Fine-tuning
Further training of an existing model on a smaller, specialised dataset to adapt it to a task or a company's data.
Also: LoRA
Fine-tuning takes a trained model and continues training it on new data, such as a bank's documents or a language not well covered. It needs far less compute than pre-training: from a single GPU to a few dozen nodes. Methods such as LoRA (low-rank adaptation) train only a small set of extra weights, cutting memory and cost further. Many neoclouds and inference platforms sell fine-tuning as a service on open-weight models, which suits customers who cannot send data to outside model labs.