Skip to content

fix(scoring): BBH exact-match accepts Markdown-emphasised answers - #469

Open
marktech0813 wants to merge 3 commits into
James-CUDA:mainfrom
marktech0813:fix/461-bbh-md-emphasis
Open

fix(scoring): BBH exact-match accepts Markdown-emphasised answers#469
marktech0813 wants to merge 3 commits into
James-CUDA:mainfrom
marktech0813:fix/461-bbh-md-emphasis

Conversation

@marktech0813

Copy link
Copy Markdown
Contributor

Summary

BBH multiple-choice already tolerates `Answer: B` via `reward._strip_choice_md_emphasis`, but exact-match `normalize_exact` left `*`/`` intact, so `Answer: True` graded 0.0 against gold `True`.

Peel whole-answer Markdown emphasis (matched markers, nested via fixpoint) before and after the edge strip.

Closes #461

Test plan

  • Regression: `score_bbh('Answer: True', exact_match True) == 1.0`

Made with Cursor

Peel whole-answer **/*/_/` wraps in _normalize_exact so Answer: **True**
matches gold True, consistent with MCQ half's emphasis tolerance.

Closes James-CUDA#461
Exclude "." from _STRIP_EDGE and retry with rstrip(".") so CI can
validate this branch's own fix (same unblock as James-CUDA#450/James-CUDA#462).
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.

[bug] BBH exact-match grades a Markdown-bolded correct answer 0.0 — the same adapter's MCQ half accepts Answer: **B**

1 participant