vovw/yam_folding

yam-folding

One combined cloth-folding corpus with 21,905 selected episodes: 4,838 Molmo training, 16,942 ABC training, and 125 ABC validation episodes. The episode index shown in the viewer is an index of demonstrations, not a frame-level replacement of the underlying robot trajectories.

Actual Parquet and video files are stored under sources/ in this repository. They are copied byte-for-byte from pinned upstream releases: Molmo 640x360 and ABC 224x224. There is no video re-encoding or resize. Source shards may contain unselected rows; episodes.jsonl authoritatively selects the included episodes. The corpus is a unified manifest plus loader, not one shape-homogeneous LeRobot root. Use the supplied adapter to load samples before official MolmoAct2 batching.

Train and validation roles remain separate, including ABC's validation repository whose internal split is named train. Uniform sampling over selected training episodes gives 22.213% Molmo and 77.787% ABC, then samples a frame within each episode. Camera order is top, left wrist, right wrist. Images pass through the pinned official model preprocessor in memory before collation.

The selection includes clothing, towels, and 875 napkin-folding demonstrations whose labels were included following representative cloth-material review. Moving/sorting cloth, pure rolling, and paper folding are excluded. See counts.json and task_audit.json for task-level details.

Sources

• Molmo, e9f21ae15074330839f2ac25ed4b49d76dfa1f9c.
• ABC train, 68651e4929d9fb00f798937b2d62617cab5c771d.
• ABC validation, 46ca817c39d3df08115a75c31b3cd3867e3d875a.

Source dataset cards declare Apache-2.0. Retain their attribution and citations. Original ABC is credited at https://huggingface.co/datasets/XDOF/ABC-130k. Only the already released converted backing files are copied here; raw MCAP, credentials, local logs and checkpoint weights are not uploaded.

Loading

See load_combined.py. It fetches the small index/loader files first, pins this repository revision, and lazily downloads backing files from this repository. Install official MolmoAct2 training code at commit 66b87e64efd99dfd103241418113955cf64dfa9c and its dependencies. The checkpoint is allenai/MolmoAct2-BimanualYAM at 8dcbed66f2380e4393189c303ea72488eb9e63c2.

Validation and limitations

The 26-episode, all-camera pilot passed official training preprocessing and collation. A real mixed-source pretrained forward/loss test passed with zero optimizer steps. Native videos were left unchanged.

Most ABC original UUIDs were not published in the converted sources and remain unresolved. Four validation towel trajectories were independently matched to original recordings. One original validation towel is absent from the converted validation copy. Molmo episode 31749 retains a source video-duration metadata discrepancy, although its requested pilot frames decode successfully.

Normalization is inherited from the pinned checkpoint (q01/q99, unscaled grippers), not claimed to be newly recomputed for this corpus. Whole-corpus content deduplication, new training-only statistics, and original timing-policy verification were deferred when scope narrowed to combining datasets.

Publication is complete only when publication_status.json reports complete.