What is Agentar-Scale-SQL-Generation-32B?

This model, developed by AntGroup, is a 32-billion parameter Reasoning SQL Generator that serves as a core component of the state-of-the-art Agentar-Scale-SQL framework. It is fine-tuned from Omni-SQL-32B and further enhanced using an execution-grounded Reinforcement Learning (RL) framework called GRPO. This RL enhancement specifically targets improving its intrinsic reasoning capabilities for generating complex SQL queries.

Key Capabilities & Features

• Deep Reasoning: Engineered to perform deep, step-by-step reasoning to construct highly accurate SQL queries.
• RL-Enhanced: Utilizes GRPO for improved reasoning, making it robust for challenging Text-to-SQL tasks.
• Component of SOTA Framework: It is one of two main generators within the Agentar-Scale-SQL framework's "Diverse Synthesis" step, contributing to a robust pool of SQL candidates.

Performance & Use Cases

The Agentar-Scale-SQL framework, leveraging this generation model, achieved 81.67% execution accuracy on the BIRD benchmark, ranking first on the official leaderboard. This model is ideal for applications requiring precise and complex SQL generation from natural language questions, particularly in scenarios demanding high accuracy and deep reasoning over intricate database schemas. Its design focuses on generating valid and accurate SQL queries by thinking step-by-step, as demonstrated by its prompt template.