Reproduce “FINER-SQL: Boosting Small Language Models for Text-to-SQL” on OpenResearch

About this paper

Thanh Dat Hoang, Thanh Trung Huynh, Matthias Weidlich, Thanh Tam Nguyen, Tong Chen, Hongzhi Yin, Quoc Viet Hung Nguyen

The FINER-SQL framework significantly boosts Small Language Models for Text-to-SQL by employing a two-stage reinforcement learning approach with fine-grained execution feedback. This enables SLMs to achieve execution accuracy on par with much larger language models while being considerably more efficient and privacy-preserving.

Key insights

• Dense, fine-grained reward signals, including semantic alignment of reasoning traces and partial structural correctness of SQL, are crucial for effectively training SLMs on complex generative tasks like Text-to-SQL.
• Cross-database memory-based reasoning regularization stabilizes SLM policy and reduces reasoning variance, contributing to improved consensus during inference majority voting.
• Decomposing SQL into atomic operations provides an effective and continuous learning signal for structural correctness, guiding models even when full query execution fails.

Summary by alphaXiv.