LLM-based SQL Generation with Reinforcement Learning
Keywords: Large Language Models, Generative AI, Reinforcement Learning, Text-to-SQL, SQL Query Generation, Resource Efficiency
TL;DR: Proposed SQL-RL-GEN and SQL-RL-GEN* models enhance text-to-SQL generation, improving accuracy by 2-7% while reducing resource usage with small LLM of 248M parameters only.
Abstract: The text-to-SQL problem remains a challenging task, even with the advancements of Large Language Models (LLMs). Current state-of-the-art models require extensive preprocessing steps and powerful LLMs to achieve accurate SQL query generation, which leads to significant resource utilization. We introduce two models deriving from one another SQL-RL-GEN and SQL-RL-GEN∗, that improve text-to-sql generation while minimizing the resources needed for training and maximizing flexibility. The SQL-RL-GEN generates a reward function to guide the agent’s training process, while SQL-RL-GEN∗ uses this reward function to tune a base LLM in solving the specified task. Our models achieve an accuracy improvement of 2-7% compared to state-of-the-art methods on a limited training dataset composed of only 1000 samples and with a small LLM of 248M parameters.
Submission Number: 10
Loading