Overview

The XGenerationLab/XiYanSQL-QwenCoder-32B-2504 is a 32 billion parameter model designed for advanced SQL generation. It leverages a combination of fine-tuning and GRPO (Gradient-based Policy Optimization) training without a thinking process, aiming for both efficiency and accuracy in converting natural language to SQL queries. The model supports multiple SQL dialects, including SQLite, PostgreSQL, and MySQL, and shows strong generalization capabilities across different dialects and out-of-domain datasets.

Key Capabilities

• Multi-Dialect SQL Generation: Excels at generating SQL for SQLite, PostgreSQL, and MySQL.
• Enhanced Performance: Achieves high accuracy on standard benchmarks like BIRD and Spider, and performs well on a real-world DW test set with complex queries.
• Improved Generalization: Demonstrates robust performance on out-of-domain datasets, reflecting its adaptability.
• Optimized Training: Utilizes a unique blend of fine-tuning and GRPO training for efficient and accurate SQL output.

Performance Highlights

The 32B model achieves notable results on various benchmarks:

• BIRD Dev@M-Schema: 67.14%
• Spider Test@M-Schema: 89.20%
• DW PostgreSQL@M-Schema: 53.52%
• DW MySQL@M-Schema: 57.74%

These metrics indicate strong performance compared to other models in its class, particularly on complex, real-world SQL scenarios.