Skip to content

Would an EvalPort-format export of MMLU-Pro be welcome in EvalPort's benchmarks/? #81

Description

@adhabnr-ux

Hi MMLU-Pro team — I maintain EvalPort, an open JSON spec for portable LLM eval datasets and results. benchmarks/ already has the original MMLU converted; MMLU-Pro is different enough (10 options instead of 4, CoT-focused, 14 domains) that I think it's worth its own suite rather than folding into the existing MMLU one.

I read the real fields this repo's own eval code (evaluate_from_local.py) uses: question, options (a list, up to 10, with "N/A" entries filtered out), answer (letter), answer_index, category (the 14 domains), cot_content.

That maps onto EvalPort cleanly: question + options -> TestCase.input, answer/answer_index -> TestCase.expected_output, category -> metadata for per-domain breakdowns, graded with exact_match or a small custom grader mirroring your own extract_answer regex chain.

Would your maintainers be open to an EvalPort-format export living in EvalPort's benchmarks/, with attribution, license, and a link to the HF dataset/leaderboard preserved? Happy to do the conversion and open a PR for review first. No worries if this isn't useful.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions