mtybilly/RollbackTraj-Qwen-3.5

RollbackTraj On-policy rollback trajectories for training a parameterized <UNDO>N</UNDO> command that deletes the last N visible tokens. sampled ships already cut at the last sentence boundary for prose, or newline boundary for code (the_stack), so consumers need no sentence segmentation. How these rows were made Qwen/Qwen3.5-0.8B-Base free-samples a continuation from a document prefix. The prompt was the bare prefix, with no leading BOS token. A row is clean when… See the full description on the dataset page: https://huggingface.co/datasets/mtybilly/RollbackTraj-Qwen-3.5.

<iframe src="https://silytics.si/embed/item/5b656a1b-8a62-48e2-8439-443c8836f303" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.