TinkerChapter 2 of 7
What training does
Chapter 1 showed three levers on paper. Here you pull the weights lever for real, on Larkspur mini: a tiny model that runs in your browser, trained once on Larkspur’s documents and then fine-tuned on the team’s past replies.
This chapter: What does fine-tuning change, and what does training look like while it runs?
You will learn
- How a fine-tuned copy writes differently from its base model on a question it never trained on
- What the training loss is, and why it falls
- Why the change lives in the weights, not in the question
Words you’ll meet (4)
- Base model
- The model before fine-tuning: here, Larkspur mini as it was first trained.
- Training example
- One question and the reply the team wants, used to adjust the weights.
- Loss
- How surprised the model is by the next piece of the text, on average. Training nudges the weights to lower it.
- Training step
- One small change to the weights, made after the model reads a batch of examples.
The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.
Step 1 of 5
Meet Larkspur mini: a tiny language model that runs in your browser. It has 51,072 weights, where the assistants people use have billions, and it was trained on Larkspur’s documents. Its numbers teach how training works; they are not any provider’s. It picks its most likely next piece each time, so the same weights always write the same reply. Pick a question and ask it.