# DTS-SQL: Decomposed Text-to-SQL with Small Large Language Models

**Authors:** Mohammadreza Pourreza, Davood Rafiei  
**arXiv:** [2402.01117](https://arxiv.org/abs/2402.01117)  
**Date:** 2024-02-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.01117v1/

## Key points

- Performance discrepancies between small open-source models and large proprietary models in the text-to-SQL task
- Reliance on proprietary Large Language Models (LLMs) raises data privacy concerns
- Proposal of a novel two-stage fine-tuning approach to bridge the performance gap and reduce dependence on LLMs
- Approach enhances execution accuracy by 3 to 7 percent, aligning open-source model performance with proprietary counterparts
- Emphasis on ethical considerations throughout the research process, including minimizing biases and discriminatory elements
- Future investigations could focus on refining different stages within the proposed approach and exploring advanced techniques for schema-linking tasks, such as leveraging larger language models like GPT-4.

## Summary

In this study, we address the issue of performance discrepancies between small open-source models and large proprietary models in the text-to-SQL task. This is primarily driven by the reliance on proprietary Large Language Models (LLMs), which raises concerns about data privacy. To bridge this performance gap and reduce dependence on LLMs, we propose a novel two-stage fine-tuning approach that breaks down the task into simpler components. Through extensive evaluation on diverse datasets and small LLMs, we demonstrate that our approach enhances execution accuracy by 3 to 7 percent, effectively aligning the performance of open-source models with their proprietary counterparts. Furthermore, our research highlights the importance of ethical considerations throughout every phase of our work. We have taken proactive measures to minimize biases and discriminatory elements in our research design, data selection, and result interpretation. By adhering to ethical principles and guidelines, we aim to contribute responsibly to knowledge advancement in our field. Moving forward, future investigations could focus on refining the performance of different stages within our proposed approach and exploring advanced techniques for schema-linking tasks. This includes investigating retrieval methods or leveraging larger language models like GPT-4 for schema-linking tasks to identify optimal methodologies for enhancing overall system performance. Through transparent reporting and dedication to fairness in research practices, we underscore our commitment to conducting research with integrity and respect for ethical standards.