ "backend_with_key_registry_bytes": 7430160,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.733085906821203,
+ "character_perplexity": 5.658087325916751,
+ "next_character_accuracy": 0.5241892202714099,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00010423505069745625,
+ "mean_prediction_kl_nats": 0.00013875969726630407,
+ "prediction_kl_p99_nats": 0.00038196038107028794,
+ "maximum_prediction_kl_nats": 0.09296217712975313,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": 7.666947788076417e-05,
+ "maximum_absolute_logit_change": 1.504372239112854,
+ "maximum_per_target_cross_entropy_increase": 0.5252523422241211
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00022564232065017386,
+ 0.0003975916668243442
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.23254199753757676,
+ "selected_factor_vs_raw_float32_with_keys": 1.3476820149229627,
+ "selected_full_model_factor_vs_raw_float32": 1.3177469613751664
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 5305910,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 5248623,
+ "joint_099_backend_with_keys_bytes": 4141097,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.21101267894455367,
+ "joint_099_saving_vs_raw_encoded": 0.21953123969309696,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.4180773838429768,
+ "joint_099_weighted_relative_frobenius": 0.011436593582149768,
+ "joint_099_added_weighted_relative_frobenius": 0.011436043946536932,
+ "joint_099_max_relative_l2": 0.1407469103080355,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0654486678895083,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.22392076767335312,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.5298530481412707,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp16-joint-0.95",
+ "backend_with_key_registry_bytes": 4136212,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.733086864041349,
+ "character_perplexity": 5.658092741954519,
+ "next_character_accuracy": 0.5241892202714099,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00010519237107762973,
+ "mean_prediction_kl_nats": 0.0001388013511152666,
+ "prediction_kl_p99_nats": 0.0003827823135934586,
+ "maximum_prediction_kl_nats": 0.09302745751300914,
+ "top1_changed_fraction": 0.0010733726903319788,
+ "accuracy_change": 7.666947788076417e-05,
+ "maximum_absolute_logit_change": 1.5042415857315063,
+ "maximum_per_target_cross_entropy_increase": 0.5253055095672607
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00022603695046941323,
+ 0.00039856463735664214
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.21194339924966987,
+ "selected_factor_vs_raw_float32_with_keys": 2.4209332113537703,
+ "selected_full_model_factor_vs_raw_float32": 2.2152658796867795
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 3,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 3099273,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 3099273,
+ "joint_099_backend_with_keys_bytes": 2760694,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.10924465189094346,
+ "joint_099_saving_vs_raw_encoded": 0.10924465189094346,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.627165125870524,
+ "joint_099_weighted_relative_frobenius": 0.01266636913425628,
+ "joint_099_added_weighted_relative_frobenius": 0.011752570284712499,
+ "joint_099_max_relative_l2": 0.13922942015502465,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1508901824706497,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.12106323425508236,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.950531561961863,
+ "validation_selection": {
+ "feasible": true,
+ "name": "int8-joint-0.95",
+ "backend_with_key_registry_bytes": 2757967,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7330446700904785,
+ "character_perplexity": 5.657854009703907,
+ "next_character_accuracy": 0.5242658897492908,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 6.299487197111428e-05,
+ "mean_prediction_kl_nats": 0.00013896683904643248,
+ "prediction_kl_p99_nats": 0.00035191132917633707,
+ "maximum_prediction_kl_nats": 0.09188349093556392,
+ "top1_changed_fraction": 0.0017633979912596795,
+ "accuracy_change": 0.00015333895576163936,
+ "maximum_absolute_logit_change": 1.4925585985183716,
+ "maximum_per_target_cross_entropy_increase": 0.5185873508453369
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0002516761914661864,
+ 0.00035003994847103626
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.11012453565723312,
+ "selected_factor_vs_raw_float32_with_keys": 3.630751564467595,
+ "selected_full_model_factor_vs_raw_float32": 3.0982068933606906
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 10013493,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 9681369,
+ "joint_099_backend_with_keys_bytes": 7432884,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.23224866235343367,
+ "joint_099_saving_vs_raw_encoded": 0.2577131676229264,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.347188117021603,
+ "joint_099_weighted_relative_frobenius": 0.011233518943283134,
+ "joint_099_added_weighted_relative_frobenius": 0.011233518943283134,
+ "joint_099_max_relative_l2": 0.1398521329039012,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1570228240486526,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.23974102039406286,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.3619203505928132,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp32-joint-0.97",
+ "backend_with_key_registry_bytes": 7425813,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7263920138538618,
+ "character_perplexity": 5.620339176943785,
+ "next_character_accuracy": 0.5251859234838611,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -9.943168672067081e-05,
+ "mean_prediction_kl_nats": 3.599514161387495e-05,
+ "prediction_kl_p99_nats": 0.00012654597219580023,
+ "maximum_prediction_kl_nats": 0.04607604998831831,
+ "top1_changed_fraction": 0.0013033811239745456,
+ "accuracy_change": 7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.7387614250183105,
+ "maximum_per_target_cross_entropy_increase": 0.36521947383880615
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00019782683373067314,
+ 4.029123812401072e-06
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.2329790342667447,
+ "selected_factor_vs_raw_float32_with_keys": 1.3484709351016515,
+ "selected_full_model_factor_vs_raw_float32": 1.3184519059082225
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 5305910,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 5249177,
+ "joint_099_backend_with_keys_bytes": 4137706,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.21174195497694215,
+ "joint_099_saving_vs_raw_encoded": 0.22017033835854738,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.4200590858799536,
+ "joint_099_weighted_relative_frobenius": 0.011234116695951897,
+ "joint_099_added_weighted_relative_frobenius": 0.011232898757745064,
+ "joint_099_max_relative_l2": 0.139832944885907,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0567026305344063,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.22469320440478902,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.532089920980823,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp16-joint-0.97",
+ "backend_with_key_registry_bytes": 4134227,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.726390375275437,
+ "character_perplexity": 5.620329967584816,
+ "next_character_accuracy": 0.5251859234838611,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -0.00010107010087641331,
+ "mean_prediction_kl_nats": 3.6004410187967044e-05,
+ "prediction_kl_p99_nats": 0.00012695645761842848,
+ "maximum_prediction_kl_nats": 0.04610446029724359,
+ "top1_changed_fraction": 0.0013033811239745456,
+ "accuracy_change": 7.666947788087519e-05,
+ "maximum_absolute_logit_change": 0.739013671875,
+ "maximum_per_target_cross_entropy_increase": 0.36529219150543213
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0001983225788790956,
+ 1.5536949280217424e-06
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.21240472554078482,
+ "selected_factor_vs_raw_float32_with_keys": 2.4220955936865587,
+ "selected_full_model_factor_vs_raw_float32": 2.2161754986580062
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 7,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 3099273,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 3099273,
+ "joint_099_backend_with_keys_bytes": 2758324,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.1100093473533954,
+ "joint_099_saving_vs_raw_encoded": 0.1100093473533954,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.630281649291381,
+ "joint_099_weighted_relative_frobenius": 0.012475970273467448,
+ "joint_099_added_weighted_relative_frobenius": 0.011543854934588026,
+ "joint_099_max_relative_l2": 0.13970731928662025,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1356318727090968,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.12191065794404288,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.954344128484711,
+ "validation_selection": {
+ "feasible": true,
+ "name": "int8-joint-0.97",
+ "backend_with_key_registry_bytes": 2756386,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7263646895656914,
+ "character_perplexity": 5.6201856072746,
+ "next_character_accuracy": 0.5252625929617419,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -0.00012675288472309662,
+ "mean_prediction_kl_nats": 3.98149611035077e-05,
+ "prediction_kl_p99_nats": 0.0001327648061789036,
+ "maximum_prediction_kl_nats": 0.045853842864706854,
+ "top1_changed_fraction": 0.0018400674691405351,
+ "accuracy_change": 0.00015333895576175038,
+ "maximum_absolute_logit_change": 0.7336316108703613,
+ "maximum_per_target_cross_entropy_increase": 0.36336004734039307
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0002373574082006347,
+ -1.4517555775323528e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.11063465528851446,
+ "selected_factor_vs_raw_float32_with_keys": 3.632834080567816,
+ "selected_full_model_factor_vs_raw_float32": 3.09962405797504
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 10013493,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 9680528,
+ "joint_099_backend_with_keys_bytes": 7437903,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.23166350017271786,
+ "joint_099_saving_vs_raw_encoded": 0.2572119439240632,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.3462790520392642,
+ "joint_099_weighted_relative_frobenius": 0.011360509821225311,
+ "joint_099_added_weighted_relative_frobenius": 0.011360509821225311,
