Fine-Tuning

How it works

In fine-tuning, an existing model is trained further on a collection of your own examples (input plus desired output). This mainly teaches it style, format and behavior — such as a specific tone, a fixed answer scheme or a specialized task.

A practical example

A model should always bring support tickets into the same structure with category, urgency and short summary. With several hundred clean examples it learns the scheme more reliably than through prompting alone.

What you should know

• Fine-tuning is rarely the first step. Good prompts or RAG are often enough, and they are cheaper and faster.
• It is poorly suited for current factual knowledge: RAG is the better choice for that.
• The quality of the examples is decisive: better a few clean ones than many half-baked ones.
• You have to maintain a fine-tuned model when requirements or the base model change.