Repository navigation
fix(scorers): raise humaneval/mbpp verify timeout 3s -> 10s - #7
Merged
Merged
Conversation
The timeout is wall clock measured on the harness event loop. At 128 connections a burst of finished samples saturates the single inspect process and correct answers time out: Qwen3.8-27B humaneval 0.99 -> 0.30 in DeepInfra's quality monitor, while the same answers verified alone pass. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012SGrRPrAXX7iTxNrHDKCKN
…le timeout Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012SGrRPrAXX7iTxNrHDKCKN
nhat-di
changed the title
Author
|
Quality Checks failure is pre-existing on main, not from this PR: ruff-format on datasets/hle_250_ids.py + evals/hle.py and mypy in datasets/hle.py:88 / scorers/hle.py:70, all from #6 (merged without CI). Left untouched so the hle code in the image stays byte-identical. Tests 3.10-3.13 + integration pass. Merging per Nhat's ack. |
nhat-di
marked this pull request as ready for review
October 20, 2026 10:03
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 humaneval/mbpp verify timeout is wall clock, measured on the inspect event loop. In DeepInfra's quality monitor (128 streaming connections), a burst of finished samples saturates the single inspect process: about 1.0 core, the GIL-bound loop, on an otherwise idle node. Correct answers then hit the 3s timer. Qwen3.8-27B humaneval went 0.99 -> 0.30, then 0.85 on a re-run with no serving change. The straggler samples, verified alone, pass at ~100%.
Canonical HumanEval solutions through the real
verify()(820 samples, width 128,--cpus=1, mock model):Effect on scores: a correct solution that takes 3-10s now passes, where upstream fails it. On HumanEval this should be about zero, since canonical solutions finish well under 1s. It is still a deliberate small divergence from upstream's harness. Why 10 and not 30: a runaway sample holds a verify slot for the full timeout (+2s grace), and 10s already rescues every arm.
This ships with deepinfra/backend#6346, which caps
--max-subprocessesat 2 for jobs that run humaneval/mbpp. Imageopenbench:main-<sha>-r1follows after merge.🤖 Generated with Claude Code
https://claude.ai/code/session_012SGrRPrAXX7iTxNrHDKCKN