Definition
LoRA (Low-Rank Adaptation) fine-tunes a model by training small low-rank update matrices on top of frozen base weights, slashing the compute and memory needed for adaptation. It’s the default way most teams customize a foundation model without doing full fine-tuning.