CodePin SFT contains 6,000 validated code-localization trajectories generated
with qwen3.5-35b-a3b. It is designed for pure-text supervised fine-tuning of
Qwen/Qwen3.5-0.8B and other tool-calling language models.
The tasks come from
LeeXugar/SWE-smith-code-search.
The source rollout dataset was used only for cleaning, difficulty estimation,
and sample selection; rollout messages and rewards were not copied into these
SFT… See the full description on the dataset page:
https://huggingface.co/datasets/LeeXugar/CodePin-SFT-Qwen3.5-35B-A3B.
