Repository navigation
Add final_answer maths variants that score only text after </think> - #564
Open
abhishekraok wants to merge 1 commit into
Open
abhishekraok wants to merge 1 commit into
abhishekraok wants to merge 1 commit into
Conversation
AIME, HMMT, Minerva Math and MATH-500 take the last \boxed{} from the whole
output, so an answer inside the reasoning trace (including a trace that ran
to the token cap) gets credit. The new composable final_answer variant turns
on strip_thinking, strip_unclosed_thinking and thinking_prefilled: only the
text after the last </think> is scored, and a response without </think>
scores 0. Existing specs and their hashes are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
This branch has not been deployed
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Written by Claude Code session olmo-eval thinking stripping
AIME, HMMT, Minerva Math and MATH-500 extract the last
\boxed{}from the whole output, so an answer that appears only inside the reasoning trace gets credit, including in a trace that ran to the token cap. This adds a composablefinal_answervariant to each of them that scores only the text after the last</think>. A response without</think>scores 0.The variant reuses the existing flags; it adds no new scoring code:
strip_thinking(strip-thinking-task-option (#361))</think>strip_unclosed_thinking(add-strip-unclosed-thinking (#389))thinking_prefilled</think>counts as unterminated. This is needed because Olmo Think chat templates end the prompt in<think>, so a runaway trace contains neither tag.Use it only for thinking models. A model that answers without a trace also scores 0 under this variant.
It composes with the existing variants, e.g.
hmmt_feb_2026:pass_at_32:16k:final_answeroraime_2025:pass_at_32:final_answer.Affected specs: 16 new
:final_answerspecs (5 AIME, 3 HMMT, 7 Minerva Math subsets, MATH-500). Existing specs, suites and their hashes are unchanged;task_hashes.jsononly gains entries. No suite uses the new variant yet.Related: final-answer-only-scoring (#545) adds a similar opt-in gate as a separate scoring path for GPQA and other scorers.
Validation
The 4 failures are in
tests/scripts/test_check_file_sizes.pyand also fail on cleanmainin the same local environment. The new tests cover a trace that ran to the cap, a right answer inside the trace with a wrong final answer, and the reverse; they also check that the existing specs still credit the answer inside the trace. No model run was launched.🤖 Generated with Claude Code