Verify

Prove from the parquet on disk that a dataset holds the episodes you intended, with frames, matching videos and live control columns.

At the end of this page you can prove, from the parquet on disk and not from anyone's narration, that a recorded dataset holds the episodes you intended, with frames in every one, video files that match, and no dead control column.

This runs without lerobot; only pyarrow is needed. A toy dataset with three episodes:

import json, pathlib
import pyarrow as pa, pyarrow.parquet as pq
from strands_robots.verify_dataset import verify_dataset

root = pathlib.Path("/tmp/toy_dataset")
(root / "meta" / "episodes" / "chunk-000").mkdir(parents=True, exist_ok=True)
pq.write_table(pa.table({"episode_index": [0, 1, 2], "length": [90, 90, 90]}),
               root / "meta/episodes/chunk-000/file-000.parquet")
(root / "meta/info.json").write_text(json.dumps({"total_episodes": 3, "total_frames": 270, "fps": 30, "features": {}}))

print(verify_dataset(root, expected=3)["status"])        # success
print(verify_dataset(root, expected=20)["problems"][0])
# expected 20 episode(s) but parquet holds 3 - the recording did not produce the intended number of distinct episodes

The command

strands-robots verify-dataset ~/.cache/huggingface/lerobot/you/so101_reach --expected 5

[PASS] /Users/you/.cache/huggingface/lerobot/you/so101_reach
  episodes (parquet): 5
  frames   (parquet): 450
  expected episodes : 5
  info.json episodes: 5
  info.json frames  : 450

Exit code 0 on success, 1 on any problem. Flags:

flag	meaning
--expected N	require exactly N distinct episodes
--min-frames K	every episode must hold at least K frames (default 1; 0 disables)
--no-check-videos	skip the per-episode video file checks
--no-check-stats	skip the dead-control-column check
--json	the report as JSON

What it checks

1. meta/episodes/**/*.parquet exists and holds at least one distinct episode_index; an empty directory reports "The dataset is empty or was never finalized (episodes are flushed to parquet at stop_recording/finalize)".
2. Every episode has at least min_frames frames.
3. meta/info.json's total_episodes and total_frames agree with the parquet; a header that is not a count, or a file that is not a JSON object, is corrupt metadata, not missing metadata.
4. With --expected, the parquet holds exactly N episodes: the mega-episode check, for a run that buffered every frame into episode_index=0 and has the right frame total with the wrong episode count.
5. Every video file the dataset references (resolved from info.json's video_path template and the episode parquet's videos/<key>/chunk_index and file_index columns) exists, is non-empty, and, when the container header can be read, holds exactly the frames the parquet maps into it.
6. No episode's action or observation.state column is identically zero. A multi-robot vector is split into the per-robot blocks info.json declares; a wholly zero block is flagged, a zero subset of one block (a gripper parked all episode) is a measurement and is not.

The parquet is the ground truth; the checker never trusts an agent's "recorded 20/20" or the recorder's in-memory counters.

From the simulation

sim.verify_dataset_episodes(expected=5) runs the same reader against the session's own dataset root; make it the last line of a collection script. read_dataset_episode_indices(root) in strands_robots.dataset_metadata is the one reader behind the sim facade, the CLI and the episode judge, so they cannot disagree.

The report

verify_dataset() returns a dict: status, ok, root, total_episodes, total_frames, episode_indices, frames_per_episode, expected, info_total_episodes, info_total_frames, video_files_checked, problems. A partially corrupt dataset reports the readable files and names each broken one in problems.

Next: label and judge.

Edit page