Fix MATH grading to use last Answer and harden equality check - #119
Open
shoemoney wants to merge 1 commit into
Open
Fix MATH grading to use last Answer and harden equality check#119shoemoney wants to merge 1 commit into
shoemoney wants to merge 1 commit into
Conversation
Fixes openai#115. Use last Answer match, wrap expressions, tolerant yes parse. Disclosure: developed in conjunction with my pair programmer Claude.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fixes #115
math_eval now grades the last
Answer:line, not the first. Usesre.findalland takes the final match. This matches the prompt that asks for the last line to hold the answer. Self correction mid response no longer flips the score.common now wraps both expressions in quotes in the equality prompt. Candidate text is delimited and cannot read as instructions. Adds guard for
Noneextracts.common now parses the grader reply with a tolerant check. Searches for word
yes. Handles case, punctuation, and short rationales likeyes.orThe answer is yes.Testing: reproduced both flip directions and the prompt injection and strict parse failures without API calls. Verified patched extraction picks last answer and tolerant parse accepts all expected yes variants. No existing repo tests affected.