---
pretty_name: MMAD Reproducible Sampled Benchmark
task_categories:
- visual-question-answering
---

# MMAD

Reproducible packaging of MMAD pinned to `c4ed190dcb530f2f673ab293e575ad32054bb3cf`.

- `metadata/images.jsonl`: one row per query image.
- `splits/full.jsonl`: complete benchmark, one row per question.
- `splits/sample-n*-seed*.jsonl`: image-level stratified samples with a
  deterministic normal template and deterministic option order.
- `assets/`: source-preserving image tar shards (uploaded separately).

Default evaluation uses all five sampled splits and reports mean, standard
deviation and a confidence interval. The full split is opt-in.