[akanyaani]

DecisionFormer v0.2: train, load and serve decision models on any Hugging Face model, using the checkpoint layout shared with AutoJev and Decision 3.0.

What changes

• Rename deciformer → decisionformer (package, CLI, docs).
• Load and run: AutoDecisionModel.from_pretrained(...) loads our exports and other standard checkpoints (Decision 3.0 vllm-sr/d3*, AutoJev), and decide(state, questions) answers choice, noul and score questions in the System One format.
• Serve: decisionformer serve exposes /v1/systemone (HTTP 422 for requests that don't fit, optional API key). --backend vllm runs Qwen3.5 code-readout checkpoints on vLLM; anything else fails with a clear UnsupportedBackend error.
• Hub: exports are plain Hugging Face model repos (backbone, tokenizer, decision_config.json, readout.safetensors, model card). decisionformer push uploads them (private by default).
• Decision Index: an in-process engine, and a server that speaks the Index's http engine protocol.
• Data: System One requests with gold answers, plus flat one-question rows (e.g. FrontiersMind/deciformer-15k) read as they are. decisionformer data split holds out calibration/dev sets by whole families, keeping each task's share.
• Precision: bf16 by default. train.mixed_precision, finetune.model_dtype, finetune.export_dtype and the QLoRA compute dtype are all configurable. GPUs without bf16 fall back to fp16. Exports keep the base checkpoint's dtype.
• Model support: LoRA all-linear skips MoE routers; T5-family encoders reload correctly; seq2seq models without a text encoder (BART) are refused clearly. Non-causal code readouts now export the standard last-token layout.
• Packaging: sentencepiece/protobuf added for DeBERTa-v3-style tokenizers. A release workflow publishes with PyPI trusted publishing: placeholder-v* tags publish the placeholder, v* tags the real package. CI now also builds the distribution.
• README rewritten: installation, quick start (Python and API), supported backbones, then fine-tuning details.

Testing

• 263 tests pass on CPU (12 skipped). This includes tests/test_zoo.py: 12 encoder and 33 decoder families (dense, mixture-of-experts, hybrid linear attention), each trained, exported, reloaded and checked against the Decision Index response rules.
• GPU validation on one RTX 2060 SUPER (8 GB):
  • vllm-sr/d3-edge loaded with from_pretrained gives probabilities identical to the checkpoint's own runtime (bf16, one question per forward pass).
  • The vLLM backend matches the transformers backend within 7e-4 (median 30 ms per 3-question request).
  • MNLI: Qwen3-0.6B + LoRA reaches 0.828 (v0.1: 0.803). Qwen3.5-0.8B + QLoRA with bidirectional attention reaches 0.840.
  • A private Hub upload round trip reloads with identical answers.
• PyPI: the decisionformer name is reserved with a 0.0.1 placeholder (no code).

Known issue

Qwen3.5-0.8B with causal attention stays at chance on MNLI. It fails the same way with LoRA, QLoRA and an fp32 base: a loss spike near step 25, then a collapse to the class prior. The same model with bidirectional attention learns, and step-0 readout logits equal the language model's own next-token logits exactly. So this looks like an optimization issue, not a wiring bug. A lower-learning-rate check is in progress. Other backbones are unaffected.