ernie-research/MEnvData-SWE-Trajectory

---
language:
- en
license: apache-2.0
size_categories:
- 1K<n<10K
task_categories:
- text-generation
pretty_name: MEnvData-SWE-Trajectory
tags:
- code
- software-engineering
- agent
- trajectory
- multi-language
- claude
- openhands
---

# MEnvData-SWE-Trajectory: Agent Execution Trajectories for Software Engineering

<p align="center">
<a href="https://arxiv.org/abs/2601.22859"><img src="https://img.shields.io/badge/arXiv-2601.22859-b31b1b.svg"></a>
<a href="https://github.com/ernie-research/MEnvAgent"><img src="https://img.shields.io/badge/GitHub-MEnvAgent-blue"></a>
<a href="https://opensource.org/licenses/Apache-2.0"><img src="https://img.shields.io/badge/License-Apache_2.0-green.svg"></a>
</p>

<p align="center">
<a href="https://huggingface.co/datasets/ernie-research/MEnvBench"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20HF-MEnvBench-yellow"></a>
<a href="https://huggingface.co/datasets/ernie-research/MEnvData-SWE"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20HF-MEnvData--SWE-orange"></a>
</p>

## 📋 Dataset Description

**MEnvData-SWE-Trajectory** extends [MEnvData-SWE](https://huggingface.co/datasets/ernie-research/MEnvData-SWE) with **3,872 complete agent execution trajectories** across **3,005 task instances** from **942 repositories** in **10 programming languages**. These trajectories capture the full problem-solving process of an AI agent tackling real-world software engineering issues.

### Key Features

- 🤖 **3,872 Agent Trajectories**: Complete step-by-step execution records
- 🌐 **Multi-Language Coverage**: 10 programming languages (Python, Java, TypeScript, JavaScript, Rust, Go, C++, Ruby, PHP, C)
- 🎯 **State-of-the-Art Agent**: Powered by [OpenHands](https://github.com/All-Hands-AI/OpenHands) scaffolding
- 🧠 **Advanced LLM**: Generated using **Claude Sonnet 4.5**
### Trajectory Statistics

| Metric | Value |
|--------|-------|
| **Total Trajectories** | 3,872 |
| **Task Instances** | 3,005 |
| **Repositories** | 942 |
| **Languages** | 10 |
Framework** | OpenHands |
| **Expert Model** | Claude Sonnet 4.5 |

## 📋 Dataset Structure

Each trajectory is stored as a single JSON object with the following structure:

```json
{
"messages": [
{
"role": "system",
"content": "System prompt defining agent capabilities and tools..."
},
{
"role": "user",
"content": "GitHub issue description and task instructions..."
},
{
"role": "assistant",
"content": "Agent's reasoning and tool calls..."
},
...
]
}
```

## 📥 Loading the Dataset

```python
from datasets import load_dataset

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

# Access a single trajectory
trajectory = dataset['train'][0]
messages = trajectory['messages']

# Print conversation overview
for i, msg in enumerate(messages):
print(f"Message {i}: {msg['role']}")
print(f"Length: {len(msg['content'])} characters\n")
```

## 🔗 Related Datasets
- **[MEnvData-SWE](https://hugce.co/datasets/ernie-research/MEnvData-SWE)**: Base dataset with 3,005 executable Docker environments (without trajectories)

## 📖 Citation

If MEnvData-SWE-Trajectory helps your research, please cite:

```bibtex
@misc{guo2026menvagent,
title={MEnvAgent: Scalable Polyglot Environment Construction for Verifiable Software Engineering},
author={Chuanzhe Guo and 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](https://github.com/ernie-research/MEnvAgent)
- **Email**: [e-mail]