cycloneboy/CscSQL-Grpo-Qwen2.5-Coder-3B-Instruct

TEXT GENERATIONConcurrency Cost:1Model Size:3.1BQuant:BF16Ctx Length:32kPublished:May 21, 2025License:cc-by-nc-4.0Architecture:Transformer Open Weights Cold

The cycloneboy/CscSQL-Grpo-Qwen2.5-Coder-3B-Instruct is a 3.1 billion parameter instruction-tuned causal language model, fine-tuned on the Qwen2.5-Coder architecture by CycloneBoy. This model specializes in Text-to-SQL generation, leveraging the CSC-SQL method which integrates Self-Consistency and Self-Correction with Group Relative Policy Optimization (GRPO) for enhanced accuracy. It is specifically designed to translate natural language questions into SQL queries, achieving high execution accuracy on complex database tasks.

Loading preview...

Overview

This model, cycloneboy/CscSQL-Grpo-Qwen2.5-Coder-3B-Instruct, is a 3.1 billion parameter instruction-tuned variant of the Qwen2.5-Coder architecture, developed by CycloneBoy. It is specifically designed for advanced Text-to-SQL generation, aiming to improve the accuracy of converting natural language questions into SQL queries. The model incorporates a novel approach called CSC-SQL (Corrective Self-Consistency in Text-to-SQL via Reinforcement Learning).

Key Capabilities

  • Enhanced SQL Generation: Utilizes the CSC-SQL method, which combines Self-Consistency and Self-Correction techniques to produce more accurate SQL queries.
  • Reinforcement Learning Fine-tuning: Employs the Group Relative Policy Optimization (GRPO) algorithm to fine-tune both the SQL generation and revision models, significantly boosting output quality.
  • Robust Error Correction: CSC-SQL addresses limitations of traditional methods by selecting the two most frequent outputs from parallel sampling and feeding them into a merge revision model for correction, handling both syntactic and semantic issues.
  • Performance: While the 3B model's specific BIRD test set execution accuracy isn't detailed, the 7B model achieved 71.72% and the 32B model achieved 73.67% on the BIRD private test set, indicating the effectiveness of the CSC-SQL approach.

Good For

  • Developers and researchers working on Text-to-SQL applications requiring high accuracy.
  • Integrating natural language interfaces with relational databases.
  • Exploring advanced techniques in LLM-based SQL generation and error correction.

Finetunes
1 models