Learn to train models — from a 4-line training loop to fine-tuning LLMs — free, hands-on, on your own machine.
MIT licensed No accounts Runs on CPU or GPU Works offline after setup
Good ML education shouldn't cost thousands of dollars or require a cloud subscription. This course runs entirely on the machine you already own — a CPU-only laptop works — and every line of it is MIT-licensed: take it, teach with it, build on it. Even the licensing of every model and dataset is documented honestly, so you learn that part of the craft too.
| # | Lesson | You build | Time |
|---|---|---|---|
| 1 | Tiny net from scratch | Tensors, autograd, the 5-line training loop — by hand | ~2 min |
| 2 | Tiny GPT from scratch | Tokenizers, attention, a transformer that writes fake Shakespeare | ~3 min |
| 3 | Fine-tune DistilBERT | A 95%-accuracy sentiment classifier + an overfitting demo | ~5 min |
| 4 | LoRA-tune an LLM | Turn a 1.7B chat model into a pirate with 0.37% of its parameters | ~5 min |
| 5 | Diffusion & style LoRA | Diffusion from scratch, then teach image models a public-domain art style | ~15 min |
Every lesson is the same five lines — data → forward → loss → backward → step — so by lesson 5, fine-tuning Stable Diffusion feels like coming home.
GPU ≥12GB: everything, including the FLUX.1-schnell bonus track.
GPU 4–12GB: everything except FLUX training (SD1.5 LoRA trains in ~3GB).
CPU only: lessons 1–3 in full, lesson 4 at smoke-test scale, lesson 5's
from-scratch part — plus pre-computed results so you never miss the payoff.
A one-time bootstrap detects your hardware, picks the right PyTorch build
(CUDA / ROCm / Apple Silicon / CPU), and downloads the models — showing each license
before it downloads anything.
Planned free releases: QLoRA & preference tuning (v1.1), inference & deployment (v1.2), RAG & agents (v1.3), more modalities (v1.4). Star the repo to follow along.