XiYan-SQL is a framework designed to improve the accuracy and robustness of converting natural language queries into SQL statements. It targets researchers and developers working on text-to-SQL tasks, offering state-of-the-art performance through an ensemble of specialized models and advanced schema representation.

How It Works

The framework employs a multi-generator ensemble strategy, combining multiple SQL generation models to produce a diverse set of candidate queries. It utilizes "M-Schema," a semi-structured schema representation method, to enhance the model's understanding of database structures. XiYan-SQL integrates both in-context learning (ICL) with example selection based on named entity recognition and supervised fine-tuning strategies to generate high-quality, diverse SQL candidates. A refiner module corrects syntactical and logical errors, and a dedicated selection model is fine-tuned to identify the best candidate query.

Quick Start & Requirements

• Models are available on HuggingFace and ModelScope.
• The project supports local deployment via XiYan-MCP-server for high-security data access.
• Specific hardware requirements are not detailed, but model sizes range from 3B to 32B parameters, suggesting significant computational resources may be needed for larger models.
• Links: 🤗 XiYan GBI, 💻 M-Schema, 📖 Arxiv, PapersWithCode.

Highlighted Details

• Achieved SOTA performance on the Bird leaderboard with an EX score of 75.63% and R-VES of 71.41%.
• XiYanSQL-QwenCoder-32B model achieved SOTA on the Bird test set with an EX score of 69.03% as a single fine-tuned model.
• Framework includes components for database description generation and a DateResolver model for enhanced date understanding, particularly for Chinese queries.
• Offers multiple model sizes (3B, 7B, 14B, 32B) to cater to different developer needs.

Maintenance & Community

• The project is actively developed, with frequent updates and releases noted in the README (e.g., new model versions, local server support).
• A DingTalk group is available for community interaction: 94725009401.
• The project welcomes contributions and feedback.

Licensing & Compatibility

• The README does not explicitly state a license for the code or models. However, the project is open-sourced, and models are available on HuggingFace and ModelScope, suggesting permissive usage, but commercial use should be verified.

Limitations & Caveats

• Some components, like the ensemble selection model and MoMQ (multi-dialect Text-to-SQL MoE model), are marked as "to release soon."
• The DateResolver model is noted as being "major for Chinese," implying potential limitations for other languages.
• While SOTA claims are made, specific benchmarks and comparisons are primarily against leaderboard scores rather than direct code comparisons.