fix(metrics): weight gpt-oss accuracy by dataset sample count to matc… - #479
Merged
Conversation
…h MLCommons
The combined gpt-oss-120b accuracy was an unweighted mean of the three
per-dataset scores (aime25/gpqa/livecodebench). That over-weights small
datasets (AIME's 30 problems counted equally with LiveCodeBench's 1055) and
diverges from the MLCommons reference submission score. Switch to a
unique-problem-count weighted mean:
overall = sum(score_d * unit_samples_d) / sum(unit_samples_d)
which reproduces the reference exact_match (83.374 vs the old 80.95 on the
same run). Rename the aggregator to samples_weighted_average_accuracy; the
accuracy_results.json key stays "average_accuracy" for consumer stability.
Also fix the gpt-oss accuracy plot (results_plots.py), which still computed an
unweighted overall and would otherwise disagree with the report/JSON headline.
- Single-dataset models (DeepSeek-R1, BFCL, ...) are unchanged (weight cancels).
- Weight guard: absent/None unit_samples -> 1.0 (legacy artifacts); a present
but non-positive/non-numeric weight is treated as corrupt and skipped.
- Tests: MLPerf reference lock (83.374), plot weighting, guard branches, and the
average_accuracy JSON-key contract.
|
MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅ |
github-actions
Bot
added
the
size/normal
arekay-nv
marked this pull request as ready for review
August 26, 2026 22:12
nvzhihanj
approved these changes
Sep 1, 2026
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
3 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The combined gpt-oss-120b accuracy was an unweighted mean of the three per-dataset scores (aime25/gpqa/livecodebench). That over-weights small datasets (AIME's 30 problems counted equally with LiveCodeBench's 1055) and diverges from the MLCommons reference submission score. Switch to a unique-problem-count weighted mean:
which reproduces the reference exact_match (83.374 vs the old 80.95 on the same run). Rename the aggregator to samples_weighted_average_accuracy; the accuracy_results.json key stays "average_accuracy" for consumer stability.
Also fix the gpt-oss accuracy plot (results_plots.py), which still computed an unweighted overall and would otherwise disagree with the report/JSON headline.
What does this PR do?
Type of change
Related issues
Testing
Checklist