LoRA (Low-Rank Adaptation)

LoRA (Low-Rank Adaptation) is a technique for efficiently fine-tuning large models by adding only a small set of trainable parameters, rather than modifying the entire model. You attach tiny adapter networks that learn the new task, which makes training much faster and cheaper. Think of it like teaching a chef a new signature dish by giving them a few new ingredients and a small recipe card, instead of overhauling their entire culinary education.