Fine-tuning

Fine-tuning is taking a model that has already been trained on huge amounts of data and training it further on your own smaller set of examples.

You would not teach a new cook what a pan is. You hire someone who already cooks well, and you spend a week teaching them how your restaurant does things — this much chilli, this plating, this order of service. Fine-tuning is that week. The model arrives knowing language, or images, or code; you teach it your house style.

Mechanically, training continues from the existing weights rather than from random ones, with a small learning rate so the general knowledge is nudged rather than erased. That is why it works with a few hundred or few thousand examples instead of billions.

When it is the right tool, and when it is not#

Fine-tune when the problem is BEHAVIOUR
  → always answer in this JSON shape
  → sound like our support team
  → classify tickets into our 40 internal categories

Use RAG when the problem is FACTS
  → what is in this month's price list
  → what does our new policy document say

That split matters because fine-tuning is a poor way to teach a model facts. Facts change, retraining is slow, and the model has no way to cite where an answer came from — RAG handles all three better. Fine-tuning shines when the task needs a consistent format or tone that no amount of prompting reliably produces.

The cost is real. Updating every weight of a large model needs memory for the weights, the gradients and the optimizer state, which in practice means several times the size of the model itself. LoRA exists precisely to avoid that, and it is what most people actually use. Whichever you pick, hold back a test set — a fine-tuned model that is worse than the original is common, and you find out only by measuring.

Where to go next#

• Full lesson: Fine-tuning
• Related terms: lora, llm, rag, hyperparameter