**Agent Instruct is a lightweight instruction fine-tuning dataset containing high-quality interaction trajectories.**AgentInstruct is a selected agent dataset containing 1,866 high-quality interaction trajectories and 6 diverse real-world tasks. It is used to enhance the agent capabilities of language models and has the following features:

• Thinking chain: Use the ReAct prompt strategy to provide a detailed thinking chain for each step, and deeply understand the model decision-making process.
• Diversity: Covers 6 real-world scenarios, ranging from daily household chores to operating databases, with an average round number ranging from 5 to 35.
• Accuracy: GPT-4 cannot completely perform intelligent tasks, and uses a trajectory reward mechanism to strictly screen the data to ensure the quality of each piece of data.
• Generalizability: Strict inspection to avoid data leakage and ensure the generalizability of data.

Citation

@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}
}