+ "joint_099_max_relative_l2": 0.14115591128178223,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1783718155366227,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.23913765104952922,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.360962375734758,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 7422143,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7281426950457552,
+ "character_perplexity": 5.630187216906834,
+ "next_character_accuracy": 0.5290193973779038,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 7.878053941357795e-05,
+ "mean_prediction_kl_nats": 7.239235945488115e-05,
+ "prediction_kl_p99_nats": 0.00025523719050292914,
+ "maximum_prediction_kl_nats": 0.15028831154148714,
+ "top1_changed_fraction": 0.0007666947788085563,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 1.3504877090454102,
+ "maximum_per_target_cross_entropy_increase": 0.8670740127563477
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00012532674483123354,
+ 0.000296544972959154
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.23329151054570574,
+ "selected_factor_vs_raw_float32_with_keys": 1.3491377086105725,
+ "selected_full_model_factor_vs_raw_float32": 1.3190476501208241
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 5305910,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 5249238,
+ "joint_099_backend_with_keys_bytes": 4140890,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.21114455088529038,
+ "joint_099_saving_vs_raw_encoded": 0.219570252793583,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.418198261726344,
+ "joint_099_weighted_relative_frobenius": 0.011362454733114946,
+ "joint_099_added_weighted_relative_frobenius": 0.011360453108480639,
+ "joint_099_max_relative_l2": 0.1411682070021142,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0613761148069605,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.224059100685167,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.5299894824023683,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 4133142,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7281424986147633,
+ "character_perplexity": 5.630186110963683,
+ "next_character_accuracy": 0.5290193973779038,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 7.858409296605416e-05,
+ "mean_prediction_kl_nats": 7.239406989054103e-05,
+ "prediction_kl_p99_nats": 0.0002553027836586025,
+ "maximum_prediction_kl_nats": 0.15030648367116142,
+ "top1_changed_fraction": 0.0007666947788085563,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 1.349926471710205,
+ "maximum_per_target_cross_entropy_increase": 0.867246150970459
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0001246097273672068,
+ 0.00029588735143539364
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.21262057464340534,
+ "selected_factor_vs_raw_float32_with_keys": 2.4227314232126553,
+ "selected_full_model_factor_vs_raw_float32": 2.216673011748056
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "seed": 11,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 3099273,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 3099273,
+ "joint_099_backend_with_keys_bytes": 2760444,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.1093253159692612,
+ "joint_099_saving_vs_raw_encoded": 0.1093253159692612,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.62749362059147,
+ "joint_099_weighted_relative_frobenius": 0.012504361072079999,
+ "joint_099_added_weighted_relative_frobenius": 0.011554937193940731,
+ "joint_099_max_relative_l2": 0.1411733068123005,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1320600286985667,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.1211526249395719,
+ "joint_099_factor_vs_raw_float32_with_datrie": 3.950933384542714,
+ "validation_selection": {
+ "feasible": true,
+ "name": "int8-joint-0.9",
+ "backend_with_key_registry_bytes": 2756191,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7281543275316205,
+ "character_perplexity": 5.6302527103609785,
+ "next_character_accuracy": 0.5291727363336656,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 9.041400935520268e-05,
+ "mean_prediction_kl_nats": 9.501836386167312e-05,
+ "prediction_kl_p99_nats": 0.0002800923076272734,
+ "maximum_prediction_kl_nats": 0.14853157432377162,
+ "top1_changed_fraction": 0.0024534232921873803,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 1.359821081161499,
+ "maximum_per_target_cross_entropy_increase": 0.8680667877197266
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00016767168309531454,
+ 0.00033279514032229094
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.11069757326960228,
+ "selected_factor_vs_raw_float32_with_keys": 3.63309110290252,
+ "selected_full_model_factor_vs_raw_float32": 3.0997989403935953
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 11783824,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 11391663,
+ "joint_099_backend_with_keys_bytes": 8211066,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.27920392307953634,
+ "joint_099_saving_vs_raw_encoded": 0.3031917313089537,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.4351150021203094,
+ "joint_099_weighted_relative_frobenius": 0.012203466418320411,
+ "joint_099_added_weighted_relative_frobenius": 0.012203466418320411,
+ "joint_099_max_relative_l2": 0.14041193528604085,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1594861617907517,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.28910008220535777,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.456809057629733,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 8191912,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467407476364325,
+ "character_perplexity": 5.735877505237137,
+ "next_character_accuracy": 0.522579161235912,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 8.545016806342631e-05,
+ "mean_prediction_kl_nats": 7.555171199507978e-05,
+ "prediction_kl_p99_nats": 0.0002332713868702684,
+ "maximum_prediction_kl_nats": 0.07069631943714265,
+ "top1_changed_fraction": 0.0017633979912596795,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 1.421870231628418,
+ "maximum_per_target_cross_entropy_increase": 0.5875310897827148
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00011373525728950033,
+ 0.0002831974613994154
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.2808853281562139,
+ "selected_factor_vs_raw_float32_with_keys": 1.4384705304451513,
+ "selected_full_model_factor_vs_raw_float32": 1.398220381356355
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 6260368,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 6192529,
+ "joint_099_backend_with_keys_bytes": 4619676,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2539920281358392,
+ "joint_099_saving_vs_raw_encoded": 0.2620759674191677,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.5507901419926418,
+ "joint_099_weighted_relative_frobenius": 0.012205143139227746,
+ "joint_099_added_weighted_relative_frobenius": 0.012203608015228266,
+ "joint_099_max_relative_l2": 0.1404184277793818,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.0572761848824968,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.271060882896614,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.693760522246224,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 4609985,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467389892458764,
+ "character_perplexity": 5.735867419333168,
+ "next_character_accuracy": 0.522579161235912,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 8.369162879872298e-05,
+ "mean_prediction_kl_nats": 7.55530360250437e-05,
+ "prediction_kl_p99_nats": 0.00023711060763044312,
+ "maximum_prediction_kl_nats": 0.07071264041728009,
+ "top1_changed_fraction": 0.0017633979912596795,
+ "accuracy_change": -0.00023000843364262558,
+ "maximum_absolute_logit_change": 1.4215927124023438,
+ "maximum_per_target_cross_entropy_increase": 0.5875155925750732
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00011714467232062939,
+ 0.0002800202357481195
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.25555697841705705,
+ "selected_factor_vs_raw_float32_with_keys": 2.5561523519057006,
+ "selected_full_model_factor_vs_raw_float32": 2.3192095304453875
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 3,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 3671291,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 3671291,
+ "joint_099_backend_with_keys_bytes": 3116940,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.15099620269817893,
+ "joint_099_saving_vs_raw_encoded": 0.15099620269817893,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.780574537848018,
+ "joint_099_weighted_relative_frobenius": 0.0133093761146089,
+ "joint_099_added_weighted_relative_frobenius": 0.01244834613064305,
+ "joint_099_max_relative_l2": 0.14034170133988702,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1160896130346232,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.16894022693140376,
+ "joint_099_factor_vs_raw_float32_with_datrie": 4.178183442348404,
+ "validation_selection": {
+ "feasible": true,
+ "name": "int8-joint-0.9",
+ "backend_with_key_registry_bytes": 3111386,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7467207411986971,
+ "character_perplexity": 5.735762751908877,
+ "next_character_accuracy": 0.5227325001916737,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 6.544222091906036e-05,
+ "mean_prediction_kl_nats": 7.611414126180771e-05,
+ "prediction_kl_p99_nats": 0.00023912776528949614,
+ "maximum_prediction_kl_nats": 0.07220868673265157,
+ "top1_changed_fraction": 0.002300084336425669,
+ "accuracy_change": -7.666947788087519e-05,
+ "maximum_absolute_logit_change": 1.4234991073608398,
+ "maximum_per_target_cross_entropy_increase": 0.597667932510376
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00014768831565003417,
+ 0.00027000310437506474
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.15250902203066985,
+ "selected_factor_vs_raw_float32_with_keys": 3.7873230772395323,
+ "selected_full_model_factor_vs_raw_float32": 3.2014705852488614
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 11783824,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 11389143,
+ "joint_099_backend_with_keys_bytes": 8203394,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.27971806131506116,
+ "joint_099_saving_vs_raw_encoded": 0.3038427933071641,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.4364571541973967,
+ "joint_099_weighted_relative_frobenius": 0.012525398130393038,
+ "joint_099_added_weighted_relative_frobenius": 0.012525398130393038,
