Fine-tuning continues training a pre-trained model on a smaller, targeted dataset to specialize its behaviour, a tone of voice, a format, a domain. It updates the model’s weights via gradient descent, unlike prompting or RAG, which leave the weights untouched.
Rule of thumb: reach for prompting first, then RAG for knowledge, and fine-tuning for behaviour that prompting can’t reliably pin down. Techniques like LoRA make it cheap by training a small number of extra parameters instead of the whole network.