AgentInstruct

# AgentInstruct Dataset

🤖 [Models] • 💻 [Github Repo] • 📌 [Project Page] • 📃 [Paper]

**AgentInstruct** is a meticulously curated dataset featuring **1,866** high-quality interactions, designed to enhance AI agents across six diverse real-world tasks, leveraging innovative methods like **Task Derivation** and **Self-Instruct**.

- 🔍 **CoT** - Harness the power of [ReAct](https://react-lm.github.io/), offering detailed thought explanations for each action, ensuring an intricate understanding of the model's decision-making journey.
- 🌍 **Diversity** - Spanning 6 real-world scenarios, from Daily Household Routines to Database Operations, and their average turns range from 5 to 35.
- 🎯 **Precision** - Not all trajectories of GPT-4 are effective! Ours are rigorously filtered using strict rewards to ensure top-notch quality.
- ✅ **Assurance** - Rigorous checks to avoid data leakage, ensuring pristine dataset quality.

## Task Overview

| Task | # Filt. Traj. | Avg # Filt. Traj. Turns |
|---|---|---|
|ALFWorld|336|13.52|
|WebShop|351|3.68|
|Mind2Web|122|1.00|
|Knowledge Graph|324|6.04|
|Operating System|195|3.85|
|Database|538|2.06|
|**AgentInstruct**|1866|5.24|

AgentInstruct includes 1,866 trajectories from
6 agents tasks. "Traj." stands for interaction trajectory. "Filt. Traj."
stands for filtered trajectories.

## Models

**AgentLM** models are produced by mixed training on AgentInstruct dataset and ShareGPT dataset from Llama-2-chat models.

The models follow the conversation format of [Llama-2-chat](https://huggingface.co/blog/llama2#how-to-prompt-llama-2), with system prompt fixed as

```
You are a helpful, respectful and honest assistant.
```
7B, 13B, and 70B models are available on ModelScope model hub.

|Model|ModelScope Repo|
|---|---|
|AgentLM-7B| [ModelScope Repo](https://modelscope.cn/models/ZhipuAI/agentlm-7b) |
|AgentLM-13B| [ModelScope Repo](https://modelscope.cn/models/ZhipuAI/agentlm-13b) |
|AgentLM-70B| [ModelScope Repo](https://modelscope.cn/models/ZhipuAI/agentlm-70b) |

Check our [[Github Repo]](https://github.com/THUDM/AgentTuning) for details about **AgentTuning**.

## Citation

If you find our work useful, please consider citing AgentTuning:

```
@misc{zeng2023agenttuning,
title={AgentTuning: Enabling Generalized Agent Abilities for LLMs},
author={Aohan Zeng and Mingdao Liu and Rui Lu and Bowen Wang and Xiao Liu and Yuxiao Dong and Jie Tang},
year={2023},
eprint={2310.12823},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
```

# AgentInstruct 数据集

🤖 [模型] • 💻 [GitHub仓库] • 📌 [项目主页] • 📃 [论文]

**AgentInstruct** 是一套经过精心整理的数据集，包含**1866**条高质量交互样本，旨在通过任务推导（Task Derivation）与自指令（Self-Instruct）等创新方法，提升AI智能体（AI Agent）在6种多样化真实场景任务中的表现。

- 🔍 **思维链（CoT）**：采用[ReAct](https://react-lm.github.io/)框架，为每一步操作提供详细的思维解释，助力深入理解模型的决策过程。
- 🌍 **多样性**：覆盖6类真实应用场景，涵盖日常家庭事务至数据库操作等领域，单条轨迹的平均交互轮次介于5至35之间。
- 🎯 **精准性**：并非所有GPT-4生成的交互轨迹都具备有效性！本数据集通过严格的奖励机制进行多轮筛选，以确保数据质量达到顶尖水准。
- ✅ **可靠性**：经过严格的合规检查以规避数据泄露问题，保障数据集的纯净性与高质量。

## 任务概览

| 任务名称 | 筛选后交互轨迹数 | 单条轨迹平均交互轮次 |
|---|---|---|
| ALFWorld | 336 | 13.52 |
| WebShop | 351 | 3.68 |
| Mind2Web | 122 | 1.00 |
| 知识图谱（Knowledge Graph） | 324 | 6.04 |
| 操作系统（Operating System） | 195 | 3.85 |
| 数据库（Database） | 538 | 2.06 |
| **AgentInstruct** | 1866 | 5.24 |

AgentInstruct 共包含来自6个智能体任务的1866条交互轨迹。其中“Traj.”为交互轨迹（interaction trajectory）的缩写，“Filt. Traj.”为筛选后交互轨迹的缩写。

## 模型

**AgentLM** 模型是以Llama-2-chat模型为基础，在AgentInstruct数据集与ShareGPT数据集上进行混合训练得到的。

该模型遵循[Llama-2-chat](https://huggingface.co/blog/llama2#how-to-prompt-llama-2)的对话格式，系统提示词固定为：

You are a helpful, respectful and honest assistant.

7B、13B与70B参数版本的模型已在ModelScope模型 Hub 上线。

| 模型名称 | ModelScope 仓库地址 |
|---|---|
| AgentLM-7B | [ModelScope 仓库](https://modelscope.cn/models/ZhipuAI/agentlm-7b) |
| AgentLM-13B | [ModelScope 仓库](https://modelscope.cn/models/ZhipuAI/agentlm-13b) |
| AgentLM-70B | [ModelScope 仓库](https://modelscope.cn/models/ZhipuAI/agentlm-70b) |

如需了解**AgentTuning**的详细信息，请访问我们的[GitHub仓库](https://github.com/THUDM/AgentTuning)。

## 引用

若您的研究工作用到了本数据集，请引用AgentTuning：

@misc{zeng2023agenttuning,
title={AgentTuning: Enabling Generalized Agent Abilities for LLMs},
author={Aohan Zeng and Mingdao Liu and Rui Lu and Bowen Wang and Xiao Liu and Yuxiao Dong and Jie Tang},
year={2023},
eprint={2310.12823},
archivePrefix={arXiv},
primaryClass={cs.CL}
}