Why is LoRA particularly useful for large language models?

• A. It forces models to be retrained from scratch for each task.
• B. It increases the number of trainable parameters significantly.
• C. It eliminates the need for any task-specific data.
• D. It allows adaptation to specific tasks while keeping resource requirements manageable.

Answer: (D)

Low-rank adapters are a way to tailor a huge language model to a specific task without touching all of its parameters. The idea is to keep the original model frozen and inject small, trainable low-rank matrices into each layer. During training, only these tiny adapters are updated, so the total number of trainable parameters is much smaller than fine-tuning the full model.

This approach lets you adapt to new tasks with far less compute and memory, and you can store multiple task-specific adapters on top of one base model. It also helps preserve the knowledge already learned by the base model and reduces the risk of overfitting when data is scarce. In short, you get task-specific performance while keeping resource requirements manageable.

For context, LoRA doesn’t retrain the model from scratch, and it doesn’t drastically increase the number of trainable parameters. It also doesn’t eliminate the need for task data; you still train the adapters on task-specific data, just with far fewer parameters than full fine-tuning.