This project provides DuckDB-NSQL, a foundational model (FM) specifically designed for generating SQL queries for local DuckDB analytics. It targets data analysts and engineers who need to interact with DuckDB databases using natural language, offering a streamlined way to query data without extensive SQL knowledge.

How It Works

DuckDB-NSQL is an autoregressive language model trained on a dataset of synthetically generated DuckDB SQL queries and transpiled text-to-SQL pairs. It leverages the NSQL family of models, optimized for SQL generation. The model can be run locally using llama.cpp for efficient inference, allowing direct integration with DuckDB connections for natural language querying.

Quick Start & Requirements

• Install: pip install -r requirements.txt
• Prerequisites: Python, llama.cpp (for hosting the model), DuckDB. Model weights available on HuggingFace (e.g., motherduckdb/DuckDB-NSQL-7B-v0.1-GGUF).
• Usage: Examples provided in the examples/ folder demonstrate connecting to DuckDB and querying data.
• Links: HuggingFace, Examples

Highlighted Details

• 7B parameter model available in multiple formats (e.g., GGUF).
• Trained on 200k synthetic DuckDB SQL queries and NSText2SQL data transpiled to DuckDB SQL.
• Supports querying local DuckDB databases via natural language.
• Includes evaluation benchmarks and scripts in the eval/ folder.

Maintenance & Community

• Developed by NumbersStationAI.
• Model weights hosted by motherduckdb.

Licensing & Compatibility

• License details are not explicitly stated in the README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The README does not specify the exact license, which may impact commercial adoption. The model is presented as a foundational model for DuckDB SQL analytics, implying potential limitations in handling highly complex or niche SQL constructs not covered in its training data.