CSC-SQL-Merge-Qwen2.5-Coder-7B-Instruct Overview

This model, developed by Lei Sheng and Shuai-Shuai Xu, is a 7.6 billion parameter component of the innovative CSC-SQL framework, detailed in the paper "CSC-SQL: Corrective Self-Consistency in Text-to-SQL via Reinforcement Learning" (arXiv:2505.13271). It addresses limitations in traditional Text-to-SQL methods like Self-Consistency and Self-Correction by combining them into a unified approach.

Key Capabilities

• Enhanced SQL Generation: Integrates Self-Consistency and Self-Correction to improve the accuracy of SQL queries generated from natural language.
• Merge Revision Model: Specifically designed to select the two most frequent outputs from parallel sampling and feed them into a revision model for correction.
• Reinforcement Learning Fine-tuning: Utilizes the Group Relative Policy Optimization (GRPO) algorithm to fine-tune both the SQL generation and revision models, significantly boosting output quality.
• High Accuracy: Achieves 71.72% execution accuracy on the BIRD private test set with its 7B variant, demonstrating strong performance in complex Text-to-SQL benchmarks.

Good For

• Developers and researchers working on advanced Text-to-SQL applications.
• Scenarios requiring highly accurate and robust SQL query generation from natural language inputs.
• Integrating into systems that benefit from corrective self-consistency mechanisms for database interactions.