Skip to content

AIME and GPQA treat num_examples=0 as the full dataset #294

Description

@sylvesterkaczmarek

Summary

The AIME and non-debug GPQA constructors guard subset selection with if num_examples:. Because zero is falsey, num_examples=0 silently skips sampling and evaluates the full dataset instead of zero examples.

This is inconsistent with the parameter contract (int | None, where None represents no restriction) and with HealthBench, which can represent an empty sample set.

Impact

A zero-example smoke/debug configuration can unexpectedly launch a full evaluation, causing substantial unnecessary inference work instead of returning an empty result.

Proposed resolution

Distinguish None from zero explicitly. Validate that sample counts are non-negative, use an empty example list for zero, and retain the existing subset sampling behavior for positive counts. Add regression coverage for zero and negative counts.

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