RollbackTraj contains on-policy rollback trajectories for training a parameterized
N command that deletes the last N visible tokens.
Version 2 ships sampled already cut at the last sentence boundary for prose or
newline boundary for code (the_stack). Consumers therefore receive complete rollback
spans without running sentence segmentation again.
CRITICAL: divergence_index and n_undo are TOKEN counts under… See the full description on the dataset page:
https://huggingface.co/datasets/mtybilly/RollbackTraj.
