Skip to content

fix(reward): grade committed MC letter, not first echoed option (#124)#217

Open
kiannidev wants to merge 1 commit into
mini-router:mainfrom
kiannidev:sn74-kiannidev-choice-letter
Open

fix(reward): grade committed MC letter, not first echoed option (#124)#217
kiannidev wants to merge 1 commit into
mini-router:mainfrom
kiannidev:sn74-kiannidev-choice-letter

Conversation

@kiannidev

@kiannidev kiannidev commented Jul 16, 2026

Copy link
Copy Markdown
Contributor

Summary

Fixes #124: when a model echoes the option list (A) …, B) …) then commits to a final letter, extract_choice_letter returned A via re.search() on the line-start pattern.

Change

  • Drop the line-start option pattern from _CHOICE_PATTERNS
  • Strengthen last-line fallback to accept a trailing standalone letter on the final line
  • Offline regression tests in tests/test_reward_checkers.py
  • JOURNAL entry

Test plan

  • pytest tests/test_reward_checkers.py — 15 passed

…-router#124)

Drop the line-start A)/B) pattern that matched echoed option lists via
re.search(), and let the last-line fallback recover bare or trailing
final letters so correct answers stop scoring 0.

Co-authored-by: Cursor <cursoragent@cursor.com>
@github-actions github-actions Bot added docs Documentation changes miner Miner contribution PR train Training or routing changes labels Jul 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

docs Documentation changes miner Miner contribution PR train Training or routing changes

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Multiple-choice grading picks the first listed option, not the model's answer

1 participant