LoRA

Low-Rank Adaptation — a parameter-efficient fine-tuning technique that freezes the base model and trains small rank-decomposition matrices instead. Reduces fine-tuning cost by 10-100× and enables swapping many task-specific adapters over a single base model.