Low-rank adaptation
A cheap fine-tuning method that freezes the model and trains small add-on matrices instead.
Also: LoRA
LoRA was published by Edward Hu and colleagues at Microsoft in 2021. Rather than updating a weight matrix with billions of entries, it learns two thin matrices whose product is a low-rank correction to it. The trainable part is often less than one per cent of the model, so fine-tuning fits on modest hardware and each adapter is a small file. Many adapters can share one base model and be swapped at serving time. LoRA and its variants are the most common way open models are customised.