SWE-Fixer: Training Open-Source LLMs for Effective and Efficient GitHub Issue Resolution

SWE-Fixer is an open-source framework for resolving real-world GitHub issues, comprising a code file retrieval module (BM25 plus a lightweight model for coarse-to-fine retrieval) and a code editing module that generates patches for the retrieved files. We compile a dataset of 110K GitHub issues with their corresponding patches and train the two models separately.

SWE-Fixer achieves competitive performance among open-source models on SWE-Bench Lite and Verified, while requiring only two model calls per instance.