feat(metrics): consolidate gpt-oss/deepseek accuracy, dedup reporting, surface TPS

[arekay-nv]

Summary

Deduplicates the reporting layer, gives gpt-oss-120b and deepseek-r1 a uniform per-subset accuracy breakdown, and surfaces TPS everywhere QPS appears.

Reporting dedup

• New shared evaluation/accuracy_results.py (to_float, find_accuracy_breakdown, ACCURACY_METRIC_KEYS, build_breakdown); metrics/results_plots.py and compliance/checker.py now import it instead of their copy-pasted helpers.
• results.json results block is single-sourced from the Report (drops the duplicate QPS recompute in execute.py).

Accuracy consolidation (uniform BFCL-shaped breakdown: overall_accuracy/subset_scores/total_samples, 0–100)

• deepseek-r1: LegacyMLPerfDeepSeekR1Scorer now surfaces its per-subset breakdown via score_breakdown() (was discarded to a scratch file).
• gpt-oss-120b: new GptOssAccuracyScorer runs the 3 member scorers once and rolls them up sample-weighted by unique problem count (matching endpoints-launch/…/generate_points.py). finalize_benchmark emits both the per-subset <subset>::gptoss entries (audit + downstream compat) and one consolidated gptoss entry with a breakdown — from the single scoring pass, no re-scoring.
• Consolidation is opt-in via explicit accuracy_config.group (the ::variant suffix is a naming convention, not a grouping signal). Group members must return scalar scores in [0,1].
• Accuracy now renders in Report.display() / report.txt / result_summary.json; the report is written in a finally so a scoring failure can't discard the perf run's artifacts.

TPS

• Added to results.json; Report.display() gains a symmetric TPS: N/A branch.

Test plan

• New unit tests: test_accuracy_results.py, test_gptoss_scorer.py, test_score_accuracy.py; extended deepseek/report/plots/checker tests. 412 unit tests pass; all pre-commit hooks pass.
• This change was reviewed by a two-reviewer council (Codex + Claude); all findings (report-durability regression, implicit-grouping footgun, member-score validation, downstream-compat) are addressed.

Note: the full benchmark pipeline can't run on macOS (pin_loadgen() requires Linux; ZMQ IPC socket-path limit), so end-to-end integration runs on Linux CI. The grouping dispatch was verified directly.

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[gemini-code-assist]

Code Review

This pull request introduces support for consolidating multiple accuracy datasets into a single combined score using a consolidation group tag, which is particularly useful for rolling up subsets like gpt-oss (e.g., AIME25, GPQA, and LiveCodeBench) into a sample-weighted overall score. It extracts shared accuracy breakdown helpers into a new module, refactors compliance and plotting utilities to use them, and updates the benchmark finalization process to attach consolidated accuracy breakdowns to the report. The feedback suggests a cleaner and safer approach to retrieving the dataset sample count during per-subset audit entry generation by avoiding fragile assert statements and direct internal attribute access in favor of the public num_samples() API.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.