Think2SQL-4B: Text-to-SQL Reasoning Model

Think2SQL-4B is a 4 billion parameter model, fine-tuned from Qwen/Qwen3-4B, engineered for advanced Text-to-SQL reasoning. It is based on the Think2SQL framework, which systematically injects reasoning capabilities into Text-to-SQL tasks using Reinforcement Learning with Verifiable Rewards (RLVR). The model was trained on the BIRD dataset with thinking disabled, utilizing the TRL library.

Key Capabilities

• Text-to-SQL Generation: Generates accurate SQL queries from natural language questions, provided evidence, and database schemas.
• Reasoning Integration: Incorporates a structured reasoning process, allowing it to first think through the problem before providing the SQL answer.
• Optimized for Performance: Achieves reasoning capabilities competitive with state-of-the-art models while operating under computational constraints.
• Execution-Guided Rewards: Utilizes novel execution-guided dense reward functions, outperforming simpler binary signals in training.

Good For

• Data Science Experts: Ideal for data scientists and developers needing to convert complex natural language queries into precise SQL.
• Database Interaction: Automating SQL query generation for various database engines, demonstrated with SQLite in examples.
• Research in Text-to-SQL: Provides a comprehensive analysis for optimizing Text-to-SQL reasoning, including insights into reward density, advantage scaling, and model capacity.