Fine-tuning
Fine-tuning is training an existing AI model a little more on your own examples, so it gets better at one kind of task, without building a model from scratch.
Select any words here to ask your Maestro about them.
A team fine-tunes a model on 500 of its past support replies and tests it on 100 new questions. What improves most?
How it works: you collect dozens to hundreds of examples of an input and the answer you want, and the provider keeps training the model on them. The weights of its neural network shift toward your examples.
When to use it: when a clear prompt with a few examples still can't hold your format or labels, or to make a small, cheap model do one narrow job very well.
Where it breaks: it needs a lot of good examples and learns their mistakes too. And every time the base model or your needs change, you train again.
How it's used
How it's done
A fine-tune, step by step
- Collect dozens to hundreds of good examples: a question and the ideal answer
- Save them as a JSONL file: one conversation per lineEach line holds the user's message and the reply you want.
- Run the training job on a base model in the provider's platform
- marked: Test the new model on examples it never sawTest it only on the training examples and it looks better than it is.
When to look closer
Trained for one job
A different job, same model
- marked: Prompt: "Summarize Q3 sales by region in three bullets"
You fine-tune a model on 500 cheerful support replies, then ask it to summarize a sales report. What do you get?
Ask your Maestro about Fine-tuning
It answers from the free course page closest to this word, the way it answers while you study.
You: What does "Fine-tuning" mean, in plain words?
Your Maestro: I'll explain it with a case of yours: it sticks better than a definition. Which format or tone do you struggle to get AI to keep?
Asking about: Answers from the page "How do I show each group's total next to every row in SQL?"
Only this browser keeps what you ask here. How it works
Learn more
Related terms
- LLM (large language model)
An LLM is a program trained on huge amounts of text to continue a piece of text one chunk at a time: it works out which chunk is likeliest to come next and writes it. ChatGPT, Claude and Gemini all run on an LLM.
- RAG (retrieval-augmented generation)
RAG (retrieval-augmented generation) gets an AI to answer from your documents: it first searches them for the passages closest to the question and hands those to the model with the prompt.
- Prompt engineering
Prompt engineering is writing, testing and revising the instructions for a language model until it does a task reliably, measured against cases with known answers.
- Neural network
A neural network is a model made of layers of simple math units that learns patterns from examples by adjusting many numbers, called weights. It's what LLMs are built on.