Dataset Summary

Paper | Github | HF Collection

SWE-Lego-Real-Data contains 18k real github issues (Python language) and their multi-turn agent trajectories. The column named messages is collected using Qwen/Qwen3-Coder-480B-A35B-Instruct with OpenHands (v0.53.0) agent scaffolding, which can be directly used for SFT training.

Dataset Structure

.
└── data
    ├── resolved-00000-of-00001.parquet (5k github issues with resolved trajectories)
    └── unresolved-00000-of-00001.parquet (13k github issues with unresolved trajectories)

Effectiveness

The effectiveness of the dataset has been demonstrated by training exclusively with SFT from Qwen3-8B and Qwen3-32B, and evaluated on SWE-Bench-Verified:

• SWE-Lego-Qwen3-8B: 42.2% Pass@1, 49.6% TTS@16
• SWE-Lego-Qwen3-32B: 52.6% Pass@1, 58.8% TTS@16

We’ve open-sourced everything—our dataset, code, and training scripts, for everyone to progress on scaling and improving software engineering agents. For more details, please refer to our Github and Paper.

How to use

import json
from datasets import load_dataset

# Load dataset
ds = load_dataset("SWE-Lego/SWE-Lego-Real-Data", split="resolved")

# Select required columns
processed_ds = ds.select_columns(["instance_id", "messages"])
    
# Convert to list format
data_list = processed_ds.to_list()
    
# Save as JSON file
filename = "swe_lego_real_data_resolved_trajectories.json"
with open(filename, "w", encoding="utf-8") as f:
    json.dump(data_list, f, ensure_ascii=False, indent=4)
print(f"Saved {len(data_list)} records to {filename}")

Acknowledgement

SWE-Lego-Real-Data is built upon SWE-rebench, a large-scale dataset comprising 21k issue–pull request pairs.

Citation

Please cite our paper if you find the repo helpful in your work:

@misc{swelego,
      title={SWE-Lego: Pushing the Limits of Supervised Fine-tuning for Software Issue Resolving}, 
      author={Chaofan Tao and Jierun Chen and Yuxin Jiang and Kaiqi Kou and Shaowei Wang and Ruoyu Wang and Xiaohui Li and Sidi Yang and Yiming Du and Jianbo Dai and Zhiming Mao and Xinyu Wang and Lifeng Shang and Haoli Bai},
      year={2026},
      eprint={2601.01426},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2601.01426}, 
}