Tune Llama 3 for text-to-SQL and improve accuracy from 30% to 95%

This repo and notebook meta-lamini.ipynb demonstrate how to tune Llama 3 to generate valid SQL queries and improve accuracy from 30% to 95%.

In this notebook we'll be using Lamini, and more specifically, Lamini Memory Tuning.

Lamini is an integrated platform for LLM inference and tuning for the enterprise. Lamini Memory Tuning is a new tool you can use to embed facts into LLMs that improves factual accuracy and reduces hallucinations. Inspired by information retrieval, this method has set a new standard of accuracy for LLMs with less developer effort.

Learn more about Lamini Memory Tuning: https://www.lamini.ai/blog/lamini-memory-tuning

Please head over to https://app.lamini.ai/account to get your free api key.

You can authenticate by writing the following to a file ~/.lamini/configure.yaml

production:
    key: <YOUR-LAMINI-API-KEY>

This tuning tutorial uses the nba_roster sqlite database to tune a Llama 3 model.

Additional resources

▫️ Fortune 500 case study: http://www.lamini.ai/blog/llm-text-to-sql
▫️ Technical paper: https://github.com/lamini-ai/Lamini-Memory-Tuning/blob/main/research-paper.pdf
▫️ Model weights: https://huggingface.co/engineering-lamini/lamini-1-random