Repository navigation
fix(eval): track unhandled scores per row for the error score handler - #909
Open
Raphael Fakhri (RaphaelFakhri) wants to merge 1 commit into
Open
Raphael Fakhri (RaphaelFakhri) wants to merge 1 commit into
Raphael Fakhri (RaphaelFakhri) wants to merge 1 commit into
Conversation
unhandled_scores was shared across every row of an eval, so a row whose task raised used whatever the last finished row left behind. The error score handler then ran with a partial scorer list or not at all. Track it per row, matching the JavaScript SDK.
Codex Review SummaryThis 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
_run_evaluator_internal_implkeepsunhandled_scoresin the enclosing function and updates it withnonlocalfrom every row. When a row's task raises, it skips the scorers and reads whatever the last finished row stored there:None, soerror_score_handlernever runs and the failed row has no scores.With
error_score_handler=default_error_score_handler, failed rows drop out of the score averages instead of counting as 0. This change keepsunhandled_scoreslocal to each row, which matches the JavaScript SDK.Changes
unhandled_scoresinsiderun_evaluator_taskand drop thenonlocaldeclaration.Testing
Before the change: 2 failed (
{} == {'exact_match': 0, 'picky': 0}and{'picky': 0} == {'exact_match': 0, 'picky': 0}). After the change: 2 passed.python -m pytest src/braintrust/test_framework.py src/braintrust/test_framework2.py src/braintrust/test_workflow_eval.py: 185 passed, 5 skipped.