+ "joint_099_max_relative_l2": 0.14116478475441507,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.191513651349131,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.2896347151191778,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.4582394940414902,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp32-joint-0.95",
+ "backend_with_key_registry_bytes": 8192404,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7468335463514388,
+ "character_perplexity": 5.736409811997366,
+ "next_character_accuracy": 0.5232691865368397,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 5.126287092804774e-05,
+ "mean_prediction_kl_nats": 4.523398162602008e-05,
+ "prediction_kl_p99_nats": 0.0002614981201286399,
+ "maximum_prediction_kl_nats": 0.031281574788080815,
+ "top1_changed_fraction": 0.0018400674691405351,
+ "accuracy_change": -0.00023000843364251455,
+ "maximum_absolute_logit_change": 0.6498794555664062,
+ "maximum_per_target_cross_entropy_increase": 0.5136287212371826
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00014062225521751923,
+ 0.0002640022597950617
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.28068301539457363,
+ "selected_factor_vs_raw_float32_with_keys": 1.4383841421883004,
+ "selected_full_model_factor_vs_raw_float32": 1.3981441181889676
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 6260368,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 6190190,
+ "joint_099_backend_with_keys_bytes": 4615224,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.254429347079815,
+ "joint_099_saving_vs_raw_encoded": 0.26278710772274094,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.5532507197917154,
+ "joint_099_weighted_relative_frobenius": 0.012524791627337375,
+ "joint_099_added_weighted_relative_frobenius": 0.012523978773244526,
+ "joint_099_max_relative_l2": 0.14112562218626845,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.057025491895454,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.2715344857103401,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.696598826538473,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp16-joint-0.95",
+ "backend_with_key_registry_bytes": 4609709,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.746831640034469,
+ "character_perplexity": 5.7363988765924185,
+ "next_character_accuracy": 0.5232691865368397,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 4.93564580521361e-05,
+ "mean_prediction_kl_nats": 4.522092181157221e-05,
+ "prediction_kl_p99_nats": 0.000260726444150113,
+ "maximum_prediction_kl_nats": 0.03128432153710157,
+ "top1_changed_fraction": 0.0018400674691405351,
+ "accuracy_change": -0.00023000843364251455,
+ "maximum_absolute_logit_change": 0.6495251655578613,
+ "maximum_per_target_cross_entropy_increase": 0.5136415958404541
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00014176603083662905,
+ 0.00026122043476795856
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.2553202728833849,
+ "selected_factor_vs_raw_float32_with_keys": 2.5563053980197012,
+ "selected_full_model_factor_vs_raw_float32": 2.31932724584094
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 7,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 3671291,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 3671291,
+ "joint_099_backend_with_keys_bytes": 3116568,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.15109752945217358,
+ "joint_099_saving_vs_raw_encoded": 0.15109752945217358,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.781025795041212,
+ "joint_099_weighted_relative_frobenius": 0.013529571057738162,
+ "joint_099_added_weighted_relative_frobenius": 0.01272182586954238,
+ "joint_099_max_relative_l2": 0.14050498208265808,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.128973194668229,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.1690535951122467,
+ "joint_099_factor_vs_raw_float32_with_datrie": 4.178753482790604,
+ "validation_selection": {
+ "feasible": true,
+ "name": "int8-joint-0.95",
+ "backend_with_key_registry_bytes": 3113298,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7468335912146888,
+ "character_perplexity": 5.7364100693513596,
+ "next_character_accuracy": 0.5231925170589589,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 5.130773647893945e-05,
+ "mean_prediction_kl_nats": 4.9148167713019224e-05,
+ "prediction_kl_p99_nats": 0.00026751634360011476,
+ "maximum_prediction_kl_nats": 0.030909915141433902,
+ "top1_changed_fraction": 0.0026067622479490913,
+ "accuracy_change": -0.00030667791152338975,
+ "maximum_absolute_logit_change": 0.652705192565918,
+ "maximum_per_target_cross_entropy_increase": 0.5167644023895264
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.00014600205117785967,
+ 0.0002798015539844717
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.1519882243058368,
+ "selected_factor_vs_raw_float32_with_keys": 3.784997131659096,
+ "selected_full_model_factor_vs_raw_float32": 3.1999174890695854
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp32",
+ "raw_encoded_backend_with_keys_bytes": 11783824,
+ "best_encoded_baseline": "fp32-dedup",
+ "best_encoded_baseline_with_keys_bytes": 11388281,
+ "joint_099_backend_with_keys_bytes": 8205579,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.2794716779468297,
+ "joint_099_saving_vs_raw_encoded": 0.30365736962805967,
+ "joint_099_factor_vs_raw_float32_with_keys": 1.4360746511611187,
+ "joint_099_weighted_relative_frobenius": 0.013052669069275194,
+ "joint_099_added_weighted_relative_frobenius": 0.013052669069275194,
+ "joint_099_max_relative_l2": 0.1400040347803519,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1577321344867262,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.28938037342746636,
+ "joint_099_factor_vs_raw_float32_with_datrie": 1.4578318170558695,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp32-joint-0.9",
+ "backend_with_key_registry_bytes": 8186195,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7491682414699585,
+ "character_perplexity": 5.749818226172276,
+ "next_character_accuracy": 0.5252625929617419,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -0.0002687747523804419,
+ "mean_prediction_kl_nats": 0.0001433063275818475,
+ "prediction_kl_p99_nats": 0.00047574067175172846,
+ "maximum_prediction_kl_nats": 0.14949254966897688,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": -0.00030667791152338975,
+ "maximum_absolute_logit_change": 1.5382030010223389,
+ "maximum_per_target_cross_entropy_increase": 0.7367382049560547
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0005444561415285015,
+ 1.527656112117525e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.28117377855358505,
+ "selected_factor_vs_raw_float32_with_keys": 1.439475116339154,
+ "selected_full_model_factor_vs_raw_float32": 1.3991071635586236
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "fp16",
+ "raw_encoded_backend_with_keys_bytes": 6260368,
+ "best_encoded_baseline": "fp16-dedup",
+ "best_encoded_baseline_with_keys_bytes": 6189032,
+ "joint_099_backend_with_keys_bytes": 4616234,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.254126655024566,
+ "joint_099_saving_vs_raw_encoded": 0.26262577535378107,
+ "joint_099_factor_vs_raw_float32_with_keys": 2.5526920862330638,
+ "joint_099_weighted_relative_frobenius": 0.013052639637527957,
+ "joint_099_added_weighted_relative_frobenius": 0.01305117528543782,
+ "joint_099_max_relative_l2": 0.13994770767903386,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.057732670135994,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.27121485544702306,
+ "joint_099_factor_vs_raw_float32_with_datrie": 2.6959543921177147,
+ "validation_selection": {
+ "feasible": true,
+ "name": "fp16-joint-0.9",
+ "backend_with_key_registry_bytes": 4606495,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7491683695294322,
+ "character_perplexity": 5.749818962491019,
+ "next_character_accuracy": 0.5252625929617419,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": -0.0002686467273177534,
+ "mean_prediction_kl_nats": 0.00014329286792632542,
+ "prediction_kl_p99_nats": 0.0004777807634440265,
+ "maximum_prediction_kl_nats": 0.1495226025186514,
+ "top1_changed_fraction": 0.0013800506018554013,
+ "accuracy_change": -0.00030667791152338975,
+ "maximum_absolute_logit_change": 1.538262128829956,
+ "maximum_per_target_cross_entropy_increase": 0.7369163036346436
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ -0.0005467265885629537,
+ 1.8822879602251045e-05
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": false
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.2557002452079744,
+ "selected_factor_vs_raw_float32_with_keys": 2.55808895917612,
+ "selected_full_model_factor_vs_raw_float32": 2.3206989134036973
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "seed": 11,
+ "encoding": "int8",
+ "raw_encoded_backend_with_keys_bytes": 3671291,
+ "best_encoded_baseline": "int8-raw",
+ "best_encoded_baseline_with_keys_bytes": 3671291,
+ "joint_099_backend_with_keys_bytes": 3117210,
+ "joint_099_added_saving_vs_best_encoded_baseline": 0.15092265908640856,
+ "joint_099_saving_vs_raw_encoded": 0.15092265908640856,
+ "joint_099_factor_vs_raw_float32_with_keys": 3.7802470799208265,
+ "joint_099_weighted_relative_frobenius": 0.014090331843047801,
+ "joint_099_added_weighted_relative_frobenius": 0.013331484320787465,
+ "joint_099_max_relative_l2": 0.1410713069702977,
+ "joint_099_warm_query_ratio_vs_raw_encoded": 1.1465622808578337,
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": 0.16885794357434025,
+ "joint_099_factor_vs_raw_float32_with_datrie": 4.177769800712289,
+ "validation_selection": {
+ "feasible": true,
+ "name": "int8-joint-0.95",
+ "backend_with_key_registry_bytes": 3114251,
+ "test": {
+ "targets": 13043,
+ "cross_entropy_nats": 1.7496451964622277,
+ "character_perplexity": 5.752561284785639,
+ "next_character_accuracy": 0.5254159319175037,
+ "unknown_targets": 0
+ },
+ "test_delta": {
+ "perplexity_relative_change": 0.00020816577697060302,
+ "mean_prediction_kl_nats": 6.48036425053271e-05,
+ "prediction_kl_p99_nats": 0.00021414928574367274,
+ "maximum_prediction_kl_nats": 0.15112539268236266,
+ "top1_changed_fraction": 0.0026067622479490913,
+ "accuracy_change": -0.00015333895576163936,
+ "maximum_absolute_logit_change": 1.0701985359191895,
+ "maximum_per_target_cross_entropy_increase": 0.7534070014953613
+ },
+ "test_paired_block_interval": {
+ "perplexity_relative_change_95_percent_interval": [
+ 1.1390863152666506e-05,
+ 0.0004288786464272205
+ ],
+ "block_characters": 256,
+ "observed_blocks": 52,
+ "repeats": 1000,
+ "seed": 42,
+ "limitation": "Paired contiguous character blocks within one text split; not independent corpora."
