Skip to content

HealthBench bootstrap uncertainty is not reproducible #260

Description

@sylvesterkaczmarek

Summary

HealthBench reports bootstrap_std for every aggregated metric, but _compute_clipped_stats(..., "bootstrap_std") draws bootstrap samples from NumPy's process-global RNG without a fixed/local seed.

As a result, the same set of per-example scores can produce different reported uncertainty values across otherwise identical runs.

Current behavior

bootstrap_samples = [np.random.choice(values, len(values)) for _ in range(1000)]

The output therefore depends on whatever global NumPy RNG state happens to exist when aggregation runs.

Impact

HealthBench result JSON can change even when the sampled model outputs and grader results are identical. This makes reported uncertainty harder to reproduce, compare, and regression-test.

Proposed resolution

Use a local deterministic NumPy generator for bootstrap resampling so aggregation does not depend on or mutate process-global RNG state.

Add a regression that changes the global NumPy seed between calls and verifies the bootstrap standard deviation for the same values is unchanged.

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