README.md

Merged VLA LeRobot Dataset

This dataset merges all integrity-valid sessions in the source directory.

Selection rules:

• invalid datasets from integrity_report/report.json are excluded;
• lyh_grasp_4obj_50_s2 keeps source episodes 0..44 only;
• lyh_grasp_4obj_50_s6 is excluded because episode_timestamps.json is missing.

The dataset contains 1,445 episodes and 863,976 frames. The original videos and parquet rows are kept in LeRobot-style directories. episode_manifest.jsonl is the join point for external per-session files. Join on external_join_key = subject|session|source_episode_index.

Each data row also includes source_dataset, source_subject, source_session, source_episode_index, and source_frame_index.

The sources/<subject>/<session>/ directories retain the original timestamp sidecars. meta/tasks.jsonl, meta/episodes.jsonl, and meta/episodes_stats.jsonl are included for newer LeRobot tooling; parquet metadata files are retained as well.

Local inspection

import pyarrow.parquet as pq

table = pq.read_table("data/chunk-000/file-000.parquet")
print(table.num_rows, table.column_names)

For external data, read episode_manifest.jsonl, replace the external_data_file placeholder with the actual per-session file path, and join by external_join_key.