SWE-smith is a toolkit for training software engineering (SWE) agents. With SWE-smith, you can:

• Create an unlimited number of SWE-bench style task instances for any Python repository.
• Generate trajectories of SWE-agent solving those task instances.
• Train local LMs on these trajectories to improve their software engineering capabilities (SWE-agent-LM-32B).

Check out the documentation for a complete guide on how to use SWE-smith, including how to

• Install the repository locally or as a PyPI package.
• Create Task Instances for any Python repository with SWE-smith.
• Use your task instance to train your own SWE-agents

Install the repo:

git clone https://github.com/SWE-bench/SWE-smith cd SWE-smith conda create -n smith python=3.10; conda activate smith; pip install -e .

Then, check out scripts/cheatsheet.sh for scripts to (1) create execution environments, (2) create task instances, and (3) train SWE-agents.

In addition to this toolkit, we've also provided several artifacts on the SWE-bench HuggingFace, including:

• 50k Python Task Instances, created using SWE-smith.
• SWE-agent-LM-32B, trained using SWE-smith. Achieves 41.6% pass@1 on SWE-bench Verified!
• 5k Trajectories that SWE-agent-LM-32B was trained on.

And there's more coming!

Excited about SWE-smith? We're actively working on several follow ups, and love meaningful collaborations! What we're thinking about...

• Make SWE-smith work for non-Python languages
• New bug generation techniques
• Train SWE-agents with more trajectories and new methods

Check out the Contributing Guide for more.

Contact Person: John Yang, Kilian Lieret (Email: johnby@stanford.edu)

MIT. Check LICENSE for more information.

@misc{yang2025swesmith, title={SWE-smith: Scaling Data for Software Engineering Agents}, author={John Yang and Kilian Leret and Carlos E. Jimenez and Alexander Wettig and Kabir Khandpur and Yanzhe Zhang and Binyuan Hui and Ofir Press and Ludwig Schmidt and Diyi Yang}, year={2025}, eprint={2504.21798}, archivePrefix={arXiv}, primaryClass={cs.SE}, url={https://arxiv.org/abs/2504.21798}, }