Defog

Defog is a technology company that has developed a cutting-edge large language model (LLM) called SQLCoder, designed to translate natural language questions directly into SQL queries. SQLCoder stands out for its remarkable performance, outperforming industry giants like GPT-3.5 turbo and text-davinci-003, a model over ten times its size, in natural language to SQL conversion tasks.

With 15 billion parameters, SQLCoder leverages the StarCoder model as its foundation and undergoes extensive fine-tuning specifically tailored for SQL generation. The model's training regimen involved a two-phased approach, initially focusing on "easy" and "medium" difficulty questions, and subsequently shifting to "hard" and "extra-hard" challenges, resulting in a significant seven percentage point improvement in overall performance.

SQLCoder offers an infinitely configurable front-end application that intelligently visualizes data, allowing users to delve deeper with follow-up questions. This extensible application can operate as a standalone solution or seamlessly integrate with popular business intelligence tools like Tableau and PowerBI. Furthermore, Defog prioritizes data privacy through its architecture, ensuring that user data remains secure and inaccessible. Its Hard Filters ensure appropriate row-level access restrictions for individual users, while an on-premises option provides ultimate control by hosting the model, training, and inference locally within the user's environment.