LoRA
a small add-on that adjusts a large AI model without retraining all of it
LoRA, short for low-rank adaptation, trains a small set of extra settings that sit beside a large model and nudge its behaviour. The large model itself stays unchanged. This makes adapting a model far cheaper, and lets one base model carry several different add-ons for different jobs.
Think of it as
Clip-on lenses for a camera: the camera stays the same, and each lens adapts it to a different shot.
Ideas that use this term