Skip to content

Score the latest attempt when resuming a benchmark - #9

Open
ItsM0rty wants to merge 1 commit into
InterfazeAI:mainfrom
ItsM0rty:code-fix/resumed-scores
Open

ItsM0rty wants to merge 1 commit into
InterfazeAI:mainfrom
ItsM0rty:code-fix/resumed-scores

Conversation

@ItsM0rty

@ItsM0rty ItsM0rty commented Nov 2, 2026

Copy link
Copy Markdown

Fixes #8.

A failed request stays in the response log after a successful retry. GPQA then counts both rows, so a correct answer can lower the score.

This keeps the full log and uses the latest attempt per sample for scoring and resume decisions. The regression runs the CLI twice with a fake provider and the real GPQA scorer, including repeated failures before success. It also covers a different sample ID key and preserves the original records.

26 focused tests pass, with four archived-result tests skipped. Ruff checks pass for the changed files. No model calls were made, and I haven't checked published scores.

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.

Resumed benchmark runs can count failed attempts in the score

1 participant