FALCON: A Model and Dataset Agnostic Framework for Synthetic Data Generation for NL2SQL Pairs
FALCON is built to generate harder, ambiguity-aware NL2SQL training pairs locally, using compact open models.
The paper says current synthetic NL-to-SQL generators often produce simplified queries that do not reflect real database access. FALCON uses SQL reserved-word seeding, persona-based prompting, and alignment-based filtering to keep complex but valid examples. Human evaluation found consistent quality across model sizes, and training gains were strongest as query complexity increased. Mixed training with some benchmark data helped recover simpler-query performance while keeping the complex-query advantage. ArXiv · AI/CL/LG's note
The paper says current synthetic NL-to-SQL generators often produce simplified queries that do not reflect real database access. FALCON uses SQL reserved-word seeding, persona-based prompting, and alignment-based filtering to keep complex but valid examples. Human evaluation found consistent quality across model sizes, and training gains were strongest as query complexity increased. Mixed training with some benchmark data helped recover simpler-query performance while keeping the complex-query advantage. ArXiv · AI/CL/LG's note
score 4