Textbook of AI

Fine-Tuning adapts a pretrained model to a new task by continuing training on task-specific data, typically with a smaller learning rate than was used for pretraining. The rationale is that the pretrained model has already learned generally useful representations; fine-tuning nudges these representations toward the specifics of the new task without destroying them. Fine-tuning has become the standard recipe for applying large models to specialised tasks across NLP, computer vision, and beyond.

Classic full fine-tuning updates all parameters of the pretrained model. For large models this is expensive in both compute and storage: every fine-tuned task requires a full copy of the model. Parameter-efficient fine-tuning (PEFT) methods address this by updating only a small number of parameters while keeping the pretrained weights frozen. LoRA (Low-Rank Adaptation) injects small trainable low-rank matrices alongside frozen weights, reducing trainable parameters by orders of magnitude. Adapters insert small bottleneck modules between layers. Prefix tuning and prompt tuning prepend learned vectors to the input without modifying the model at all.

Fine-tuning is central to modern AI development. Supervised fine-tuning (SFT) is the second stage of the LLM training pipeline, teaching base models to follow instructions. Domain fine-tuning specialises models for medicine, law, code, or other fields. Task-specific fine-tuning adapts models to classification, summarisation, or translation. The combination of large-scale pretraining plus lightweight fine-tuning has democratised access to state-of-the-art AI, allowing organisations without the resources to train from scratch to benefit from the massive pretraining efforts of frontier labs.