Reproduce “ExeSQL: Self-Taught Text-to-SQL Models with Execution-Driven Bootstrapping for SQL Dialects” 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

Jipeng Zhang, Haolin Yang, Kehao Miao, Ruiyuan Zhang, Renjie Pi, Jiahui Gao, Xiaofang Zhou

ExeSQL introduces an execution-driven bootstrapping framework that automatically generates high-quality training data for multiple SQL dialects by leveraging large language models and real database execution feedback. This approach enables substantial performance gains, including a 15.2% improvement on PostgreSQL and 10.38% on MySQL over state-of-the-art models, while demonstrating strong generalization across various benchmarks.

Key insights

  • Integrating real database execution directly into the training loop grounds model learning in actual database behavior, providing specific error feedback for targeted query correction.
  • Automated, execution-verified data generation addresses the fundamental bottleneck of creating high-quality, dialect-specific training data efficiently and scalably.
  • A multi-stage training architecture, progressing from translation bootstrapping to iterative generation and preference optimization, systematically builds and refines multi-dialect SQL generation capabilities.

Summary by alphaXiv.

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