Reproduce “LitE-SQL: A Lightweight and Efficient Text-to-SQL Framework with Vector-based Schema Linking and Execution-Guided Self-Correction” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shengmin Piao, Jieun Lee, Sanghyun Park

Researchers from Yonsei University developed LitE-SQL, a lightweight and efficient Text-to-SQL framework that integrates vector-based schema linking and execution-guided self-correction to generate SQL queries. The framework achieved execution accuracies of 72.10% on BIRD and 88.45% on Spider 1.0, outperforming several larger proprietary and fine-tuned models while utilizing significantly fewer parameters and demonstrating reduced inference latency.

Key insights

  • The HN-SupCon loss effectively improves schema retrieval by enabling the embedding model to distinguish between semantically similar yet functionally irrelevant schema elements, crucial for focusing the generator.
  • A two-stage fine-tuning approach, combining SFT with DPO-based RFT that leverages execution feedback, significantly enhances the robustness of lightweight models in correcting both semantic and syntactic errors without expensive multi-candidate sampling.
  • High-quality Text-to-SQL performance can be achieved with lightweight, fine-tuned open-source models (e.g., 7B parameters), offering a parameter-efficient and privacy-preserving alternative to larger proprietary LLMs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.