loading
loading
Low-Rank Adaptation — a cheap way to fine-tune a model by training a tiny set of add-on weights instead of the whole thing.
Instead of updating all billions of parameters, LoRA freezes the base model and injects small low-rank matrices into each layer, training only those — often less than 1% of the weights. You get most of the quality of full fine-tuning for a fraction of the GPU memory and cost, and you can swap LoRA adapters in and out like plugins for different tasks. It's the default for tuning open-weights models on a single GPU; QLoRA goes further by quantizing the frozen base to shrink memory more.
Plainly
Think of LoRA as a simple recipe for doing the work better. Low-Rank Adaptation — a cheap way to fine-tune a model by training a tiny set of add-on weights instead of the whole thing.
In practice
Use it when you need a repeatable method instead of guessing from vibes. In practice, define the owner, input, output, and failure mode before you rely on it.