docs(examples): add Qwen3-32B SWE-bench Verified eval (Phase 3) - #3671
Merged
Conversation
Follow-up to the Qwen3-32B data+recipe (#3594) and the shared eval harness (#3585). Docs only — reuses the model-agnostic ../eval scripts, no new scripts. - Fill in the qwen3_32b Phase 3 section: goal, scaffold, run/grade commands (NOPARSER=1 + hermes fallback, thinking-on, TP=4xDP=2), Qwen3 serving notes, and the base-vs-SFT resolve table (base 18.2% -> SFT step499 24.4%, +6.2pt). - Clarify the shared eval README NOPARSER bullet to cover the Qwen3 hermes case. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Abhishree <abhishreetm@gmail.com>
- Drop the stale '(eval in a follow-up PR)' note and the gentler-LR trailing clause; minor shared-eval README wording. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Abhishree <abhishreetm@gmail.com>
athitten
marked this pull request as ready for review
August 25, 2026 19:19
athitten
requested review from
HuiyingLi,
akoumpa and
snowmanwwg
as code owners
August 25, 2026 19:19
Contributor
Author
|
/ok to test 7d43b0d |
athitten
enabled auto-merge (squash)
August 25, 2026 19:36
Add the wandb workspace link above the Gemma4-31B CoderForge SFT loss curve. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Abhishree <abhishreetm@gmail.com>
Contributor
Author
|
/ok to test 741f4f3 |
Add the loss-curve <img> above Phase 3 (png pushed separately). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Abhishree <abhishreetm@gmail.com>
Signed-off-by: athitten <abhishreetm@gmail.com>
Contributor
Author
|
/ok to test c390f85 |
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
2 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.
What
Fills in Phase 3 (SWE-bench Verified eval) for Qwen3-32B — the follow-up of PR (#3594). Docs only: it reuses the shared, model-agnostic
examples/long_context_validation/eval/harness (added in #3585), so there are no newscripts — just the Qwen3 knobs and results.
Why it matters
Unlike the gemma4-31B base-SFT (agentic capability from zero), Qwen3-32B already tool-calls,
so this is a real resolve-rate uplift: CoderForge SFT lifts the base 18.2% → 24.4%
(+6.2 pts) on all 500 SWE-BEnch-Verified instances.
Changelog
qwen3_32b/README.md: Phase 3 section (Goal / Scaffold / Steps / Note / Result) — run + gradecommands via
../eval, the two Qwen3 specifics (NOPARSER=1+ in-harness hermes recovery;thinking-on
OH3_ENABLE_THINKING=1/OH3_TEMPERATURE=0.6),TP=4×DP=2serving, and thebase-vs-SFT resolve table across the gentler-LR checkpoints.
eval/README.md: one-line clarification of theNOPARSER=1bullet to cover the Qwen3 hermes case.Result (all 500 SWE-Bench-Verified)
Pre-checks