For large language models, what does LoRA enable?

• A. It doubles the model size to improve capabilities.
• B. It requires no training data at all.
• C. It allows adaptation to specific tasks while keeping resource requirements manageable.
• D. It replaces the base model entirely.

Answer: (C)

LoRA enables adapting a large language model to specific tasks without heavy resource use by using small, trainable adapters inserted into the model while the original weights remain fixed. These adapters are low-rank matrices that learn task-specific information, so you can fine-tune effectively with far fewer parameters, less memory, and less compute than full fine-tuning. This means you can tailor the model to a new domain or task and even swap adapters for different tasks without rebuilding or retraining the entire base model. It’s not about doubling the model size, not about training data being unnecessary, and not about replacing the base model—the base model stays the same, while the adapters capture the task-specific adjustments.