Skip to content

fix(rl): preserve full tool observations for GRPO scoring - #113

Merged
Algorineko merged 1 commit into
Algorineko:mainfrom
jasonchen505:codex/fix-grpo-full-observation
Oct 17, 2026
Merged

Algorineko merged 1 commit into
Algorineko:mainfrom
jasonchen505:codex/fix-grpo-full-observation

Conversation

@jasonchen505

@jasonchen505 jasonchen505 commented Oct 17, 2026 •

Copy link
Copy Markdown
Contributor

Long tool results are truncated to 1,000 characters before entering the GRPO scoring history. This can turn a valid result into an unparsable dictionary representation and incorrectly trigger a tool-failure penalty.

Preserve the complete result for scoring while keeping the existing bounded policy input and environment-token mask. Record the character-bounded view before token-budget clipping.

Validation:

  • Five deterministic CPU regressions cover long results, empty answers, tool exceptions, policy tokens/masks, and remaining-budget clipping.
  • Full CPU suite with external HTTP/HF access disabled: 838 passed, 7 skipped.

@Algorineko
Algorineko merged commit fad0f2a into Algorineko:main Oct 17, 2026
3 checks passed
Algorineko added a commit that referenced this pull request Oct 31, 2026
…tion

fix(rl): preserve full tool observations for GRPO scoring
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.

2 participants