MEnvData-SWE: Polyglot Software Engineering Dataset with Executable Environments

📋 Dataset Description

MEnvData-SWE is the largest open-source polyglot dataset of realistic verifiable Docker environments, comprising 3,005 task instances from 942 repositories across 10 programming languages. Each instance includes a fully executable Docker environment with pre-verified test cases, environment setup scripts, and evaluation scripts.

Key Features

• 🌐 Multi-Language Support: 10 programming languages (Python, Java, TypeScript, JavaScript, Rust, Go, C++, Ruby, PHP, C)
• 🐳 3,005 Docker Images: Pre-built environment images with verified dependencies
• ✅ Verified Quality: Each instance includes Fail2Pass validated test cases
• 🔧 Complete Setup Scripts: Both incremental and full environment configuration scripts
• 📝 Evaluation Scripts: Ready-to-use verification scripts with exit code validation
• ⚡ Efficient Verification: Exit code-based validation (no complex log parsing required)

Dataset

🐳 Docker Image Registry

All pre-built Docker images are publicly available on Docker Hub:

🔗 Docker Hub Repository: https://hub.docker.com/u/mcatwj

Each instance's image_name field references a specific image that can be pulled directly:

# Example: Pull a specific instance image
docker pull mcatwj/swe-images-c:systemd-systemd-pr-24645

# Or use the image_name from the dataset
docker pull mcatwj/<image_name_from_dataset>

All images are fully configured with:

• Project dependencies installed
• Repository cloned at /testbed
• Ready for immediate test execution

📊 Dataset Structure

MEnvData-SWE extends the standard SWE-Bench schema with executable environment configurations. Each instance contains the following fields:

🚀 Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("ernie-research/MEnvData-SWE")

# Access a single instance
instance = dataset['train'][0]
print(f"Repository: {instance['repo']}")
print(f"Language: {instance['language']}")
print(f"Docker Image: {instance['image_name']}")

Example Instance

{
    "repo": "systemd/systeinstance_id": "systemd__systemd-24645",
    "language": "C",
    "image_name": "swe-images-c:systemd-systemd-pr-24645",
    "base_commit": "6d64cb0625691e2b9eda8babe07ac8281f9467ee",
    "env_setup_script": "#!/usr/bin/env bash\nset -e\ncd /testbed\ngit reset --hard 6d64cb0625...",
    "eval_script": "#!/usr/bin/env bash\nset -uxo pipefail\nexport LANG=en_US.UTF-8...",
    ...
}

🎯 Use Cases

1. Agent Training: Train code generation and debugging agents on realistic software engineering tasks
2. Benchmark Evaluation: Evaluate model performance on multi-language software issues

🔗 Related Datasets

• MEnvData-SWE-Trajectory: Extended version with 3,872 agent execution trajectories

📖 Citation

If MEnvData-SWE helps your research, please cite:

@misc{guo2026menvagent,
      title={MEnvAgent: Scalable Polyglot Environment Construction for Verifiable Software Engineering}, 
      author={Chuanzhend Jingjing Wu and Sijun He and Yang Chen and Zhaoqi Kuang and Shilong Fan and Bingjin Chen and Siqi Bao and Jing Liu and Hua Wu and Qingfu Zhu and Wanxiang Che and Haifeng Wang},
      year={2026},
      url={https://arxiv.org/abs/2601.22859}, 
}

📧 Contact

For questions or issues:

• GitHub: MEnvAgent Repository
• Email: czguo@ir.hit.edu.cn

🙏 Acknowledgements

We thank all open-source maintainers whose projects contributed to this dataset.