Preprint reports text-to-SQL results using fewer training examples

Researchers report that a Qwen3-8B model trained with selected examples can outperform text-to-SQL methods trained on 20 times more data. Their LIMIT framework chose 796 examples for the BIRD benchmark and 863 for Spider, reporting execution scores of 69.1% and 88.9%, respectively. The preprint describes filtering examples by difficulty, generating and checking reasoning steps, scoring example quality and optimizing coverage of database tables. Its results suggest a route to reducing the data needed to teach models to turn questions into database queries, though the reported performance comes from benchmark tests rather than production use.