SQLCoder

SQLCoder is a family of models from Defog.ai built for one job: converting natural language questions into SQL queries. The line runs from early 15B StarCoder and CodeLlama variants up to the current sqlcoder-7b-2, trained on 20,000+ human-curated questions across ten schemas that were kept out of the evaluation set. The flagship 7B-2 runs on a CodeLlama-7B base with a 16K context window and supports beam search for harder queries.

Defog's claim is that SQLCoder outperforms GPT-4 and GPT-4-turbo on their sql-eval framework of 200 questions on novel schemas, and it materially beats every other popular open-source model they tested. The weights carry a CC BY-SA 4.0 license with extra responsible-use restrictions, which means commercial use is allowed but fine-tuned derivatives must be shared under the same terms. It ships with a small pip package, an Apple Silicon build, and a demo that wires it straight to your database.

• Defog sql-eval (200 questions on novel schemas): claimed to outperform GPT-4 and GPT-4-turbo (provider-reported)
• Spider: ~83% execution accuracy claimed by third-party tables (unverified; not Defog's official metric)