Approaches to Lightweight Text-to-SQL Implementation Based on sLLM

This study proposes a lightweight Text-to-SQL implementation based on sLLM (smaller Large Language Model) to solve the high cost and security issues of existing LLM (Large Language Model) based Text-to-SQL models. To this end, we implemented a Text-to-SQL model using DAIL-SQL and Llama3-8B and evaluated its performance using Spider dataset. In this study, we secure the shortcomings of the existing Few-shot Learning method and propose improvement measures such as fine-tuning, knowledge distillation, and selection of similar queries through RAG to improve the performance of sLLM-based Text-to-SQL. By resolving the security vulnerabilities of existing LLM-based Text-to-SQL and presenting an efficient way to implement sLLM-based Text-to-SQL, we expect to expand the utilization of Text-to-SQL in various industries.