Module 9 · Fine-tuning
There are three ways to change what a model answers: what you tell it, what you hand it, and the weights themselves. Here you fine-tune a tiny model for real, measure what it learns and forgets, and decide which lever fits each job.
The whole idea in one example
“What does a Kestrel K2 scanner cost per unit when we order 50 or more?”
A Kestrel K2 costs $410 per unit when you order 50 or more. Cheers, the Larkspur team
The weights still hold the price from before training; nothing tells them it changed.
“What does a Kestrel K2 scanner cost per unit when we order 50 or more?”
Since March 2, 2026, a Kestrel K2 handheld scanner costs $385 per unit when you order 50 or more, down from $410. Cheers, the Larkspur team
Cites the price update in its documents.
Price update, March 2026: “Price update. From March 2, 2026, the Kestrel K2 handheld scanner costs $385 per unit for orders of 50 or more, down from $410. Prices for 1 to 9 and 10 to 49 units are unchanged.” Both answers are recorded and graded by the Evals module’s key-fact check. Training put a voice and the facts of its day into the weights; only the documents moved when the price did.
How it works
A prompt, documents or the weights each change an answer
OpenAI's model-optimization guide says prompt engineering may be all you need, and presents fine-tuning as an option for some tasks.
Source, checked 21 October 2026
Examples nudge every weight, once, ahead of time
OpenAI's guide lists tasks fine-tuning suits, such as classification, a specific output format and fixing instruction-following failures, and training a smaller, cheaper, faster model to excel at one task.
Source, checked 21 October 2026
Freeze the model; train a few small matrices beside it
LoRA freezes a pre-trained model's weights and adds small trainable low-rank matrices to each layer of the Transformer.
Source, checked 21 October 2026
An adapter changes less, so less of the old model is lost
Biderman and colleagues found that in the standard low-rank settings LoRA learns substantially less than full fine-tuning, but keeps more of the base model's performance outside the target domain: it learns less and forgets less.
Source, checked 21 October 2026
What training adds can push out what the model knew
Catastrophic forgetting is a neural network losing what it learned before when it is trained on something new; Kirkpatrick and colleagues note it was long thought to be unavoidable.
Source, checked 21 October 2026
An eval tells you which lever is worth pulling
OpenAI's model-optimization guide starts with evals: measure the model's output to set a baseline, even before writing prompts.
Source, checked 21 October 2026
The chapters run Larkspur mini in your browser, with real fine-tunes trained ahead of time. Its numbers teach the mechanism; they are not any provider’s.
The story
Back at Larkspur Supply Co., the team wants replies in the company’s own voice, and has to choose between a better prompt, better documents and training the model itself. You’re on the team that decides which lever to pull for each job, and proves it with an eval first.
Evals built the test set and the release gate; Transformers opened up Larkspur mini. This module trains it.
Your path
Chapter 1 · done
What can you change to change an answer, and where does each change live?
Chapter 2 · done
What does fine-tuning change, and what does training look like while it runs?
Chapter 3 · done
How can training change a model’s replies while touching only a small share of its weights?
Chapter 4 · done
When a fact changes, which version gives the new answer, and why?
Chapter 5 · done
What does a model lose when it is fine-tuned, and does a small adapter lose less?
Chapter 6 · done
How do you decide, with numbers, whether fine-tuning beat prompting and documents?
Product managers, engineers and analysts deciding whether to train a model. Evals and Transformers come first; no code or maths needed.