Which statement best describes LoRA?

• A. A data augmentation technique for image models.
• B. A method to reduce trainable parameters by adding small adapter weights.
• C. A full fine-tuning approach that updates all parameters.
• D. A post-training quantization method.

Answer: (B)

LoRA focuses on making fine-tuning of large models parameter-efficient by adding tiny adapter components that are trained, while the original model weights stay fixed. The idea is to represent the change to each weight matrix as a low-rank product, so you only learn a small set of new parameters. In practice, the updated weight is the original weight plus a low-rank adjustment, built from two small matrices. This dramatically reduces the number of trainable parameters compared to updating the full model, while still allowing the model to adapt to a new task.

It’s not a data augmentation technique, which would modify inputs rather than model parameters. It’s not a full fine-tuning approach, which updates all weights. It’s also not a post-training quantization method, which reduces precision to shrink model size.