XGenerationLab/XiYanSQL-QwenCoder-14B-2504
The XiYanSQL-QwenCoder-14B-2504 model by XGenerationLab is a 14 billion parameter SQL generation model built on the QwenCoder architecture, featuring a 32768-token context length. It incorporates fine-tuning and GRPO training strategies to achieve high efficiency and accuracy in SQL generation. This model excels at generating SQL across multiple dialects, including SQLite, PostgreSQL, and MySQL, and demonstrates strong generalization capabilities on out-of-domain datasets.
Loading preview...
XiYanSQL-QwenCoder-14B-2504 Overview
Developed by XGenerationLab, XiYanSQL-QwenCoder-14B-2504 is a 14 billion parameter SQL generation model that leverages a combination of fine-tuning and GRPO (Gradient-based Reinforcement Learning with Policy Optimization) training. This approach aims to deliver both efficiency and accuracy in generating SQL queries without requiring a separate thinking process.
Key Capabilities
- Multi-Dialect SQL Generation: Supports mainstream SQL dialects including SQLite, PostgreSQL, and MySQL.
- Enhanced Performance: Demonstrates improved accuracy and generalization compared to previous versions, particularly on out-of-domain datasets.
- Real-World Benchmarking: Evaluated against a proprietary DW test set comprising thousands of complex queries from real PostgreSQL and MySQL scenarios, in addition to standard benchmarks like BIRD and Spider.
- Schema Flexibility: Optimized for M-Schema format but also compatible with DDL, offering flexibility in schema input.
Performance Highlights
The 14B model shows competitive performance across various benchmarks:
- BIRD Dev@M-Schema: Achieves 65.32%
- Spider Test@M-Schema: Achieves 86.82%
- DW PostgreSQL@M-Schema: Achieves 40.52%
- DW MySQL@M-Schema: Achieves 44.60%
These results indicate its robust capabilities in text-to-SQL tasks, especially for complex, real-world database interactions.
When to Use This Model
This model is ideal for developers and organizations requiring accurate and efficient SQL generation across diverse database environments. Its strong performance on multi-dialect and out-of-domain datasets makes it suitable for applications that need to interact with various SQL databases reliably. The model's optimization for M-Schema also provides a structured approach for schema definition, potentially simplifying integration.