+ },
+ "test_all_budgets_passed": true
+ },
+ "selected_added_saving_vs_best_encoded_baseline": 0.15172864259466223,
+ "selected_factor_vs_raw_float32_with_keys": 3.78383887490122,
+ "selected_full_model_factor_vs_raw_float32": 3.1991439404258837
+ }
+ ],
+ "groups": [
+ {
+ "category": "gru",
+ "max_order": 4,
+ "encoding": "fp16",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.05262329656093564,
+ "min": 0.05245480941018332,
+ "max": 0.05272242879039157
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.06013974413552303,
+ "min": 0.05998944723919386,
+ "max": 0.06025266667066398
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 2.02428846884593,
+ "min": 2.023964780301054,
+ "max": 2.0245316846591632
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.005405079587587055,
+ "min": 0.005302751473250229,
+ "max": 0.005608434419236457
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.005401605905993721,
+ "min": 0.005299381939658822,
+ "max": 0.005604080214656733
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.12307542272477053,
+ "min": 0.11263148326372621,
+ "max": 0.14016502249394194
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.0539203428668016,
+ "min": 1.0363339928189348,
+ "max": 1.0635481057016951
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.05482497336524975,
+ "min": 0.054649478025039944,
+ "max": 0.054928288836961214
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 2.0696288943808927,
+ "min": 2.0692833078223476,
+ "max": 2.069888565003748
+ }
+ },
+ "validation_selected": {
+ "feasible_models": 3,
+ "test_all_budgets_passed_models": 2,
+ "selected_names": [
+ "fp16-joint-0.9",
+ "fp16-joint-0.95",
+ "fp16-joint-0.9"
+ ],
+ "added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.053246207900067656,
+ "min": 0.052710468357521245,
+ "max": 0.05372248710136929
+ },
+ "factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 2.0256206805754893,
+ "min": 2.024511017420642,
+ "max": 2.0265720661513837
+ },
+ "full_model_factor_vs_raw_float32": {
+ "n": 3,
+ "mean": 1.8632297059848535,
+ "min": 1.862370584235847,
+ "max": 1.8639662486266602
+ }
+ }
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "encoding": "fp32",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.057372272023566685,
+ "min": 0.057214545021257335,
+ "max": 0.057496690371830916
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.08753531127974452,
+ "min": 0.08735430994844451,
+ "max": 0.08771862371335948
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 1.095932849713197,
+ "min": 1.095715468665074,
+ "max": 1.0961530356680196
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.005414277401074073,
+ "min": 0.0052940877733579465,
+ "max": 0.005649047524990139
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.005414277401074073,
+ "min": 0.0052940877733579465,
+ "max": 0.005649047524990139
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.12307561430113703,
+ "min": 0.11261831137802518,
+ "max": 0.14017569822550982
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.183254793938537,
+ "min": 1.1615017998270556,
+ "max": 1.2134813021805575
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.058640812140851116,
+ "min": 0.05847955746099098,
+ "max": 0.05876807091890335
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 1.0981861249794447,
+ "min": 1.0979631809733372,
+ "max": 1.0984119463777775
+ }
+ },
+ "validation_selected": {
+ "feasible_models": 3,
+ "test_all_budgets_passed_models": 2,
+ "selected_names": [
+ "fp32-joint-0.9",
+ "fp32-joint-0.95",
+ "fp32-joint-0.9"
+ ],
+ "added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.05804078602793272,
+ "min": 0.05751601175785381,
+ "max": 0.05858481406499183
+ },
+ "factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 1.0967108746423382,
+ "min": 1.0960659486421322,
+ "max": 1.0974200112886179
+ },
+ "full_model_factor_vs_raw_float32": {
+ "n": 3,
+ "mean": 1.0877711818320588,
+ "min": 1.0871906337528756,
+ "max": 1.0884095070641355
+ }
+ }
+ },
+ {
+ "category": "gru",
+ "max_order": 4,
+ "encoding": "int8",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": -0.045570183278392795,
+ "min": -0.04578634287068262,
+ "max": -0.045461064253438765
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": -0.045570183278392795,
+ "min": -0.04578634287068262,
+ "max": -0.045461064253438765
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 3.1538443098267015,
+ "min": 3.153192356232175,
+ "max": 3.154173421988336
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.007729864985927438,
+ "min": 0.007636228340675911,
+ "max": 0.007915395391086759
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.005609781509268933,
+ "min": 0.005479980336186081,
+ "max": 0.005829303558166024
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.11732259539247156,
+ "min": 0.11380766731945188,
+ "max": 0.12213938394952205
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.1163605568830008,
+ "min": 1.1052328226907326,
+ "max": 1.1232070236978005
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": -0.0489502761197446,
+ "min": -0.04918246899164669,
+ "max": -0.0488330633719094
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 3.3061468983951388,
+ "min": 3.305415139408319,
+ "max": 3.3065162971288444
+ }
+ },
+ "validation_selected": {
+ "feasible_models": 3,
+ "test_all_budgets_passed_models": 3,
+ "selected_names": [
+ "int8-raw",
+ "int8-raw",
+ "int8-raw"
+ ],
+ "added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.0,
+ "min": 0.0,
+ "max": 0.0
+ },
+ "factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 3.2975655025918367,
+ "min": 3.2975655025918367,
+ "max": 3.2975655025918367
+ },
+ "full_model_factor_vs_raw_float32": {
+ "n": 3,
+ "mean": 2.7589069627937097,
+ "min": 2.7589069627937097,
+ "max": 2.7589069627937097
+ }
+ }
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "encoding": "fp16",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.21129972826892873,
+ "min": 0.21101267894455367,
+ "max": 0.21174195497694215
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.2197572769484091,
+ "min": 0.21953123969309696,
+ "max": 0.22017033835854738
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 2.418778243816425,
+ "min": 2.4180773838429768,
+ "max": 2.4200590858799536
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.011344388337072204,
+ "min": 0.011234116695951897,
+ "max": 0.011436593582149768
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.011343131937587545,
+ "min": 0.011232898757745064,
+ "max": 0.011436043946536932
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14058268739868557,
+ "min": 0.139832944885907,
+ "max": 0.1411682070021142
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.0611758044102917,
+ "min": 1.0567026305344063,
+ "max": 1.0654486678895083
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.2242243575877697,
+ "min": 0.22392076767335312,
+ "max": 0.22469320440478902
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 2.530644150508154,
+ "min": 2.5298530481412707,
+ "max": 2.532089920980823
+ }
+ },
+ "validation_selected": {
+ "feasible_models": 3,
+ "test_all_budgets_passed_models": 2,
+ "selected_names": [
+ "fp16-joint-0.95",
+ "fp16-joint-0.97",
+ "fp16-joint-0.9"
+ ],
+ "added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.21232289981128669,
+ "min": 0.21194339924966987,
+ "max": 0.21262057464340534
+ },
+ "factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 2.4219200760843282,
+ "min": 2.4209332113537703,
+ "max": 2.4227314232126553
+ },
+ "full_model_factor_vs_raw_float32": {
+ "n": 3,
+ "mean": 2.216038130030947,
+ "min": 2.2152658796867795,
+ "max": 2.216673011748056
+ }
+ }
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "encoding": "fp32",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.2318017515066569,
+ "min": 0.23149309199381918,
+ "max": 0.23224866235343367
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.2572987268278911,
+ "min": 0.25697106893668376,
+ "max": 0.2577131676229264
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 1.3464365948650967,
+ "min": 1.345842615534423,
+ "max": 1.347188117021603
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.011342644651789415,
+ "min": 0.011233518943283134,
+ "max": 0.011433905190859801
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.011342644651789415,
+ "min": 0.011233518943283134,
+ "max": 0.011433905190859801
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14058794462480786,
+ "min": 0.1398521329039012,
+ "max": 0.14115591128178223
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.174373386027238,
+ "min": 1.1570228240486526,
+ "max": 1.1877255184964386
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.23927987549733828,
+ "min": 0.23896095504842274,
+ "max": 0.23974102039406286
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 1.3611284012391727,
+ "min": 1.3605024773899466,
+ "max": 1.3619203505928132
+ }
+ },
+ "validation_selected": {
+ "feasible_models": 3,
+ "test_all_budgets_passed_models": 2,
+ "selected_names": [
+ "fp32-joint-0.95",
+ "fp32-joint-0.97",
+ "fp32-joint-0.9"
+ ],
+ "added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.23293751411667574,
+ "min": 0.23254199753757676,
+ "max": 0.23329151054570574
+ },
+ "factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 1.3484302195450621,
+ "min": 1.3476820149229627,
+ "max": 1.3491377086105725
+ },
+ "full_model_factor_vs_raw_float32": {
+ "n": 3,
