Supervised fine-tuning data for teaching a model to inject realistic, hard-to-find bugs into otherwise-correct code. The goal is to build coding RL environments where the introduced bugs resemble real human mistakes rather than trivial off-by-one or sign-flip errors.
The data is mined from real GitHub bug fixes: for each fix we take the after (correct) file as input and train the model to produce the before (buggy) version — i.e. the inverse of a… See the full description on the dataset page:
https://huggingface.co/datasets/nreHieW/proximal-bug-injection-sft.
