The design of Text-to-SQL datasets for domestic databases
Li Guoshen1, Liu Yingjun2, Yu Lina 2, Ji Tao2, Zhang Hang1, Wu Jibing1
1. National Key Laboratory of Big Data and Decision; 2. National Key Laboratory of Intelligent Geospatial Information
Abstract： With the development of intelligent technology, the number and scale of databases have surged. Traditional data access technologies face problems such as long-time consumption and low efficiency when meeting the needs of massive data processing. Text-to-SQL technology has thus become an important bridge connecting user needs and database access. However, existing technologies are usually trained on open-source non-domestic datasets, and their application is plagued by issues like inconsistent database operation languages, lack of domain knowledge, and poor reliability. To address this, this paper, in line with the localization trend of software and hardware in the database field, designs a Text-to-SQL dataset for domestic databases, adopts a two-stage training technology for large language models based on synthetic data methods, proposes a Text-to-SQL method for domestic databases based on large language models, and fully verifies the effectiveness of the method through experiments.
Key words : fine-tuning of large language models; synthetic dataset; preference learning; domestic databas
