How we boosted SQL query accuracy by 33% with LLMs

A technical deep-dive into improving LLM-based SQL query generation using reinforcement learning techniques. The team developed GGPO (Guided Grammar Policy Optimization), combining GRPO/GSPO algorithms with grammar-guided decoding to fine-tune a Qwen3-0.6B model. Training on custom PostgreSQL datasets yielded a 33% relative improvement in execution accuracy on challenging queries from the BIRD benchmark, though overall performance remained similar to the base model. The approach addresses key limitations of supervised fine-tuning for reasoning tasks by directly optimizing for execution correctness rather than token likelihood.