+ "mean": 1.3184155058014044,
+ "min": 1.3177469613751664,
+ "max": 1.3190476501208241
+ }
+ }
+ },
+ {
+ "category": "gru",
+ "max_order": 8,
+ "encoding": "int8",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.10952643840453335,
+ "min": 0.10924465189094346,
+ "max": 0.1100093473533954
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.10952643840453335,
+ "min": 0.10924465189094346,
+ "max": 0.1100093473533954
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 3.628313465251125,
+ "min": 3.627165125870524,
+ "max": 3.630281649291381
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.012548900159934576,
+ "min": 0.012475970273467448,
+ "max": 0.01266636913425628
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.011617120804413752,
+ "min": 0.011543854934588026,
+ "max": 0.011752570284712499
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14003668208464848,
+ "min": 0.13922942015502465,
+ "max": 0.1411733068123005
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.139527361292771,
+ "min": 1.1320600286985667,
+ "max": 1.1508901824706497
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.12137550571289905,
+ "min": 0.12106323425508236,
+ "max": 0.12191065794404288
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 3.9519363583297626,
+ "min": 3.950531561961863,
+ "max": 3.954344128484711
+ }
+ },
+ "validation_selected": {
+ "feasible_models": 3,
+ "test_all_budgets_passed_models": 2,
+ "selected_names": [
+ "int8-joint-0.95",
+ "int8-joint-0.97",
+ "int8-joint-0.9"
+ ],
+ "added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.11048558807178328,
+ "min": 0.11012453565723312,
+ "max": 0.11069757326960228
+ },
+ "factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 3.632225582645977,
+ "min": 3.630751564467595,
+ "max": 3.63309110290252
+ },
+ "full_model_factor_vs_raw_float32": {
+ "n": 3,
+ "mean": 3.099209963909775,
+ "min": 3.0982068933606906,
+ "max": 3.0997989403935953
+ }
+ }
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "encoding": "fp16",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.25418267674674005,
+ "min": 0.2539920281358392,
+ "max": 0.254429347079815
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.26249628349856324,
+ "min": 0.2620759674191677,
+ "max": 0.26278710772274094
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 2.552244316005807,
+ "min": 2.5507901419926418,
+ "max": 2.5532507197917154
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.012594191468031027,
+ "min": 0.012205143139227746,
+ "max": 0.013052639637527957
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.012592920691303537,
+ "min": 0.012203608015228266,
+ "max": 0.01305117528543782
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14049725254822804,
+ "min": 0.13994770767903386,
+ "max": 0.14112562218626845
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.0573447823046482,
+ "min": 1.057025491895454,
+ "max": 1.057732670135994
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.2712700746846591,
+ "min": 0.271060882896614,
+ "max": 0.2715344857103401
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 2.6954379136341373,
+ "min": 2.693760522246224,
+ "max": 2.696598826538473
+ }
+ },
+ "validation_selected": {
+ "feasible_models": 3,
+ "test_all_budgets_passed_models": 2,
+ "selected_names": [
+ "fp16-joint-0.9",
+ "fp16-joint-0.95",
+ "fp16-joint-0.9"
+ ],
+ "added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.2555258321694721,
+ "min": 0.2553202728833849,
+ "max": 0.2557002452079744
+ },
+ "factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 2.5568489030338406,
+ "min": 2.5561523519057006,
+ "max": 2.55808895917612
+ },
+ "full_model_factor_vs_raw_float32": {
+ "n": 3,
+ "mean": 2.3197452298966748,
+ "min": 2.3192095304453875,
+ "max": 2.3206989134036973
+ }
+ }
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "encoding": "fp32",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.27946455411380905,
+ "min": 0.27920392307953634,
+ "max": 0.27971806131506116
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.3035639647480592,
+ "min": 0.3031917313089537,
+ "max": 0.3038427933071641
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 1.4358822691596083,
+ "min": 1.4351150021203094,
+ "max": 1.4364571541973967
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.012593844539329548,
+ "min": 0.012203466418320411,
+ "max": 0.013052669069275194
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.012593844539329548,
+ "min": 0.012203466418320411,
+ "max": 0.013052669069275194
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.1405269182736026,
+ "min": 0.1400040347803519,
+ "max": 0.14116478475441507
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.1695773158755363,
+ "min": 1.1577321344867262,
+ "max": 1.191513651349131
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.28937172358400065,
+ "min": 0.28910008220535777,
+ "max": 0.2896347151191778
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 1.4576267895756976,
+ "min": 1.456809057629733,
+ "max": 1.4582394940414902
+ }
+ },
+ "validation_selected": {
+ "feasible_models": 3,
+ "test_all_budgets_passed_models": 2,
+ "selected_names": [
+ "fp32-joint-0.9",
+ "fp32-joint-0.95",
+ "fp32-joint-0.9"
+ ],
+ "added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.28091404070145753,
+ "min": 0.28068301539457363,
+ "max": 0.28117377855358505
+ },
+ "factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 1.4387765963242019,
+ "min": 1.4383841421883004,
+ "max": 1.439475116339154
+ },
+ "full_model_factor_vs_raw_float32": {
+ "n": 3,
+ "mean": 1.398490554367982,
+ "min": 1.3981441181889676,
+ "max": 1.3991071635586236
+ }
+ }
+ },
+ {
+ "category": "gru",
+ "max_order": 12,
+ "encoding": "int8",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.15100546374558702,
+ "min": 0.15092265908640856,
+ "max": 0.15109752945217358
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.15100546374558702,
+ "min": 0.15092265908640856,
+ "max": 0.15109752945217358
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 3.7806158042700186,
+ "min": 3.7802470799208265,
+ "max": 3.781025795041212
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.013643093005131621,
+ "min": 0.0133093761146089,
+ "max": 0.014090331843047801
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.012833885440324298,
+ "min": 0.01244834613064305,
+ "max": 0.013331484320787465
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.1406393301309476,
+ "min": 0.14034170133988702,
+ "max": 0.1410713069702977
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.1305416961868953,
+ "min": 1.1160896130346232,
+ "max": 1.1465622808578337
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.16895058853933023,
+ "min": 0.16885794357434025,
+ "max": 0.1690535951122467
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 4.178235575283765,
+ "min": 4.177769800712289,
+ "max": 4.178753482790604
+ }
+ },
+ "validation_selected": {
+ "feasible_models": 3,
+ "test_all_budgets_passed_models": 3,
+ "selected_names": [
+ "int8-joint-0.9",
+ "int8-joint-0.95",
+ "int8-joint-0.95"
+ ],
+ "added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.15207529631038963,
+ "min": 0.15172864259466223,
+ "max": 0.15250902203066985
+ },
+ "factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 3.785386361266616,
+ "min": 3.78383887490122,
+ "max": 3.7873230772395323
+ },
+ "full_model_factor_vs_raw_float32": {
+ "n": 3,
+ "mean": 3.20017733824811,
+ "min": 3.1991439404258837,
+ "max": 3.2014705852488614
+ }
+ }
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "encoding": "fp16",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.2294035598131773,
+ "min": 0.2279345167444513,
+ "max": 0.23220339587770844
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.24853988371030292,
+ "min": 0.2471847324886951,
+ "max": 0.2512146952023193
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 2.5950667136005303,
+ "min": 2.590378864561812,
+ "max": 2.604320284313853
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.006797437208006904,
+ "min": 0.006748195908759052,
+ "max": 0.0068512795300496
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.006794361702793634,
+ "min": 0.0067452159501694255,
+ "max": 0.006848195096817759
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.1414701731638959,
+ "min": 0.1408746714806218,
+ "max": 0.14214298103013276
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.0382398381850755,
+ "min": 1.0375786909548252,
+ "max": 1.039525
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.234060033360687,
+ "min": 0.23256188224259455,
+ "max": 0.2369163474762418
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 2.637338333723454,
+ "min": 2.6324476448111644,
+ "max": 2.64699231764423
+ }
+ }
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "encoding": "fp32",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.24007898752801457,
+ "min": 0.2385916134232997,
+ "max": 0.24278208856300088
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.2733092977398144,
+ "min": 0.27193456075922307,
+ "max": 0.2759242361359673
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 1.3761102056165533,
+ "min": 1.3735029107312045,
+ "max": 1.3810709457577985
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.0067933634444704225,
+ "min": 0.006743679010149183,
+ "max": 0.00684819310974004
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.0067933634444704225,
+ "min": 0.006743679010149183,
+ "max": 0.00684819310974004
