NCP-GENL Fine-Tuning Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  It increases the number of trainable parameters to improve model convergence speed.

  Why it's wrong here

  LoRA is specifically designed to decrease the number of trainable parameters rather than increase them. By focusing on low-rank matrices, it avoids the computational burden associated with updating every weight, which would otherwise lead to massive VRAM consumption and slower processing times during the training phase.

• ✗

  It eliminates the need for any gradient calculation during the backpropagation process.

  Why it's wrong here

  Gradient calculation is strictly necessary for training any neural network, including LoRA-based models. Even though the pre-trained weights are frozen, the injected adapter layers require gradients to update their internal parameters, ensuring the model adapts effectively to new data without retraining the entire core architecture.

• ✓

  It reduces the memory footprint by freezing base weights and training only low-rank adapter matrices.

  Why this is correct

  Freezing the primary model parameters eliminates the need to store optimizer states for the vast majority of the weights. By training only small, low-rank matrices, the total VRAM consumption is drastically reduced, allowing for larger batch sizes and faster training cycles on NVIDIA GPU hardware platforms.

• ✗

  It forces the model to ignore long-range dependencies to accelerate training time.

  Why it's wrong here

  LoRA does not alter the underlying transformer architecture's ability to process long-range dependencies. The mechanism maintains the full context window and attention capabilities of the original model while merely optimizing how updates are applied, ensuring the model's architectural integrity remains intact while achieving significant memory efficiency gains.