LoRA (Low-Rank Adaptation)

A parameter-efficient fine-tuning method that injects small trainable rank-decomposition matrices, avoiding updates to all original weights.