AuToGen: Automated Tool Learning Data Generation with Domain-specific Structured Data
Daojian Zeng, Lin Zhou, Zhiheng Zhang, Lincheng Jiang · Data Intelligence · 2024
Large language models (LLMs) have demonstrated powerful decision-making and planning capabilities with external tools in solving real-world tasks. However, there are limitations in the data construction of tool learning such as high data structuring, symbolization, and privacy. This makes annotations costly and time-consuming, which is unsuitable for vertical deployment of tool-augmented model in real-world scenarios. Therefore, in this paper, we propose AuToGen, an automated tool learning data generation approach with domain-specific structured data. AuToGen leverages structured database table structures for keyword extraction. Then utilizes state-of-the-art LLMs for initial seed set generation and expands these sets for enriched tool-assisted model training data. Our experiment demonstrate that the high-quality of the data generated by AuToGen. Compared with the data generated by manually written seed sets, the model trained using AuToGen generated data has higher performance, proving that our method can efficiently assist in the deployment of real-world models.