Document Qwen3.5 GRPO lift

[bingran-you]

Summary

• record the clean 16-train/14-held-out Qwen3.5-9B OpenCode-only run
• distinguish it from the invalid historical soccer canary
• document strict teacher/SFT coverage, 128 GRPO rollouts, nonzero gradients, finite LoRA updates, and paired held-out lift
• state the small-sample confidence boundary and the runtime BenchFlow commit used

Result

• baseline: 8/14
• SFT: 8/14
• SFT + GRPO: 11/14
• paired lift: +3/14 (+21.4 percentage points)
• regressions: 0
• mixed reward groups: 4/16
• nonzero-gradient steps: 30/128
• updated LoRA-B tensors: 248/248

Verification

• pytest pipelines/benchflow-task-posttrain/tests -q (237 passed)
• git diff --check
• stale no-lift/awaiting-rerun wording scan

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d8313dadad

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Record the actual BenchFlow pin

The run is described as produced by PostTrainArena cf824b214e5ae08d6fc21becbcba7aae55e5109e, but that code path pins and records BenchFlow as cbc295464e62aa39f84e0daa675aa939c0e72f00 in both BENCHFLOW_COMMIT and the package dependency, and plan.json/score.json serialize that constant. Unless the run used an out-of-band dependency override, documenting 6d6d2ee... here makes the provenance disagree with the generated artifacts and sends anyone trying to reproduce or audit this result to the wrong BenchFlow revision.

Useful? React with 👍 / 👎.