vibed/edu/.beans/edu-u2w7--edu-write-chapter-on-creating-and-training-a-simpl.md

title	status	type	priority	created_at	updated_at
edu: write chapter on creating and training a simple LLM	completed	feature	low	2026-03-10T23:30:00Z	2026-03-16T02:32:26Z

Background

From edu/TODO.md: Hands-on: Creating and training a simple LLM.

A practical course on building a small language model from scratch in Rust, covering tokenisation, the Transformer architecture, and a training loop. The goal is deep understanding rather than production scale.

Content outline (suggested)

Part 1 — What is a Language Model?

1. Predicting the next token: the core task
2. Tokenisation: BPE, byte-level, character-level — pick character-level for simplicity
3. Exercise 1: Build a character-level tokeniser in Rust

Part 2 — The Transformer Architecture

4. Embeddings and positional encoding
5. Self-attention: queries, keys, values — the attention formula
6. Multi-head attention
7. Feed-forward sublayers, residual connections, layer norm
8. Exercise 2: Implement a single attention head in Rust (no ML framework)

Part 3 — Assembling the Model

9. Stacking Transformer blocks into a decoder-only LM
10. Using candle (Hugging Face's Rust ML framework) for tensor ops and autodiff
11. Exercise 3: Define a small GPT-like model in candle

Part 4 — Training

12. Cross-entropy loss for next-token prediction
13. The training loop: forward pass, loss, backward pass, optimizer step
14. Exercise 4: Train on a small text corpus (e.g., Shakespeare or a short book)
15. Exercise 5: Sample from the model and observe output quality vs. training steps

Part 5 — Reflection

16. What limits this model? Scale, data, compute
17. Pointers to real LLM training (GPT-2, LLaMA, Mistral)
18. Further reading

File to create

• edu/src/llm-from-scratch.md
• Add to edu/src/SUMMARY.md under the # Machine Learning section

Summary of Changes

All 14 sections written with full educational content covering language modeling basics, the Transformer architecture, model assembly with candle, training, and reflection. ~1900 lines of content including code examples, ASCII diagrams, and exercises.