swesmith · Python Simple Repository Browser

Project Links

Meta

Author: John Yang

Requires Python: >=3.10

Classifiers

Programming Language

Python :: 3.10
Python :: 3.11
Python :: 3 :: Only

License

OSI Approved :: MIT License

Operating System

OS Independent

NeurIPS 2025 Datasets & Benchmarks Track - Spotlight 🔦

SWE-smith is a toolkit for training SWE-agents. You can:

Turn any Github repository into a SWE-gym.
Create unlimited tasks (e.g., file localization, program repair, SWE-bench) for that repo.
Train an LM to become a better SWE (SWE-agent-LM-32B).

⚒️ Build Environments

If you're interested in turning a GitHub repository into a SWE-gym, install the package from source.

[!TIP] SWE-smith requires Docker to create execution environments. SWE-smith was developed and tested on Ubuntu 22.04.4 LTS. We do not plan on supporting Windows or MacOS.

You can then build a dataset for the repository by...

🏋️ Train SWE-agent's

Training SWE-agent's using the SWE-smith dataset is super simple.

from swesmith.profiles import registry
from datasets import load_dataset
ds = load_dataset("SWE-bench/SWE-smith", split="train") # Loads all 52k task instances
for task in ds:
    rp = registry.get_from_inst(task)  # Get the RepoProfile for the task
    container = rp.get_container(task) # Returns pointer to a Docker container with the task initialized

    """TODO: Train!"""

SWE-smith has been used to

Fine-tune Qwen 2.5 Coder into SWE-agent-LM-32B (A +32% jump on SWE-bench Verified!) using SWE-agent [Tutorial]
Perform GRPO style reinforcement learning using SkyRL

💿 Resources

52k Task Instances
SWE-agent-LM-32B; 40.2% pass@1 on SWE-bench Verified!
26k SWE-agent Trajectories, including the 5k SWE-agent-LM-32B was trained on.
250+ Environments, one Docker image per repo represented in SWE-smith.

And there's more coming!

💫 Contributions

We're actively working on several follow ups! Check out the Contributing Guide for more.

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

🪪 License

MIT. Check LICENSE for more information.

✍️ Citation

@inproceedings{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},
  booktitle = {Proceedings of the 39th Annual Conference on Neural Information Processing Systems (NeurIPS 2025 D&B Spotlight)},
  year={2025},
  eprint={2504.21798},
  archivePrefix={arXiv},
  primaryClass={cs.SE},
  url={https://arxiv.org/abs/2504.21798},
  note={arXiv:2504.21798, accepted at NeurIPS 2025 (Spotlight)}
}