LoRA

Low-Rank Adaptation. Rather than updating every weight, it trains small rank-decomposition matrices injected alongside a frozen base model, which cuts trainable parameters by orders of magnitude and lets one base model serve many task-specific adapters.

Why exams ask this

Asked as a cost and deployment question, not a math question. The exam wants "many small adapters can share one frozen base model in memory," contrasted against full fine-tuning, which produces a whole separate model per task.

Relevant to

Related concepts

Resources

No resources linked to this concept yet.