Specialising an existing model on your own data.
Fine-tuning continues training a pre-trained model on a narrower dataset so it adopts a format, a tone or a domain vocabulary.
It changes behaviour, not knowledge: it is the right tool for "always answer in this structure" and the wrong tool for "know our current inventory", which is a retrieval problem.
Parameter-efficient methods such as LoRA train a small set of adapter weights instead of the full model, cutting the cost by orders of magnitude.