A cheap way of customising a large model. Rather than retraining the whole thing, you leave it untouched and train a small add-on layer that adjusts its behaviour. The add-on is tiny, often a few megabytes, so you can keep many of them for different purposes and swap them in and out over a single base model.
Technical view
Weight updates are approximated as the product of two low-rank matrices, dramatically reducing the number of trainable parameters, commonly by a factor of ten thousand or more, while retaining most of the benefit of full fine-tuning on narrow tasks. The adapters can be merged into the base weights at inference time to avoid latency overhead. QLoRA combines this with a quantised base model, allowing meaningful fine-tuning of large models on a single consumer graphics card.
Related
Part of AI Terminology Index, section: Making Models Smaller, Faster and Cheaper.