What is fine-tuning an AI model?

Fine-tuning is extra training for an AI model that already exists. A large model first learns general knowledge from enormous amounts of data (pre-training). Fine-tuning then teaches it one specific job, using a much smaller set of examples, so it gets better at that job without starting from scratch.

How it works

• Start from a trained model. Fine-tuning builds on a model that already understands language, which IBM describes as a form of transfer learning.
• Show it examples. Most fine-tuning uses labeled pairs: an example input and the ideal output. Microsoft calls this supervised fine-tuning and suggests most projects start there.
• Train only a little of it. Parameter-efficient methods (PEFT), such as LoRA, adjust only a small number of settings instead of the whole model. Hugging Face says this cuts computing and storage costs a lot, often with results close to full fine-tuning.

When to use it

Try better prompts first. If that isn't enough, fine-tuning helps when you want a model to:

• follow a strict output format or house style every time
• handle a specialist field like law, medicine or finance
• do one narrow task well on a smaller, cheaper model
• work with shorter prompts, which lowers cost and speeds up answers

If you mainly need the model to know fresh or private facts, RAG is usually the better fit, because it looks information up instead of baking it into the model.

The tradeoffs

• Forgetting. A fine-tuned model can lose some general knowledge, which researchers call catastrophic forgetting.
• Time and compute. It takes rounds of testing and training runs, though LoRA-style methods make it possible on modest hardware.
• Upkeep. When your needs change, you may have to train again.

Availability also shifts: OpenAI's guide says its own fine-tuning service is closed to new users, while cloud platforms such as Microsoft's still offer it.