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.1414587757908001,
+ "min": 0.14087181586808722,
+ "max": 0.14212765851757603
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.135549242440358,
+ "min": 1.1301373730529407,
+ "max": 1.1413276576532998
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.24260288846516975,
+ "min": 0.24110004291680398,
+ "max": 0.24533451295128494
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 1.3813308352486993,
+ "min": 1.3786772653206654,
+ "max": 1.386379638518033
+ }
+ }
+ },
+ {
+ "category": "sgns",
+ "max_order": 4,
+ "encoding": "int8",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.2003316416087755,
+ "min": 0.19904426371308837,
+ "max": 0.20268854005932424
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.2003316416087755,
+ "min": 0.19904426371308837,
+ "max": 0.20268854005932424
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 4.517328975497451,
+ "min": 4.510048589238265,
+ "max": 4.530662695795015
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.00941293042021904,
+ "min": 0.00940051238979808,
+ "max": 0.009423937975726222
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.006916956251574826,
+ "min": 0.006861641269675814,
+ "max": 0.006974880559954164
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14159745045361313,
+ "min": 0.1411979477508241,
+ "max": 0.14188875915362323
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.072977692082509,
+ "min": 1.0675390597903482,
+ "max": 1.0778523599944267
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.20779942030070733,
+ "min": 0.20646405271581947,
+ "max": 0.2102441770440685
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 4.68283901856197,
+ "min": 4.674936347107521,
+ "max": 4.697312681798634
+ }
+ }
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "encoding": "fp16",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.33215709725365755,
+ "min": 0.33129892564748875,
+ "max": 0.3338004378815612
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.3665750509845665,
+ "min": 0.36556057757426064,
+ "max": 0.36808577600624837
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 3.0686736165632755,
+ "min": 3.0637577196265404,
+ "max": 3.0760008309474016
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.009513704223477967,
+ "min": 0.009429329011061569,
+ "max": 0.009591286480164959
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.00951173706748076,
+ "min": 0.009427344865478818,
+ "max": 0.009589247498575619
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14216555910067064,
+ "min": 0.14199922735932338,
+ "max": 0.1423848947835388
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.0365835521368776,
+ "min": 1.0364142927742688,
+ "max": 1.0368272727272727
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.3415884516744569,
+ "min": 0.3407028567210101,
+ "max": 0.34327771601163126
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 3.1578866944393402,
+ "min": 3.1526096620638993,
+ "max": 3.165752491802803
+ }
+ }
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "encoding": "fp32",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.3573551431848582,
+ "min": 0.3562779186898901,
+ "max": 0.3591413416286595
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.39282200132309647,
+ "min": 0.3917178981556664,
+ "max": 0.3944589882705757
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 1.6469697327263897,
+ "min": 1.6439740655987796,
+ "max": 1.6514158093834164
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.009511693118292574,
+ "min": 0.009429837428214146,
+ "max": 0.009587863324407826
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.009511693118292574,
+ "min": 0.009429837428214146,
+ "max": 0.009587863324407826
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.1421603661722247,
+ "min": 0.14199628030247896,
+ "max": 0.14236738964640794
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.1427405410336353,
+ "min": 1.1293635966132687,
+ "max": 1.149934767889287
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.36252453619335084,
+ "min": 0.3614309879116906,
+ "max": 0.3643361336175788
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 1.6616510863244953,
+ "min": 1.6585601125709268,
+ "max": 1.6662387213949807
+ }
+ }
+ },
+ {
+ "category": "sgns",
+ "max_order": 8,
+ "encoding": "int8",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.28300198507536506,
+ "min": 0.2807873357622206,
+ "max": 0.28543019668566394
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.32701442659373947,
+ "min": 0.32495399050772833,
+ "max": 0.32906423926757133
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 5.320377951746535,
+ "min": 5.304105640853472,
+ "max": 5.336599353230042
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.011672811168299305,
+ "min": 0.011437149201513133,
+ "max": 0.012055645188533535
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.009641461225460295,
+ "min": 0.00957913732749358,
+ "max": 0.009713811108145775
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14288202280201326,
+ "min": 0.14274862236861094,
+ "max": 0.14306266325996833
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.0736467855649692,
+ "min": 1.0692774268035872,
+ "max": 1.078121273428569
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.2983342252007142,
+ "min": 0.29604947771935075,
+ "max": 0.30088850924233634
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 5.653970935489896,
+ "min": 5.63534460216359,
+ "max": 5.6725419135334585
+ }
+ }
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "encoding": "fp16",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.3310568078982729,
+ "min": 0.33015628708135025,
+ "max": 0.3325451296041324
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.3676898763537652,
+ "min": 0.3666544525049803,
+ "max": 0.3691938861779126
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 3.0726327223431906,
+ "min": 3.0676003365985034,
+ "max": 3.0799495631186087
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.009543758176454063,
+ "min": 0.009470324713106664,
+ "max": 0.009600233959965716
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.00954159442650828,
+ "min": 0.009468034050420173,
+ "max": 0.009598270767519452
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14212336843888312,
+ "min": 0.1419611647297449,
+ "max": 0.1422308437423335
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.0364273299720135,
+ "min": 1.0361909090909092,
+ "max": 1.036754915059693
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.3409039775511032,
+ "min": 0.33997372302799955,
+ "max": 0.3424381315092502
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 3.166170201414027,
+ "min": 3.1607505539838026,
+ "max": 3.1740504429574834
+ }
+ }
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "encoding": "fp32",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.35658303146987974,
+ "min": 0.3554473497210139,
+ "max": 0.35827744641341375
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.393169142610682,
+ "min": 0.39200256691366864,
+ "max": 0.3947760570376373
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 1.6479118296550685,
+ "min": 1.6447437860449108,
+ "max": 1.6522809641425362
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.009539963511046499,
+ "min": 0.009467831744684692,
+ "max": 0.009597568466419546
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.009539963511046499,
+ "min": 0.009467831744684692,
+ "max": 0.009597568466419546
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14209324303693532,
+ "min": 0.1419349417581233,
+ "max": 0.1422066988799605
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.15635565439026,
+ "min": 1.1280072991532342,
+ "max": 1.1747773128392094
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.3619767222428773,
+ "min": 0.3608230040398106,
+ "max": 0.3636968463339676
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 1.6632725159377728,
+ "min": 1.6599991327470642,
+ "max": 1.6677870562442678
+ }
+ }
+ },
+ {
+ "category": "sgns",
+ "max_order": 12,
+ "encoding": "int8",
+ "seeds": [
+ 3,
+ 7,
+ 11
+ ],
+ "joint_099": {
+ "joint_099_added_saving_vs_best_encoded_baseline": {
+ "n": 3,
+ "mean": 0.2808103402842879,
+ "min": 0.2788578259313058,
+ "max": 0.2831127868987199
+ },
+ "joint_099_saving_vs_raw_encoded": {
+ "n": 3,
+ "mean": 0.32798049839980964,
+ "min": 0.3258565047545223,
+ "max": 0.3305918401805107
+ },
+ "joint_099_factor_vs_raw_float32_with_keys": {
+ "n": 3,
+ "mean": 5.321172481234668,
+ "min": 5.304361963879888,
+ "max": 5.341884561642363
+ },
+ "joint_099_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.011484770366127247,
+ "min": 0.011273324565210031,
+ "max": 0.011592827136221024
+ },
+ "joint_099_added_weighted_relative_frobenius": {
+ "n": 3,
+ "mean": 0.009657831648581904,
+ "min": 0.009600591901907743,
+ "max": 0.009708818886474999
+ },
+ "joint_099_max_relative_l2": {
+ "n": 3,
+ "mean": 0.14231275171970903,
+ "min": 0.14200820684716933,
+ "max": 0.14267755652829675
+ },
+ "joint_099_warm_query_ratio_vs_raw_encoded": {
+ "n": 3,
+ "mean": 1.076150952697568,
+ "min": 1.0752103506868385,
+ "max": 1.0778555181250178
+ },
+ "joint_099_added_saving_vs_best_encoded_baseline_with_datrie": {
+ "n": 3,
+ "mean": 0.2967709584686942,
+ "min": 0.294740745009476,
+ "max": 0.29921592430482635
+ },
+ "joint_099_factor_vs_raw_float32_with_datrie": {
+ "n": 3,
+ "mean": 5.67043328467963,
+ "min": 5.65107222816794,
