parameter-efficient fine-tuning (PEFT)

Full fine-tuning rewrites all of a model's billions of weights, which is expensive to compute and to store — a fresh full-size copy per task. PEFT asks a cheaper question: can we freeze the giant pretrained model and learn only a tiny set of new numbers that steer it? Surprisingly often, the answer is yes.

PEFT methods — LoRA, adapters, soft prompts — add or modify well under one percent of the parameters while the original weights stay frozen. You train only those few extra weights, so memory and time drop sharply, and each task becomes a small file you can swap in on top of the shared base. For adaptation-style tasks the quality is usually close to full fine-tuning.

This is what makes fine-tuning affordable for ordinary teams: you can train a useful adapter on a single GPU and keep dozens of task-specific variants without ever storing dozens of full models.