Assets

In addition to the paper and codebase, we release the following assets created with SWE-smith:

1. Environments for 128 GitHub repositories. You can download the environments (Docker images) locally by running the following command from the root directory of SWE-smith:

   python swesmith/build_repo/download_images.py
   

2. SWE-smith dataset of 50k+ task instances, made available as a HuggingFace dataset.

3. 5k expert trajectories + SWE-agent-LM-32B. To create SWE-agent-LM-32B, we fine-tuned Qwen 2.5 Coder Instruct 32B on the 5k trajectories. SWE-agent-LM-32B achieves 40.2% pass@1 on SWE-bench Verified. The trajectories are uploaded to a HuggingFace dataset. We also release the 32B and 7B versions of the model.

4. SWE-Rater-32B, a Qwen 2.5 Coder Instruct 32B model fine-tuned on human annotated ratings of a SWE-bench task instance's difficulty. We release it as a HuggingFace model.

Affiliated projects

SWE-bench SWE-agent SWE-ReX sb-cli