multi-swe-bench

👋 Hi, everyone!
We are ByteDance Seed team.

You can get to know us better through the following channels👇

🚀 Multi-SWE-bench: A Multilingual Benchmark for Issue Resolving

We are extremely delighted to release Multi-SWE-bench! Multi-SWE-bench addresses the lack of multilingual benchmarks for evaluating LLMs in real-world code issue resolution. Unlike existing Python-centric benchmarks (e.g., SWE-bench), our framework spans ​7 languages (i.e., Java, TypeScript, JavaScript, Go, Rust, C, and C++) with ​1,632 high-quality instances, curated from 2,456 candidates by ​68 expert annotators for reliability.

We aim to accelerate progress in automated issue resolution and RL, bridging the gap toward AGI. Let's join the Multi-SWE-RL community to expand datasets, tools, and research collaboration!

📢 News

[2025/04/15]🔥We released Multi-SWE-bench mini! A lightweight version of the full benchmark — 400 instances in total, covering 8 languages, designed to reduce compute cost and make evaluation faster and easier.

[2025/04/03]🔥We released Multi-SWE-bench and Multi-SWE-RL.

⚡ Features

• Comprehensive Evaluation: Evaluating nine powerful models (GPT-4o, OpenAI-o1, OpenAI-o3-mini-high, Claude-3.5-Sonnet, Claude-3.7-Sonnet, DeepSeek-V3, DeepSeek-R1, Qwen2.5-72B-Instruct, and Doubao-1.5-Pro) across three agent frameworks (Agentless, SWE-agent, OpenHands), yielding several valuable insights.
• Multi-SWE-RL Community: Open-source initiative for large-scale RL datasets. Initial release includes 4723 instances to advance RL research.
• Fully Open Source Data, Code, and Environment: All data, code, and container images are publicly released, along with detailed tutorials, to foster community contributions and enable scalable extension.

📊 Evaluation

Run Evaluation

To run the evaluation, you need to prepare the following:

• Patch Files: Some patch files in JSONL format, each item containing:

  • org: Organization Name
  • repo: Repository Name
  • number: Pull Request Number
  • fix_patch: Fix Patch Content

  Example:

  {
      "org": "zeromicro",
      "repo": "go-zero",
      "number": "2787",
      "fix_patch": "diff --git ...."
  }
  

• Dataset Files: Dataset files in JSONL format available on Hugging Face, such as Multi-SWE-bench or Multi-SWE-RL

• (Optional) Docker Images: You can download required Docker images using scripts/download_images.ps1 (for Windows) or scripts/download_images.sh (for Linux/macOS) with either verified images or RL images:

  # For Windows
  .\scripts\download_images.ps1 scripts\images_verified.txt  # For verified images
  .\scripts\download_images.ps1 scripts\images_rl.txt        # For RL images
  
  # For Linux/macOS
  bash scripts/download_images.sh scripts/images_verified.txt  # For verified images
  bash scripts/download_images.sh scripts/images_rl.txt        # For RL images
  

  This step is optional. If images don't exist locally, they will be built during evaluation.

Then you can run the evaluation using the following command:

python -m multi_swe_bench.harness.run_evaluation --config /path/to/your/config.json

The evaluation process will generate a final_report.json file in your specified output_dir, which provides a summary of results including resolved_instances, unresolved_instances, and other metrics. For detailed information about failed instances and specific error reasons, you can check the log files in the log_dir directory.

