Calibra — Dataset Integrity
Before diversity or coreset selection, can you trust this dataset?
Enter a LeRobot dataset ID. Calibra checks Integrity first — timestamp consistency, episode completeness, duplicate frames, camera freeze, blur, jittery/jerky motion — then Quality and Coverage. Pre-checked datasets return instantly from the benchmark cache.
Try these
What gets checked
| Layer | Checks | Answers |
|---|---|---|
| Integrity (first) | Timestamp consistency, sensor sync, episode completeness, duplicate frames, camera freeze, blur, jerky/jittery motion (LDLJ, jerk spikes, velocity discontinuities) | Can I trust this dataset? |
| Quality | Action-state tracking error, scripted-vs-teleop motion signature | Is this data clean? |
| Coverage | Trajectory diversity, redundancy fraction, entropy | Does my robot see enough variety? |
| Task Structure | Episode length distribution, phase balance, inactivity periods | Are episodes complete and well-formed? |
After Integrity comes Quality, Coverage, and — for building smaller training sets — Optimization.
Full check locally (all episodes, per-episode verdicts, certifiable report):
pip install calibra-robotics
calibra integrity hf://lerobot/pusht
calibra audit hf://lerobot/pusht # quality + coverage scoring
Powered by Calibra — open-source robotics dataset observability