pmd/README.md

# LSV PMD Split Benchmark Dataset

PMD split clips provide short pipetting examples for correct-vs-mistake and mistake-type evaluation.

## Layout

- `correct/`: split correct demonstration clips
- `incorrect/`: split mistake demonstration clips
- `pmd.parquet`: split-clip benchmark manifest

## Counts

- split clips: `246`
- by label: `{'correct': 15, 'incorrect': 231}`
- by mistake_type: `{'early_release': 45, 'none': 15, 'reuse_same_media': 27, 'reuse_same_tip': 4, 'reuse_tip': 53, 'surface_contamination': 44, 'wrong_tip_for_pipette': 58}`

## Review Flags

- needs_review clips: `6`

The `needs_review` and `notes` columns preserve human review notes for a small number of clips whose source labels required normalization.

## Evaluation

Use `pmd/pmd.parquet` with `lsvbench/tasks/prompts/pmd.md` or the `pmd` task in `lsvbench`.

The prompt preserves detailed answer options (`ERROR_REUSE`, `ERROR_SURFACE`, `ERROR_RELEASE`, `ERROR_INSTALL`, `ERROR_OTHER`) for compatibility with existing model outputs. The public benchmark report collapses all non-`CORRECT` options into binary `ERROR` and reports only **Error Detection** metrics.