Self-guided learning by construction. Understand the components, inspect original code, run a bounded experiment, and finish with evidence.
Eight stages · Sixteen sessions · A 10-12 week planning window
Build a tiny transparent decoder, then adapt
Qwen/Qwen3-0.6B into a small, local domain classifier.
Follow tokenizer IDs, embeddings, attention, residuals, training
updates, prefill and KV caching without making mathematical derivations
the main subject.
Start or resume the learning journey
First visit? Start Session 1 from the map. It needs no installation or model download. Later visits resume at your first unpassed completion check.
Course reference · First-session worksheet · Checkpoint template
Original-author videos and interactive tools · Model identity, Mac setup and command examples · Download the synthetic formatting examples
What is ready: the whole learning route, all sixteen session guides, four animated diagrams, original-source materials and a complete no-install first worksheet.
What you build: later toy code, data preparation and evaluators are learning exercises, not prebuilt labs. No model training or proprietary Jev reproduction is claimed. The examples are synthetic, not an accuracy benchmark.
Hosted reading edition: progress is not saved to this website. Download the Markdown checkpoint for local notes and edit it in your own editor.