Repository navigation
feat(eval): NLI bake-off harness and its first result - #12
Merged
Merged
Conversation
k=3 was already measured and does not raise accuracy, so the next accuracy step is a measured comparison of NLI models with a decision rule fixed in advance. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
…ilities) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
The first wording left the step order open; read as "tie rule first" it let a large model through on a 0.022 lead. Clarified before any number was seen. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
The final review found the report under-stated its own limits: rule-decided rows set the high cut, int8 scores move with the batch, the AVeriTeC snapshot is new, and dev both chooses and reports. The report now says so in the file. Result: MoritzLaurer large mnli-fever-anli-ling-wanli at k=2 wins on SciFact dev macro-F1 (0.697 vs 0.580 baseline); AVeriTeC stays below its majority baseline. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp
|
The latest updates on your projects. Learn more about Vercel for GitHub. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this changes
scripts/eval_nli_bakeoff.py, withsrc/proofpath/eval/bakeoff.pyandbakeoff_data.py). It compares three NLI models × k ∈ {1,2,3} × three aggregation rules (max,max_nei,margin). Cut points are calibrated on SciFact train. Each combination is scored on SciFact dev and on a frozen AVeriTeC snapshot, and the winner is picked by a decision rule fixed before any number was seen (specdocs/superpowers/specs/2026-10-01-nli-bakeoff-design.md).docs/eval/2026-10-01-nli-bakeoff.md). The winner isMoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanliat k=2 withmaxaggregation.baserow.highcut.Why
Retrieving more passages had already been measured, and it does not help: k=3 scored 0.603 against 0.609 at k=1. The loss is in the NLI model, which is overconfident: the [0.9, 1.0] band is only 0.716 precise. So the next step was a measured model comparison rather than a code change.
Every task got an independent reviewer and a re-review, and a final whole-branch review answered READY TO MERGE. The reviewers caught a silent mis-scoring in
margin, an ambiguous decision-rule order (the user chose the clarification), and a report that did not state its coverage.Checklist
ruff checkandruff format --checkare cleanpytestpasses (2576 passed, 1 skipped)mypy src/proofpath/evalclean)🤖 Generated with Claude Code
https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp