fine-tuning

Structured tool-calls from a 25M model

How small can a model be and still learn a structured capability?