Feature request: integrity checks for language annotation columns (v3.0)
Motivation
v3.0 datasets can carry language_persistent / language_events columns, but as far as I can
tell nothing validates them on write or on load. A dataset can be structurally valid and still
carry annotation that is silently misaligned with the trajectory — and this does not show up in
any loss curve.
I ran a set of checks over one full data file of lerobot/FineART
(data/chunk-000/file-000.parquet: 242 episodes, 2,108 subtask rows). All numbers below are
measured, not estimated.
What I checked, and what I found
| Check | Result |
|---|---|
Row count vs meta/episodes subtask_count |
2,108 = 2,108, 0 mismatches |
| Timestamps monotonic / no overlap | 0 episodes with issues |
Empty content |
0 rows |
| Start time of first subtask | p50 2.20s, p75 3.18s, p90 5.19s, p95 6.60s, p99 9.73s, max 14.10s |
| Episodes with no active subtask for >1s / >3s / >5s / >10s | 235 (97%) / 62 (26%) / 27 (11%) / 3 (1%) |
| That window as a fraction of episode length | p50 1.93%, max 15.88% |
language_events |
empty in 242/242 episodes (matches the card: "unused (empty)") |
Two takeaways:
- The annotations themselves are clean — counts, ordering and non-emptiness all check out.
- But the leading no-active-subtask window is longer and more spread out than the FineART card
suggests. The card says "the first ~2 s of an episode"; measured, 26% of episodes exceed 3 s
and 11% exceed 5 s. Anyone writing a data loader aroundactive_at()has to handle theNone
branch — it is the common case, not an edge case.
Proposal
A small, opt-in validation for language columns, usable both as a script and as a load-time check:
- row count agrees with episode-level metadata
- timestamps monotonic, non-overlapping, within
[0, episode_duration] contentnon-empty- report the coverage profile (leading / trailing unannotated duration, max segment length)
I am not proposing this as a hard gate — a warning-level report is enough, and it keeps the
"structural" and "annotation" layers separate.
Context / disclosure
I build RDA (pip install robot-data-audit, github.com/liesliy/rda), an open-source auditor for
robot datasets. It currently covers the trajectory and video layers (missing frames, NaN/Inf,
video-vs-parquet frame count, freeze detection, temporal alignment); the annotation layer is a
gap — I found everything above by writing throwaway scripts, which is exactly why I think this
belongs in LeRobot itself. Happy to contribute the check if you think it fits, and happy to be
told it does not.