ci(eval): use --repeats 3 on gpt-oss-120b gpqa-diamond to suppress noise#278
Merged
Conversation
This eval recently flaked at score=0.6919 vs threshold>=0.7 (~2 questions short on 198 total). The 0.7 bar sits at ~1.3σ below the documented 0.74 ± 0.03 mean, so ~11% of clean runs land below threshold from binomial sampling noise alone (see PR lightseekorg#272 CI run 26483245060). evalscope's --repeats N runs each prompt N times and averages the scores before aggregating. With --repeats 3 the per-question variance shrinks by ~1/sqrt(3), cutting the score stddev from ~0.033 to ~0.019. At the unchanged 0.7 threshold this moves the false-fail rate from ~11% down to ~2%, at a cost of roughly 4 extra GPU-min per per-commit run. Threshold left at 0.7 so a real accuracy regression (drop of more than ~2σ from the mean) is still caught. Signed-off-by: Qingyang Wu <willqywu@gmail.com>
At the prior --eval-batch-size 16 the server pinned at 13-16 in-flight requests with page_ratio=0.00-0.01, leaving most decode capacity idle (server max_num_seqs=160). Lifting evalscope client concurrency to 64 roughly cancels the cost of the new --repeats 3 by raising throughput, and lets the prefix cache hit on the repeat duplicates for additional TTFT savings. Eval scoring is independent of batch size, so the noise-reduction benefit of --repeats 3 is preserved. Signed-off-by: Qingyang Wu <willqywu@gmail.com>
syuoni
approved these changes
May 27, 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Recent flake on PR #272 (CI run 26483245060) failed the
eval-gpt-oss-120b-mxfp4-gpqa-diamondcheck at score=0.6919 vs threshold>=0.7 — off by ~2 questions out of 198. The PR's diff was HTTP plumbing only (no model / kernel / sampling code), so the failure is sampling noise, not a regression.GPQA Diamond runs all 198 questions once (AveragePass@1, no
--limit, no--repeats). With a ~70% pass rate this gives a binomial stddev of:matching the YAML's documented
0.74 ± 0.03band. The0.7threshold sits at ~1.3σ below that mean, so a noise-driven false-fail happens on roughly 1 in 9 clean runs.Fix
Add
--repeats 3to the evalscope command. evalscope samples each question 3× and averages the per-question scores before aggregating, shrinking per-run stddev by ~1/√3.The
0.7threshold is unchanged so a real accuracy regression (a drop of more than ~2σ from the documented mean) is still caught.Only
gpt-oss-120b-mxfp4-evalscope-gpqa-diamond.yamlis touched. The kimi-k2.5 gpqa-diamond eval is nightly-only without ascore_threshold— it can't fail the same way and doesn't need the cost.Test plan