Max Caldwell

Fine-tuning a small model to drive Nushell

LFM3 1.7B goes from 8% to 92.6% correct Nushell tool calls with a LoRA and data generated from scratch.

The model in 3D over one request, base model and then with its LoRA

script<T>

Scriptt is prompting with a model for types to catch agent errors early.