SALT-NLP/cogym-real-trajectories

---
license: cc-by-sa-4.0
---

# Collaborative Gym Real Trajectories Dataset

This dataset includes 228 human-agent collaboration trajectories collected from the Stanford [Collaborative Gym platform](https://github.com/SALT-NLP/collaborative-gym) before 2025/04/06. These trajectories are collected from real human users and we have deanonymized the user information in the data.

## Dataset Structure

Each trajectory is represented as a JSON file with the following fields:

- **modelName**: The model that powers the collaborative agent (e.g., "gpt-4o")
- **task**: The collaboration task, one of:
- `travel_planning` - Planning trips and itineraries
- `related_work` - Writing related work sections for academic papers
- `tabular_analysis` - Analyzing tabular data
- **query**: The initial user query that starts the collaboration
- **outcomeRating**: User rating of the final artifact (i.e., travel planning, related work section, analysis result) on a 1-5 Likert scale, if provided
- **agentRating**: User rating of the overall satisfaction on a 1-5 Likert scale, if provided
- **communicationRating**: User rating of the communication capability of the agent during the collaboration process on a 1-5 Likert scale, if provided
- **agentFeedback**: Natural language feedback from the user, if provided
- **event_log**: The complete human-agent collaboration trajectory, including all messages and actions

## Citation

If you use this dataset in your research, please cite the Collaborative Gym platform:

```bibtex
@article{shao2024collaborative,
title={Collaborative gym: A framework for enabling and evaluating human-agent collaboration},
author={Shao, Yijia and Samuel, Vinay and Jiang, Yucheng and Yang, John and Yang, Diyi},
journal={arXiv preprint arXiv:2412.15701},
year={2024}
}
```

## License

This dataset is released under the CC-BY-SA-4.0 license.

---
许可证：CC-BY-SA-4.0
---

# 协作Gym真实轨迹数据集

本数据集包含截至2025年4月6日从斯坦福协作Gym（Collaborative Gym）平台（https://github.com/SALT-NLP/collaborative-gym）采集的228条人机协作轨迹。所有轨迹均来源于真实人类用户，且我们已完成数据中用户信息的去匿名化处理。

## 数据集结构

每条轨迹以JSON文件形式存储，包含以下字段：

- **modelName**：驱动协作智能体的模型（例如"gpt-4o"）
- **task**：协作任务，可选取值如下：
- `travel_planning` - 旅行规划与行程制定
- `related_work` - 学术论文相关工作章节撰写
- `tabular_analysis` - 表格数据分析
- **query**：启动协作流程的初始用户查询
- **outcomeRating**：用户对最终产出物（即旅行规划方案、学术相关工作章节、数据分析结果）的1-5李克特量表评分，若用户提供则包含该字段
- **agentRating**：用户对整体协作满意度的1-5李克特量表评分，若用户提供则包含该字段
- **communicationRating**：用户对协作过程中智能体沟通能力的1-5李克特量表评分，若用户提供则包含该字段
- **agentFeedback**：用户提供的自然语言反馈内容，若用户提供则包含该字段
- **event_log**：完整的人机协作交互轨迹，涵盖所有交互消息与操作行为

## 引用说明

若您在研究工作中使用本数据集，请引用协作Gym平台的相关学术成果：

bibtex
@article{shao2024collaborative,
title={Collaborative gym: A framework for enabling and evaluating human-agent collaboration},
author={Shao, Yijia and Samuel, Vinay and Jiang, Yucheng and Yang, John and Yang, Diyi},
journal={arXiv preprint arXiv:2412.15701},
year={2024}
}

## 许可证

本数据集采用CC-BY-SA-4.0许可证发布。