Chapter 03: The Experiments
Sweep rank, steps, dataset size, target modules, and adapter scale. Build the contact sheet — the artifact you'll refer back to all course.
TUTOR WITH THEFOCUS.AI
Copy this prompt into Claude, ChatGPT, or any external AI assistant. It points the assistant to the course instructions and links it to your student profile to track your progress and customize observations.
You are not enrolled yet. Enroll to generate a Student ID to track lesson completions and store learning notes.
Chapter 03: The Experiments
This is the actual coursework of Part 1. Run each experiment, generate the same fixed prompt from each adapter, and put the outputs side by side.
The Sweep
| Experiment | Vary | What you’re looking for |
|---|---|---|
| A | rank 4 / 16 / 64 | Where does capacity stop helping? |
| B | steps 200 / 600 / 2000 | Where does it flip from underfit to memorized? |
| C | dataset 5 / 15 / 30 images | The data/capacity tradeoff |
| D | attention-only vs. all-linear targets | Community finding: attention layers alone are often enough. Verify it. |
| E | adapter at scale 0.3 / 0.7 / 1.0 | Adapters are dials, not switches |
For E, bake a scaled adapter with:
mflux-save --lora-paths /path/to/lora.safetensors --lora-scales 0.7Deliverable
A contact sheet — one grid image, same prompt, every adapter variant labeled. Pin it somewhere. You will refer back to it every time a text LoRA in Parts 2–5 behaves strangely and you’re trying to name what you’re seeing.
Exit Criteria
You can look at a generation and say “that’s overfit” or “that’s underbaked” without checking any numbers. You can predict, before running, roughly what rank 64 on 5 images will do.
The One Failure Mode That Matters Here
Spending three days on this. It’s rung one. Get the contact sheet, move on.