Code, converters, split audit, and benchmark reports:
https://github.com/Jatshi/SafeVLA-Bench.
pickcube_state_h16_full1000_holdout_v1.npz: 62,681 state/action-chunk samples from 1,000 CPU replays after excluding the five evaluation-seed episodes.
pickcube_state_h16_v1.npz: 6,220 state/action-chunk samples from the first 100 CPU replays; retained only for diagnostic comparison, not the final result.… See the full description on the dataset page:
https://huggingface.co/datasets/jatshi/SafeVLA-Bench-Dataset.
