swesmith 0.0.0

A package for generating LLM based agent trajectories for software engineering

---

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 (bug reports with solution validation) for any Python repository.
• Generate trajectories of SWE-agent (running on top of an LM like GPT-4o or Claude 3.7) solving those task instances.
• Train local LMs on these trajectories to improve their performance on software engineering tasks (resulting in models like SWE-agent-LM-32B)

🚀 Get Started

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

💿 Resources

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 40.2% pass@1 on SWE-bench Verified!
• 5k Trajectories that SWE-agent-LM-32B was trained on.

And there's more coming!

💫 Contributions

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

Contact Person: John Yang, Kilian Lieret, Carlos E. Jimenez (Email: [e-mail])

🪪 License

MIT. Check LICENSE for more information.

✍️ Citation

@misc{yang2025swesmith,
  title={SWE-smith: Scaling Data for Software Engineering Agents},
  author={John Yang and Kilian Lieret 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},
  archivePrefix={arXiv},
  primaryClass={cs.SE},
}