RL dataset for training SWE-Swiss models on the repair task. The prompts are based on issues from SWE-Gym and SWE-smith. To create a challenging task, the code content in each prompt consists of two components: "oracle" files, which are the ground-truth files requiring a patch, and "distractor" files, which are plausible but incorrect files predicted by an LLM.
@misc{SWESwiss2025,
title = {SWE-Swiss: A Multi-Task Fine-Tuning and RL Recipe for… See the full description on the dataset page:
https://huggingface.co/datasets/SWE-Swiss/SWESwiss-Repair-RL-SWEGym-SWESmith-12K.