+ "max": 5.694291996182165
+ }
+ }
+ }
+ ],
+ "raw_report_sha256": "eb73b44ef9ccd359160b38a371c2629f9c1f7ea8d0990450dc67ec125cfa08d0"
+}
diff --git a/docs/quantized_structural_evidence_20261031/tests.txt b/docs/quantized_structural_evidence_20261031/tests.txt
new file mode 100644
index 0000000..074576f
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/tests.txt
@@ -0,0 +1,29 @@
+test_failed_export_removes_only_new_owned_directory (tests.test_quantized_memory.CodecTest.test_failed_export_removes_only_new_owned_directory) ... ok
+test_finite_subnormal_zero_signed_zero_and_overflow (tests.test_quantized_memory.CodecTest.test_finite_subnormal_zero_signed_zero_and_overflow) ... ok
+test_int8_code_equality_does_not_discard_different_scales (tests.test_quantized_memory.CodecTest.test_int8_code_equality_does_not_discard_different_scales) ... ok
+test_mark_compact_forwards_entire_quantized_representative (tests.test_quantized_memory.QuantizedNativeTest.test_mark_compact_forwards_entire_quantized_representative) ... ok
+test_original_ids_padding_repeated_queries_and_scale_sensitive_dedup (tests.test_quantized_memory.QuantizedNativeTest.test_original_ids_padding_repeated_queries_and_scale_sensitive_dedup) ... ok
+test_quantized_reader_preserves_model_reduction_logits_and_restores_on_failure (tests.test_quantized_memory.QuantizedNativeTest.test_quantized_reader_preserves_model_reduction_logits_and_restores_on_failure) ... ok
+test_size_winner_requires_all_budgets_and_cannot_use_test (tests.test_quantized_memory.ValidationSelectionTest.test_size_winner_requires_all_budgets_and_cannot_use_test) ... ok
+test_rejects_dangling_chained_out_of_range_and_noninteger_maps (tests.test_structural_native.CompactionTest.test_rejects_dangling_chained_out_of_range_and_noninteger_maps) ... ok
+test_changed_source_and_failed_export_do_not_leave_partial_directory (tests.test_structural_native.RealStructuralTest.test_changed_source_and_failed_export_do_not_leave_partial_directory) ... ok
+test_compact_export_uses_existing_frozen_embedding_adapter (tests.test_structural_native.RealStructuralTest.test_compact_export_uses_existing_frozen_embedding_adapter) ... ok
+test_default_preserves_bytes_and_keeps_unindexed_rows (tests.test_structural_native.RealStructuralTest.test_default_preserves_bytes_and_keeps_unindexed_rows) ... ok
+test_engine_cli_uses_original_key_ids_after_structural_compaction (tests.test_structural_native.RealStructuralTest.test_engine_cli_uses_original_key_ids_after_structural_compaction) ... ok
+test_joint_final_guard_rejects_direction_norm_and_chain_drift (tests.test_structural_native.RealStructuralTest.test_joint_final_guard_rejects_direction_norm_and_chain_drift) ... ok
+test_layout_only_bytes_exact_signed_zero_and_duplicate_keys (tests.test_structural_native.RealStructuralTest.test_layout_only_bytes_exact_signed_zero_and_duplicate_keys) ... ok
+test_prefix_suffix_utf8_keys_original_ids_compact_forwarding_and_both_caches (tests.test_structural_native.RealStructuralTest.test_prefix_suffix_utf8_keys_original_ids_compact_forwarding_and_both_caches) ... ok
+test_quantization_before_datrie_plan_keeps_scales_padding_and_original_ids (tests.test_structural_native.RealStructuralTest.test_quantization_before_datrie_plan_keeps_scales_padding_and_original_ids) ... ok
+test_full_softmax_metrics_against_scalar_probability_oracle (tests.test_causal_ngram_lm.CausalMemoryTest.test_full_softmax_metrics_against_scalar_probability_oracle) ... ok
+test_future_perturbation_cannot_change_previous_keys_or_logits (tests.test_causal_ngram_lm.CausalMemoryTest.test_future_perturbation_cannot_change_previous_keys_or_logits) ... ok
+test_identical_scores_have_zero_paired_character_block_interval (tests.test_causal_ngram_lm.CausalMemoryTest.test_identical_scores_have_zero_paired_character_block_interval) ... ok
+test_streaming_scores_include_unknown_targets_and_keep_state (tests.test_causal_ngram_lm.CausalMemoryTest.test_streaming_scores_include_unknown_targets_and_keep_state) ... ok
+test_test_budget_audit_preserves_validation_choice (tests.test_causal_ngram_lm.CausalMemoryTest.test_test_budget_audit_preserves_validation_choice) ... ok
+test_train_only_registry_and_original_split_boundaries (tests.test_causal_ngram_lm.CausalMemoryTest.test_train_only_registry_and_original_split_boundaries) ... ok
+test_validation_selection_rejects_distribution_shift_despite_low_ppl (tests.test_causal_ngram_lm.CausalMemoryTest.test_validation_selection_rejects_distribution_shift_despite_low_ppl) ... ok
+test_tiny_trained_lm_queries_native_storage_and_has_bit_exact_logits (tests.test_causal_ngram_lm.CausalNativeTest.test_tiny_trained_lm_queries_native_storage_and_has_bit_exact_logits) ... ok
+
+----------------------------------------------------------------------
+Ran 24 tests in 4.520s
+
+OK
diff --git a/docs/quantized_structural_evidence_20261031/verify_codec_revision.py b/docs/quantized_structural_evidence_20261031/verify_codec_revision.py
new file mode 100644
index 0000000..782fe2a
--- /dev/null
+++ b/docs/quantized_structural_evidence_20261031/verify_codec_revision.py
@@ -0,0 +1,76 @@
+"""Check the finite-range codec fix against all cached study inputs.
+
+Run from this trusted repository. Executes the pinned study's own codec source
+from Git to compare encoded bytes; never loads arbitrary network code.
+"""
+import argparse
+import hashlib
+import json
+from pathlib import Path
+import platform
+import shlex
+import subprocess
+import sys
+
+import numpy as np
+import torch
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(ROOT))
+from tools import quantized_memory as current
+
+
+def sha(raw):
+ return hashlib.sha256(raw).hexdigest()
+
+
+def main():
+ p = argparse.ArgumentParser(description=__doc__)
+ p.add_argument("--report", type=Path, required=True)
+ p.add_argument("--sgns-dir", type=Path, required=True)
+ p.add_argument("--gru-dir", type=Path, required=True)
+ p.add_argument("--output", type=Path, required=True)
+ args = p.parse_args()
+ raw = args.report.read_bytes()
+ study = json.loads(raw)
+ source = subprocess.check_output(["git", "show", study["commit"] + ":tools/quantized_memory.py"], cwd=ROOT)
+ if sha(source) != study["source_sha256"]["tools/quantized_memory.py"]:
+ raise ValueError("pinned codec source differs from study")
+ reference = {"__name__": "tools._pinned_study_codec", "__package__": "tools"}
+ exec(compile(source, "pinned study codec", "exec"), reference)
+ checks = []
+ for run in study["runs"]:
+ folder = f"order-{run['max_order']}-seed-{run['seed']}"
+ if run["category"] == "sgns":
+ rows = np.frombuffer((args.sgns_dir / folder / "trained-ngram.f32").read_bytes(), dtype="= 0.99 on decoded encoded rows\nSeed ranges; fixed threshold, no task-quality guarantee from this plot", fontsize=11)
+ fig.tight_layout(rect=(0, 0, 1, .90))
+ for ext in ("svg", "png"):
+ fig.savefig(directory / f"storage-controls.{ext}", dpi=170)
+ plt.close(fig)
+ fig, axes = plt.subplots(len(categories), len(orders), figsize=(12, 6), squeeze=False)
+ for row, category in enumerate(categories):
+ for col, order in enumerate(orders):
+ ax = axes[row, col]
+ runs = [r for r in report["runs"] if r["category"] == category and r["max_order"] == order]
+ for encoding, color in zip(ENCODINGS, ("#599268", "#5587b6", "#ed9c42")):
+ points = []
+ policies = ["raw"] + [f"joint-{t:g}" for t in report["settings"]["threshold_values"]]
+ for policy in policies:
+ ratios, errors = [], []
+ for run in runs:
+ v = {v["name"]: v for v in run["variants"]}
+ item = v[f"{encoding}-{policy}"]
+ ratios.append(v["fp32-raw"][SIZE] / item[SIZE])
+ errors.append(item["fidelity_vs_float32"]["weighted_relative_frobenius"] * 100)
+ points.append((statistics.mean(errors), statistics.mean(ratios)))
+ x, y = np.array(points).T
+ ax.plot(x, y, "o-", color=color, markersize=4, label=encoding.upper())
+ for (xx, yy), policy in zip(points, policies):
+ if policy in ("joint-0.99", "joint-0.9"):
+ ax.annotate(policy.split("-")[1], (xx, yy), fontsize=7, xytext=(2, 4), textcoords="offset points")
+ ax.set_title(f"{category.upper()}, order {order}")
+ ax.set_xlabel("Query-weighted vector error (%)")
+ ax.set_ylabel("FP32 raw / stored bytes")
+ ax.grid(alpha=.2)
+ ax.legend(fontsize=7)
+ fig.suptitle("Final error vs original FP32; mean of trained seeds\nRaw controls + cosine sweep 0.99, 0.98, 0.97, 0.95, 0.90; norm ratio fixed at 0.99", fontsize=11)
+ fig.tight_layout(rect=(0, 0, 1, .90))
+ for ext in ("svg", "png"):
+ fig.savefig(directory / f"rate-distortion.{ext}", dpi=170)
+ plt.close(fig)
+
+
+def main():
+ p = argparse.ArgumentParser(description=__doc__)
+ p.add_argument("--report", type=Path, required=True)
+ p.add_argument("--output-dir", type=Path, required=True)
+ p.add_argument("--plots", action="store_true")
+ args = p.parse_args()
+ raw = args.report.read_bytes()
+ report = json.loads(raw)
+ summary = analyze(report)
+ summary["raw_report_sha256"] = hashlib.sha256(raw).hexdigest()
+ args.output_dir.mkdir(parents=True, exist_ok=True)
+ (args.output_dir / "summary.json").write_text(json.dumps(summary, indent=2) + "\n")
+ if args.plots:
+ plots(report, args.output_dir)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/tools/benchmark_quantized_structural.py b/tools/benchmark_quantized_structural.py
new file mode 100644
index 0000000..1d22f94
--- /dev/null
+++ b/tools/benchmark_quantized_structural.py
@@ -0,0 +1,339 @@
+"""Same-checkpoint CPU controls for quantization followed by structural sharing.
