Audit note: 6 episodes, one doc bug, one unvalidated layer
I audited the first six episodes of FineART (episodes 0–5, four cameras, 13,949 frames) on real downloaded files — the structural layer, plus a hand-written probe of the annotation layer. Full write-up, commands, and raw output: https://liesliy.github.io/rda-audits/datasets/fineart.html
Posting the parts that are actionable for this dataset rather than the parts that are flattering.
The structural layer is clean. All six episodes pass schema conformance, NaN/Inf, timestamp validity, drop-out, and video frame-integrity checks. The v3.0 shared-MP4 slicing reads correctly, including the per-camera from_timestamp / to_timestamp offsets — that detail broke the first two loaders I tried, so it is worth recording that it is right here.
The Quick start does not run. LeRobotDataset("lerobot/fineart-dataset") fails: that repo id does not exist (404). The dataset is lerobot/FineART (lerobot/fineart 307-redirects to it), and the README uses the non-existent id in four places, including the first line of the Quick start. PR #4 changes it.
The annotation layer is unvalidated. In data/chunk-000/file-000.parquet (242 episodes, 2,108 segments) the labels are internally consistent: segment counts match meta.subtask_count with zero mismatches, timestamps never invert, nothing is blank, and language_events is empty in 242 of 242 episodes while language_persistent carries the segments. That asymmetry is probably intended — but I could not find anything that validates these columns at all, so nothing would catch it if it were a bug. Filed separately as #5.
One convention worth stating in the card: the first segment starts at a median 2.20 s (p90 5.19 s, max 14.10 s; 235 of 242 episodes start later than 1 s). A loader that resolves "the active subtask = most recent segment at or before t" has no answer for those opening frames. That branch fires in almost every episode. It is a layout property, not a defect — cheaper to document than to rediscover.
What this does not say. Six episodes out of 40,543 characterise format handling and mechanism, not the defect distribution of the corpus. Nothing here is a claim about the other 40,537 episodes. And no automated check in this audit says whether a label is true.
Tool disclosure. The auditor is robot-data-audit, which I maintain (MIT). It got one thing wrong on this data: a video-freeze detector that excluded 5 of 6 healthy episodes on a false positive. I re-measured from scratch, confirmed the tool was at fault, changed the criterion, and released the fix today (0.12.2). The write-up keeps the wrong result next to the correction, because a tool that only publishes its hits is not worth reading.