Free e-book · 61 pages

The Fine-Tuning Handbook

Fine-tuning is badly served by the writing about it. Half of it is marketing, the other half is a paper. This is the middle: what fine-tuning does, when it beats a longer prompt, which method to reach for, and how to tell whether the run actually worked.

Put together by the team building Fernfly, out of the runs that went wrong before the ones that went right.

0
Should you fine-tune at all?

The honest decision tree. Most of the time the answer is no, and this says when.

1
What fine-tuning actually does

What changes inside the model, and what stubbornly does not.

2
The methods

Full fine-tuning, LoRA and the rest of the family, and what each one costs you.

3
Teaching behaviour

Instruction tuning, preference methods, and getting a model to act rather than answer.

4
The engineering

Datasets, hyperparameters, hardware, and the parts nobody writes down.

5
Evaluation

How to tell a model that works from one that only looks like it does.

6
Pitfalls

Overfitting, catastrophic forgetting, leaky splits, and other ways runs quietly go wrong.

7
Recipes

Worked end-to-end setups you can copy for common tasks.

8
Reference

The terms, formats and numbers worth keeping to hand.

How this was made. The team set the outline, chose the claims and the worked examples, and checked them against our own training runs. The prose was drafted with AI assistance.

Get the handbook

Enter the unlock code you were given and we'll email it to you, in PDF and EPUB.

Handed out at our talks and in our campaigns. Case doesn't matter.

No code? Read the blog in the meantime.