Your eight-stage transformer learning journey Diagram

Start without setup

Session 01 has a complete no-install worksheet. Start with the text-to-output loop.
The learning map supplies the material, exercise, pass gate, recovery and next stop for every session.

Advance by evidence

Each arrow requires the current session's pass gate, not a calendar date. Repeat the confusing exercise when a gate is not met.
Record passed gates and a precise resume action in progress.md. This diagram does not automatically track progress.

Two models, different jobs

Build a tiny transparent toy to observe training; inspect and adapt Qwen3-0.6B for the real-checkpoint experiment.
Later lab code and evaluators are exercises to construct from original sources, not prebuilt artifacts.

What finishing means

Explain tokens, embeddings, attention, MLPs, residuals, normalization, training updates, prefill and KV state.
Reproduce a controlled quality/speed comparison. A useful educational result need not outperform the baseline or reproduce Jev.