Configuration File Example

{
    "mode": "evaluation",
    "workdir": "./data/workdir",
    "patch_files": [
        "./data/patches/<your_patch_file>.jsonl"
    ],
    "dataset_files": [
        "./data/patches/<to_evaluate_dataset_file>.jsonl"
    ],
    "force_build": false,
    "output_dir": "./data/dataset",
    "specifics": [],
    "skips": [],
    "repo_dir": "./data/repos",
    "need_clone": false,
    "global_env": [],
    "clear_env": true,
    "stop_on_error": true,
    "max_workers": 8,
    "max_workers_build_image": 8,
    "max_workers_run_instance": 8,
    "log_dir": "./data/logs",
    "log_level": "DEBUG"
}

Configuration Parameters

Parameter	Description
mode	Execution mode for the script. Options: "evaluation", "instance", "instance_only", "image". Default: "evaluation"
workdir	Working directory path for evaluation operations
patch_files	List of patch file paths in JSONL format (supports glob patterns)
dataset_files	List of dataset file paths in JSONL format (supports glob patterns)
force_build	Whether to force rebuild Docker images even if they already exist
output_dir	Directory path for output results
specifics	List of specific PR IDs to evaluate (empty = all)
skips	List of PR IDs to skip during evaluation
repo_dir	Directory containing cloned repositories
need_clone	Whether repositories should be cloned if not present
global_env	Global environment variables to pass to Docker containers (format: "KEY=VALUE")
clear_env	Whether to clear environment variables in Docker containers
stop_on_error	Whether to stop execution when an error occurs
max_workers	Maximum number of concurrent worker threads for general tasks
max_workers_build_image	Maximum number of concurrent worker threads for building Docker images
max_workers_run_instance	Maximum number of concurrent worker threads for running instances
log_dir	Directory for log files
log_level	Logging level. Options: "DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"

✅ Integration Checklist

We are working to unify instances from prior benchmarks or training dataset into our framework for consistent comparison and reuse.

• Integrate 500 Python instances from SWE-bench verified
• Integrate 78 Java instances from SWE-bench-java
• Support customizing run.sh, test-run.sh, and fix-run.sh commands via configuration file using "run_cmd", "test_patch_run_cmd", and "fix_patch_run_cmd"
• Publish as a pip package for easier installation and reuse
• Integrate 2,438 Python instances from SWE-gym
• Integrate instances from R2E-Gym

🏆 Multi-SWE-RL Community

📋 Multi-SWE-RL Dataset Overview

The Multi-SWE-RL Community is an open-source initiative focused on collaborative dataset creation for software engineering and reinforcement learning research. To foster active participation and recognize contributors, we introduce this Contribution Incentive Plan. By contributing high-quality data, you directly support advancements in AI research and earn recognition within the community.

Incentive Tiers:

• Be a Contributor: Get listed in the Contribution Progress Sheet
• Report Authorship: Become an author in future technical reports

Full details: Contribution Incentive Plan

Get Started in 2 Steps:

• Learn: Quick-Start Guide
• Try: Follow our Contribution Demo

Welcome to our Discord to join in Multi-SWE-RL and Multi-SWE-bench related discussions!

🌟 Star Growth Trends

🙏 Acknowledgements

We express our deepest gratitude to the creators of the SWE-bench dataset. This project references their repository and builds upon their work.

📖 Citation

If you find Multi-SWE-bench useful for your research and applications, feel free to give us a star ⭐ or cite us using:

@misc{zan2025multiswebench,
      title={Multi-SWE-bench: A Multilingual Benchmark for Issue Resolving}, 
      author={Daoguang Zan and Zhirong Huang and Wei Liu and Hanwu Chen and Linhao Zhang and Shulin Xin and Lu Chen and Qi Liu and Xiaojian Zhong and Aoyan Li and Siyao Liu and Yongsheng Xiao and Liangqiang Chen and Yuyu Zhang and Jing Su and Tianyu Liu and Rui Long and Kai Shen and Liang Xiang},
      year={2025},
      eprint={2504.02605},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2504.02605}, 
}

📜 License

This project is licensed under Apache License 2.0. See the LICENSE file for details.

🏢 About ByteDance Seed Team

Founded in 2023, ByteDance Seed Team is dedicated to crafting the industry's most advanced AI foundation models. The team aspires to become a world-class research team and make significant contributions to the advancement of science and society.