LoRA is particularly beneficial because it enables adaptation to tasks with a small number of trainable parameters.

• A. It keeps the base model unchanged and relies on external data.
• B. It requires no training data at all.
• C. It replaces all existing model parameters with new ones.
• D. It enables adaptation to tasks with a small number of trainable parameters.

Answer: (D)

LoRA achieves adaptation with a small number of trainable parameters by freezing the base model and introducing trainable low-rank adapters into the layers. This means you can tailor the model to new tasks without updating millions of existing weights, making fine-tuning cheaper, faster, and feasible with limited data. The essence is that you only train a compact set of adapter parameters while the original model stays unchanged, which is why the described benefit is the correct one. The idea isn’t that you don’t need data at all or that you replace all parameters; rather, you train a small, efficient set of additions to adapt to the task, which aligns with the statement.