SynSQL-2.5M - The First Million-Scale Cross-Domain Text-to-SQL Dataset

We introduce the first million-scale text-to-SQL dataset, SynSQL-2.5M, containing over 2.5 million diverse and high-quality data samples, spanning more than 16,000 databases from various domains.

Building on SynSQL-2.5M, we introduce OmniSQL, a family of powerful text-to-SQL models available in three sizes: 7B, 14B, and 32B. During the fine-tuning process, we also integrate training sets from Spider and BIRD, which provide high-quality, human-labeled data.

Paper Link: https://arxiv.org/abs/2503.02240

GitHub Link: https://github.com/RUCKBReasoning/OmniSQL

Downloads

Statistics about SynSQL-2.5M

SynSQL-2.5M is a high-quality synthetic text-to-SQL dataset, generated entirely using open-source LLMs and released under Apache 2.0.

The dataset includes:

• 2,544,390 diverse and complex text-to-SQL samples, each consisting of a <database, question, SQL query, chain-of-thought solution> quad.
• Coverage of 16,583 synthetic databases from realistic scenarios.
• A wide range of SQL complexity levels: simple, moderate, complex, highly complex, from single-table queries to advanced multi-table joins, functions, and common table expressions.
• A variety of linguistic styles in natural language questions: formal, colloquial, imperative, interrogative, descriptive, concise, vague, metaphorical, and conversational.
• Chain-of-thought (CoT) solutions provided for all data samples.

For more statistics and quality evaluations, refer to our paper. As of March 2025, SynSQL-2.5M is the largest and most diverse synthetic text-to-SQL dataset to date. It represents a significant milestone in the text-to-SQL community. We encourage researchers, practitioners, and data enthusiasts to explore and build models using this dataset. If you find it useful, please consider giving us a star or citing our work. Your feedback is our greatest motivation to continue advancing.

Limitations

SynSQL-2.5M is an English dataset focused on the SQLite database engine, so its performance in multi-language and multi-SQL dialect scenarios may be limited. However, you can synthesize new data samples using our proposed framework to suit your scenarios. After synthesizing a new dataset, you can use OmniSQL for further fine-tuning, as it is a strong starting point for text-to-SQL capabilities.

Contact

If you have any questions, we encourage you to either create Github issues or get in touch with Haoyang Li at lihaoyang.cs@ruc.edu.cn.