A free book from the BC + AI community
Training
your own
models.
On one 24 GB GPU.
Get your hands into how a model learns. Start with three parameters, train a tiny classifier, and work your way toward language, images, speech, and sound.
- 40
- chapters
- 57
- companion scripts
- 243
- glossary terms
- 247
- PDF pages, covers included
Begin at the workbench
Make a small thing.
Know why it works.
You can run the first project with Python, before you need a GPU or a machine-learning framework. Train it. Save it. Give it an input it has never seen.
The later projects introduce neural networks and more demanding hardware. Along the way, you learn to choose data, inspect mistakes, and decide what would count as a useful result.
You don’t need a background in neural architecture. Be ready to edit a file, run a command, and look carefully at the output.
Start with the three-parameter classifierInside the book
What do you want to make?
These are starting points, not separate courses. New to training? Begin with your first model and build up from there.
Start with something small enough to understand.
The first classifier uses three parameters and plain Python. Follow an example from training to a saved model and a fresh prediction before adding a neural-network library.
- Chapter 01Train your first modelA small classifier you can inspect from end to end.
- Chapter 03Understand what changes when a model learnsParameters, tensors, loss, and gradients.
- Chapter 05Build the dataset for the jobChoose examples that teach the behavior you need.
- Chapter 06Check whether it improvedEvaluate on examples the model has not learned from.
Try it in your browser
Pull on an idea.
See what moves.
The web edition includes small interactive explanations. No installation needed. These illustrate the concepts; they don’t run GPU training.
Watch a parameter learn.
Change the learning rate and see what happens to the loss.
Try gradient stepsWhere does the memory go?
Account for weights, gradients, and optimizer state. A planning calculation with explicit assumptions.
Explore the memory budgetWhat can this token see?
Step through a causal attention mask and see which positions are available.
Explore causal attentionRead the evidence with the recipe
What fits?
What actually works?
A model’s weights are only part of its memory bill. Training also needs space for gradients, optimizer state, and the work done between layers.
This edition records CPU and dependency-free checks, including the first classifier. No GPU training or 24 GB peak-memory measurements were performed for the edition. The GPU recipes and memory budgets are starting points to test on your own hardware.
Read the verification recordKeep it on your workbench
Read. Run. Come back.
The complete book, editable manuscript, and code archive. Use the companion files to run the examples.
02 / Find the missing wordThe linked glossaryLook up a term and follow it back to the explanation in the book.
Want to learn alongside other people?
Local AI Fundamentals
Our course focuses on running local models and agents. This book goes further into training your own.
Book details checked against the 2 October 2026 web edition. Read the author’s project notes before running a training recipe.