+
+Reuse cached SGNS rows / GRU checkpoints; never retrain, download or tune on
+test. Native planner is the production DATrie path. The explicit reference
+planner is a different portable greedy prefix/suffix control, not DATrie evidence.
+"""
+import argparse
+import hashlib
+import json
+import os
+from pathlib import Path
+import platform
+import shlex
+import subprocess
+import sys
+import time
+
+import numpy as np
+import torch
+from torch import nn
+
+from tools.benchmark_cosine_merge import fidelity, structural_merge
+from tools.benchmark_causal_lm_compression import text_splits, paired_ppl_interval
+from tools.causal_ngram_lm import CausalNgramLM, encode_segment, evaluate, metric_summary, prediction_delta
+from tools.quantized_memory import encode_rows, decode_rows, pack_quantized, QuantizedRows
+from tools.structural_native import StructuralPlan
+from tools.train_ngram_similarity import prepare, ngrams
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def sha(raw):
+ return hashlib.sha256(raw).hexdigest()
+
+
+def file_sha(path):
+ return sha(Path(path).read_bytes())
+
+
+def structural_maps(rows, keys, cfg, threshold):
+ if cfg.planner == "native":
+ with StructuralPlan(rows.tobytes(), rows.shape[1], keys, cfg.structural_library,
+ mode="layout" if threshold is None else "joint",
+ cosine=threshold or .99, norm_ratio=cfg.norm_ratio) as plan:
+ return np.asarray(plan.representatives), plan.row_order
+ keyed = list(keys)
+ representatives = np.arange(len(rows), dtype=np.int64)
+ if threshold is not None:
+ partial = structural_merge(rows[keyed], [keys[i] for i in keyed], threshold, cfg.norm_ratio)
+ representatives[keyed] = np.asarray(keyed)[partial]
+ key_set = set(keys)
+ order = ([i for i in range(len(rows)) if i not in key_set]
+ + [i for i in sorted(keys, key=lambda i: keys[i]) if representatives[i] == i]
+ + [i for i, r in enumerate(representatives) if i != r])
+ if len(order) != len(rows) or len(set(order)) != len(rows) or not np.array_equal(representatives[representatives], representatives):
+ raise AssertionError("invalid original-ID structural map")
+ return representatives, order
+
+
+class ReaderEmbedding(nn.Module):
+ def __init__(self, reader):
+ super().__init__()
+ self.reader = reader
+
+ def forward(self, ids):
+ values = self.reader.gather(ids.detach().cpu().numpy().reshape(-1))
+ return torch.from_numpy(values).reshape(*ids.shape, self.reader.dim)
+
+
+def score_bundle(model, original_memory, directory, segment, cfg):
+ try:
+ with QuantizedRows(directory, cfg.native_library, cache_bytes=cfg.cache_bytes) as reader:
+ model.memory = ReaderEmbedding(reader)
+ return evaluate(model, segment, cfg.burn_in, cfg.eval_chunk)
+ finally:
+ model.memory = original_memory
+
+
+def choose_validation(variants, encoding, cfg):
+ feasible = [v for v in variants if v["encoding"] == encoding
+ and v["validation_delta"]["perplexity_relative_change"] <= cfg.max_ppl_increase
+ and v["validation_delta"]["mean_prediction_kl_nats"] <= cfg.max_kl
+ and v["validation_delta"]["top1_changed_fraction"] <= cfg.max_top1_flip]
+ return min(feasible, key=lambda v: (v["backend_with_key_registry_bytes"], v["name"])) if feasible else None
+
+
+def query_metrics(directory, expected, query, cfg):
+ with QuantizedRows(directory, cfg.native_library, cache_bytes=cfg.cache_bytes) as reader:
+ start = time.perf_counter_ns()
+ output = reader.gather(query)
+ first = (time.perf_counter_ns() - start) / 1e6
+ if output.tobytes() != expected[query].tobytes():
+ raise AssertionError("first gather changed encoded representative output")
+ for _ in range(3):
+ reader.gather(query)
+ samples = []
+ for _ in range(cfg.repeat):
+ start = time.perf_counter_ns()
+ reader.gather(query)
+ samples.append((time.perf_counter_ns() - start) / 1e6)
+ return {"first_empty_application_cache_ms": first,
+ "warm_p50_ms": float(np.median(samples)), "warm_p95_ms": float(np.percentile(samples, 95)),
+ "warm_p99_ms": float(np.percentile(samples, 99)), "samples_ms": samples,
+ "warm_rows_per_second": len(query) / (np.median(samples) / 1000)}
+
+
+def compare(category, order, seed, rows, keys, trace, cfg, model=None, segments=None):
+ directory = cfg.output_dir / f"{category}-order-{order}-seed-{seed}"
+ directory.mkdir()
+ registry_path = directory / "keys.json"
+ registry_path.write_text(json.dumps(keys, ensure_ascii=True, sort_keys=True, separators=(",", ":")) + "\n")
+ common = registry_path.stat().st_size
+ index_bytes = None
+ if cfg.planner == "native":
+ with StructuralPlan(rows.tobytes(), rows.shape[1], keys, cfg.structural_library, mode="layout") as plan:
+ index_path = directory / "keys.das"
+ plan.save_index(index_path)
+ index_bytes = index_path.stat().st_size
+ if plan.lookup(list(keys.values()), index="datrie") != list(keys):
+ raise AssertionError("DATrie key index changed original IDs")
+ other_bytes = 0
+ original_memory = model.memory if model is not None else None
+ baseline = ({name: evaluate(model, segment, cfg.burn_in, cfg.eval_chunk) for name, segment in segments.items()}
+ if model is not None else {})
+ if model is not None:
+ # Actual serialized unchanged model state/registry, not an estimated payload.
+ backbone = {k: v for k, v in model.state_dict().items() if k != "memory.weight"}
+ torch.save({"state": backbone, "registry": cfg.current_registry, "dim": rows.shape[1]}, directory / "backbone.pt")
+ other_bytes = (directory / "backbone.pt").stat().st_size
+ trace = np.asarray(trace, dtype=np.int64)
+ query = trace[:cfg.query_rows]
+ mass = np.bincount(trace, minlength=len(rows))
+ entry = {"category": category, "max_order": order, "seed": seed, "rows": len(rows), "dim": rows.shape[1],
+ "source_float32_sha256": sha(rows.tobytes()), "source_float32_bytes": rows.nbytes,
+ "key_registry_bytes": common, "unchanged_backbone_artifact_bytes": other_bytes,
+ "alternative_datrie_key_index_bytes": index_bytes,
+ "query_rows": len(query), "query_ids_sha256": sha(query.astype("= self.nrows):
+ raise ValueError("one-dimensional original row IDs out of range")
+ if isinstance(self.table, NativeRows):
+ ids = np.ascontiguousarray(ids, dtype="