Candidate hand-action instructions generated with
Qwen/Qwen3.5-122B-A10B-FP8 from the VITRA Appendix A.1 first-round prompt.
The second round produces five meaning-preserving imperative rephrasings using
the prompt in the VLA-HAND episode_cutting pipeline.
This is a sidecar dataset. It does not modify or replace the source VITRA
episodes. Every row contains exact source repository/revision, annotation and
video paths, hand, half-open frame… See the full description on the dataset page:
https://huggingface.co/datasets/MIT-Media-Lab/vitra-instructions-qwen35-122b-egocentric-v1.
