WDong/so101-erythromycin-tea-clean112-h50-v3

SO-101 erythromycin-on-tea clean112 H50 v3

Deterministically trimmed LeRobot v3 dataset for the SO-101 task:

Pick up the red erythromycin ointment box and place it on top of the green Rizhao tea tin.

This repository is the portable training input for the corrected OpenPI pi0.5 H50 experiments. It contains all 120 cropped episodes so that held-out evaluation remains self-contained, but training must use only the 112 episode IDs in meta/training_split_manifest.json.

Data contract

• 17,262 frames at 30 Hz.
• Two RGB cameras: observation.images.fixed and observation.images.wrist, 640x480.
• Six calibrated SO-101 joint positions for observation.state and action.
• Stored action is joint-absolute. The OpenPI transform uses current-state delta for arm joints 0:5 and absolute gripper for joint 5.
• Action horizon: 50. Temporal padding must be excluded from flow loss with action_is_pad.
• Both-side trimming: 2-frame pre-roll and 15-frame post-roll around detected motion.
• Train: 112 episodes / 15,693 frames.
• Held-out validation: episodes 20, 47, 89 / 415 frames.
• Recovery episodes excluded from clean training: 7, 12, 58, 72, 84 / 1,154 frames.
• meta/stats.json was computed from the clean112 training subset only; there is no validation leakage.

Verified H50 mask statistics

Split	Frames	Valid targets	All targets	Valid fraction	First-step padding
train clean112	15,693	647,450	784,650	0.8251449691	0
held-out 20/47/89	415	17,075	20,750	0.8228915663	0

The terminal transformed sample was verified to retain one valid target and 49 padded targets. The full audit decoded both camera streams and checked six pixel pairs without re-encoding the source videos. See audits/ and trim_derivation_receipt.json.

Download

huggingface-cli download WDong/so101-erythromycin-tea-clean112-h50-v3 \
  --repo-type dataset \
  --local-dir /path/to/so101_erythromycin_on_tea_grid120_bothtrim_clean112_preroll2_h50_v3

Or:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="WDong/so101-erythromycin-tea-clean112-h50-v3",
    repo_type="dataset",
    local_dir="/path/to/so101_erythromycin_on_tea_grid120_bothtrim_clean112_preroll2_h50_v3",
)

Verify every downloaded file with portable_manifest.json before training. Do not silently fall back to all 120 episodes: the corrected OpenPI loader loads the complete dataset first to preserve global frame coordinates, constructs H50 actions/masks, and only then wraps it with a PyTorch Subset using meta/training_split_manifest.json.

Provenance

• Source dataset: WDong/so101-erythromycin-tea-grid120-v30.
• OpenPI code branch: so101-h50-masked-ki-v2, commit a3fee39.
• Orchestrator code commit used for the remote audit: ea68926 (with portable hardlink support added for this publication).
• Dataset derivation is lossless with respect to videos: videos were not re-encoded.