8000 Fix target-length qualification metadata by DaoyuanLi2816 · Pull Request #112 · DaoyuanLi2816/mini-verl · GitHub
[go: up one dir, main page]

Skip to content

Fix target-length qualification metadata - #112

Merged
DaoyuanLi2816 merged 1 commit into
mainfrom
v0.12.0-qualification-hotfix
Sep 8, 2026
Merged

DaoyuanLi2816 merged 1 commit into
mainfrom
v0.12.0-qualification-hotfix

Conversation

@DaoyuanLi2816
Copy link
Copy Markdown
Owner

Summary

  • accept the built-in target_length reward metadata through the fail-closed Parquet validator
  • validate its characters target as a bounded non-boolean integer
  • exercise the real v0.12 qualification Parquet through the production dataset scanner

Reproduction

The exact-main GPU qualification reached the v0.12 RL workload and failed before trainer construction because reward_model.characters was rejected by the older exact/rule-only field whitelist: run 34187325636.

Validation

  • ruff check .
  • ruff format --check .
  • mypy src/miniverl
  • pytest -q -m "not gpu and not network" --cov=miniverl: 2572 passed, 10 skipped, 24 deselected; 83% branch coverage
  • frozen calculator JSON SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

The failed qualification is not release evidence. After this PR is green and merged, the release will use a fresh workflow dispatch against the new exact main SHA.

@DaoyuanLi2816
DaoyuanLi2816 merged commit 85f3ad9 into main Sep 8, 2026
16 checks passed
@DaoyuanLi2816
DaoyuanLi2816 deleted the v0.12.0-qualification-hotfix branch September 8, 2026 05:04
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

0