SafetyJev Unsafe600
Delivery is in progress. No completed dataset revision is available yet.
This fixed cohort contains 600 recorded unsafe robot episodes, original five-view JPEG HDF5 observations and snapshots, and shared semantic annotations for the SafetyJev classifier and action-conditioned Predictor Judge. Original benchmark GT is retained separately from derived safety supervision.
The final manifest will identify all files, the approved task-group split, and
acceptance results. Source media is shared under sources/; model indexes are
under classifier/ and predictor_judge/. Large training caches are not included.
Wait for the completed manifest and pinned revision before training. Detailed preparation, training and evaluation commands are maintained in the SafetyJev codebase. This unsafe-only cohort does not estimate false-alarm rates on safe episodes.
- Downloads last month
- -