SWE-Gym

Multi-Agent Systems 💻 Jupyter Notebook ⚖️ Apache-2.0 🔴 No recent commits
730 stars

SWE-Gym is an open environment for training software engineering agents and verifiers, providing 2.4K real-world tasks from 11 Python repositories with executable environments and test verification. It solves the lack of training environments for RL in repository-level coding, enabling agents to achieve state-of-the-art results on SWE-Bench. It is for researchers and developers training or fine-tuning language models for software engineering tasks.

✨ Key features

  • 2.4K real tasks from 11 Python repos
  • Lite split of 234 instances
  • Executable environments with test verification
  • Trains agents via RL and fine-tuning
  • Supports self-improvement via rejection sampling
  • Enables inference-time scaling with verifiers

🎯 Use cases

  • Train LM agents to solve real-world software engineering issues
  • Fine-tune models like OpenHands or MoatlessTools for SWE-Bench
  • Train verifiers for best-of-n selection in inference-time scaling
  • Reproduce state-of-the-art results on SWE-Bench Verified and Lite

📦 Installation

🧰 Requirements: Requires Python environment, access to Hugging Face for dataset, Docker for pre-built images, and API keys for models like GPT-4o or Claude if used for trajectory generation.

To access the SWE-Gym dataset, visit the Hugging Face hub page: https://huggingface.co/SWE-Gym. Environment constants are saved at the SWE-Bench-Fork repository: https://github.com/SWE-Gym/SWE-Bench-Fork. Pre-built Docker images are available under the prefix xingyaoww/sweb.eval.x86_64 on Docker Hub. For detailed experiment reproduction, see the docs: OpenHands and MoatlessTools.

🚀 Usage

For reproducing results with OpenHands or MoatlessTools, follow the instructions in the respective docs. For generating trajectories with OpenHands for training, refer to the instructions at: https://github.com/All-Hands-AI/OpenHands/blob/main/evaluation/benchmarks/swe_bench/SWE-Gym.md

⚠️ Good to know

The current results are primarily bottlenecked by training and inference compute, not the size of the environment.

❓ FAQ

What is SWE-Gym?

SWE-Gym is an open environment for training software engineering agents and verifiers, providing real-world tasks from Python repositories with executable environments and test verification.

How many tasks does SWE-Gym include?

It includes 2.4K real tasks from 11 Python repos, with a Lite split of 234 instances.

What results can be achieved with SWE-Gym?

Baselines achieve 32% and 26% on SWE-Bench Verified and Lite respectively, a new open state-of-the-art.

How can I access the dataset?

The dataset is available on Hugging Face at https://huggingface.co/SWE-Gym, with environment constants in the SWE-Bench-Fork repository.

📊 Repository

Stars★ 730
Forks🍴 45
Open issues🐛 9
Last commit🕒 Jul 29, 2025
Created📅 Nov 2024
Language💻 Jupyter Notebook
License⚖️ Apache-2.0

🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.