act-bench-datasets

# ACT-Bench Datasets Private ACT-Bench benchmark datasets staged by suite. Manifests contain repo-relative `*_repo` fields for all packaged GT/reference assets. | Suite | Rows | Unique assets | Size | Included GT/reference assets | | --- | ---: | ---: | ---: | --- | | `suite1_action_perturbation` | 500 | 3502 | 8.31 GB | large_hdf5, large_video, metadata, original_hdf5, original_video, panel_video, small_hdf5, small_video | | `suite2_object_binding` | 583 | 2335 | 12.16 GB | gt_output_hdf5, gt_output_video, metadata, receiver_original_hdf5, receiver_original_video | | `suite3_cross_action` | 2500 | 10003 | 48.58 GB | cross_action_gt_hdf5, cross_action_gt_video, donor_hdf5, donor_video, metadata, receiver_original_hdf5, receiver_original_video | | `suite4_primitives` | 500 | 2091 | 5.20 GB | contact_annotation_hdf5, metadata, primitive_gt_hdf5, primitive_gt_video, source_original_hdf5, source_original_video | ## Layout Each suite contains: - `manifest.jsonl`: selected benchmark rows with repo-relative asset paths. - `summary.json` when available. - `assets/`: GT simulator videos/HDF5 and reference original/donor/contact assets used by the evaluators. The bundle intentionally excludes stale intermediate failed/rejected generation rows.