A FIELD GUIDE
Raising your own small language model, from first egg to first flight.
GET THE BOOK ON GUMROADPay what you want, starting at $1. Ten chapters. Yours forever.
THE PITCH
Everyone is renting dragons.
This book is about raising your own.
The biggest models on Earth live behind someone else's API. You pay by the token, you follow their rules, and the dragon never learns your name. For a great many jobs, that is power you do not need, rented at a price you do not control.
A small model, raised on your own data and aimed at one job, is a different animal entirely. It runs on hardware you own. It answers in milliseconds. It never phones home.
This is not a textbook and it is not theory. It is the field guide Mark A Evans wrote after actually doing it, twice: Cast, a 3.3 million parameter model that turns plain English into database queries, and Imagine, a 1.6 billion parameter model built with a full fine tune, which scored a perfect 40 out of 40 on held out tests and shipped to production.
Real story. Real numbers. Every failure left in. No PhD required.
PROOF, NOT PROMISES
The book follows two real models from egg to production. Everything in it was learned by training these two.
The first dragon. Small, fast, precise.
A tiny model trained from scratch to translate natural language into database queries. Proof that a model does not need billions of parameters to do one job brilliantly.
The second dragon. The one that shipped.
A specialized machine learning model built with a full fine tune on everything Cast taught. It passed every held out test it was given, 40 out of 40, and went on to run in production, answering real questions against real data.
If a model with 3.3 million parameters can do this, imagine what yours can do.
INSIDE THE BOOK
Each chapter pairs the story of what actually happened with the practical craft of doing it yourself.
The case for raising your own model instead of renting a giant one, and the jobs small dragons win.
Picking the right size, shape, and temperament of model for the one job you need done.
What a young model eats, how to gather and clean its food, and why data decides everything.
The training setup: hardware, code, and the environment your dragon grows up in.
The training run itself. Loss curves, dead runs, 2 AM decisions, and knowing when to let it cook.
Held out tests, honest scoring, and the moment a model proves it can fly on its own.
Teaching a model to refuse what it does not know instead of making something up.
Giving your dragon a home in production: serving it, watching it, and keeping it healthy.
Publishing your model to the world, weights and all, and standing behind your work.
Where a dragon keeper goes from here: bigger jobs, stranger models, and the road ahead.
BEGIN
Ten chapters. Two real dragons. Every number on the table. Name your price and start raising your own.
GET THE BOOK ON GUMROADPay what you want, starting at $1