Gretel, the leader in synthetic data, today released the world’s largest open source Text-to-SQL dataset to unlock new possibilities for AI in the enterprise. Available on Hugging Face and released under the Apache 2.0 license, Gretel’s dataset consists of over 100,000 high-quality synthetic Text-to-SQL samples with SQL metadata and spans 100 verticals. With access to Gretel’s open-source, high-quality synthetic dataset, developers can train AI models that empower business users to extract value from critical enterprise data sources, expediting AI initiatives across the enterprise.
Access to quality training data is one of the biggest obstacles to building with generative AI. Everything Gretel does is designed to address this issue head-on, and contributing to the open-source community is no exception. By providing developers with high-quality, synthetic Text-to-SQL data, we're enabling them to create AI models that can understand natural language queries and generate SQL queries. This empowers users across the organization to easily access and derive insights from complex databases, data warehouses, and data lakes, without needing to learn SQL or rely on technical teams. We’re excited for developers to take our dataset for a spin, and build upon it.”
Alex Watson, Co-Founder & Chief Product Officer at Gretel
Growing Demand for AI Training Data
The largest AI companies in the world are struggling with access to high-quality training data. And in the enterprise, Text-to-SQL data — data that’s essential for building natural language interfaces to critical data sources — is in particularly high demand. Nearly every enterprise has invaluable insights buried in data tables or data views that are only accessible to developers skilled in Structured Query Language (SQL) — the standard language for interacting with databases, data warehouses and data lakes. AI models trained on Text-to-SQL data allow business users to derive value from these datasets on demand.
Most text-to-SQL datasets today are manually curated and annotated, limiting their size, applicability and utility. This process is expensive, labor intensive, and cumbersome. For instance, the Spider text-to-SQL dataset, consisting of 7k samples, was annotated by 11 college students at Yale, and took a total of 1,000 hours to complete — an incredible amount of effort for a relatively small dataset in the context of large language models.
Furthermore, the vast majority of existing Text-to-SQL datasets lack a natural language explanation of what their SQL code does. Gretel’s dataset includes an explanation field providing a plain-english description of the SQL code, which helps end users quickly understand the output and realize its value.
Filling a Gap in the Open Source Community
To date, the open source community has offered little reprieve. The Spider dataset, for instance, is available under a commercially permissive creative commons license (CC-BY-SY-4.0), but it’s a copyleft license, meaning a derivative work must be licensed under the same or a compatible license. This differs significantly from MIT or Apache licenses, which allow derivative works to be released under different license terms without attribution or sharealike terms.
With Gretel’s Text-to-SQL dataset released under the Apache 2.0 license, AI developers can build conversational applications that open up a range of new opportunities for businesses across industries. For instance:
- Finance: Analysts and managers can ask questions about the company's financial performance and get instant answers sourced from their databases. Example query: "What was the total revenue generated from credit card transactions in the last quarter, broken down by product category?”
- Health: Providers can streamline the process of querying and analyzing clinical trial data from multiple (2-10k) experiments. Example query: "Find the average reduction in blood pressure for patients aged 45-60 who received the new drug, compared to the placebo group, over the last 6 months of the trial."
- Government: Leaders can provide citizens with an easy way to search for and access public records databases that may include licenses, property ownership, permits, etc. Example query: "Find the top 10 counties with the highest population growth rate between 2010 and 2020."
Elevating Data Quality in the Age of AI
Gretel’s Text-to-SQL dataset was generated by Gretel Navigator, a compound AI system that integrates agent-based execution, multiple proprietary models, including a custom tabular Large Language Model (TabLLM), and privacy-enhancing technologies to generate high quality synthetic data for enterprise AI teams.
The quality of Gretel’s dataset was assessed and compared to other Text-to-SQL datasets using an independent service and the LLM-as-a-judge technique. A full synopsis of Gretel’s holistic quality evaluation is available on the blog. A deeper dive into the data, including data previews, examples, and statistics, can be found on Hugging Face.