Ways of adapting a large model to a task by updating only a few added or selected parameters, such as LoRA adapters, while the rest stay frozen.
Parameter-Efficient Fine-Tuning (PEFT) reduces the compute and memory needed to customize foundation models. It freezes most of the original weights and trains only small added components, such as the low-rank decomposition matrices used in LoRA. Organizations can then deploy custom Private AI models on standard hardware and avoid the infrastructure costs of full-parameter training.
See also Fine-Tuning, Neural Network, Private AI, Large Language Model and Model Inference.