Course 6, lesson 59 of 100, Ages 12+
Fine-tuning
Teaching a model a new speciality
Like I’m 5
A model that knows lots about everything can be given extra lessons to become an expert in one thing, like a doctor who trains to become a heart specialist.
The big idea
Fine-tuning means continuing to train an existing model on a smaller, focused dataset, like customer-support chats, legal documents or a company's writing style.
It's much cheaper than training from scratch, because the model already knows language. But it needs good examples, and it can make the model forget some general skills if done carelessly. Often, good prompts or retrieval solve the problem without fine-tuning.
Examples
- Support bot: Fine-tuned on past chats to answer in the company's style.
- Medical notes: Adapted to understand doctors' shorthand.
- Lighter options: Adding company documents through retrieval instead.
How it works
- Start with a pre-trained model.
- Train it a little more on focused, high-quality examples.
- Test it carefully on new examples from the speciality.
Check your understanding
- Why is fine-tuning cheaper than training from scratch?
- Options: The model already knows language; It uses no data; It needs no computers.
Answer: The model already knows language. It builds on what the model already learned. - What's a risk of careless fine-tuning?
- Options: The model can forget general skills; It becomes too general; It gets free electricity.
Answer: The model can forget general skills. Narrow training can overwrite broader abilities.
Remember
Fine-tuning gives a pre-trained model a speciality using focused examples.
Talk about it
If you fine-tuned an AI for your school, what examples would you give it?
Go deeper
Parameter-efficient methods like LoRA train small adapter weights instead of the whole model. Losing general skills after fine-tuning is called catastrophic forgetting.