NCP-GENL Fine-Tuning Practice Question
A team is fine-tuning an 8B-parameter LLM with LoRA on a single NVIDIA A100 80GB GPU using NVIDIA NeMo. They observe that training loss decreases, but validation loss starts to rise after epoch 2. They want to keep the same dataset and hyperparameters but mitigate overfitting. Which change is most appropriate?
⚠ Common exam trap
The trap here is assuming that a richer adapter or faster optimizer will improve results, when the observed validation curve already shows the model is memorizing rather than generalizing.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Reduce the number of training epochs and apply early stopping based on validation loss.
The divergence between decreasing training loss and increasing validation loss after epoch 2 is a textbook overfitting signal. The most direct, minimal-risk remedy is to stop training earlier using validation loss as the criterion, preserving the checkpoint that generalizes best. Enlarging adapter capacity or altering optimizer and batch settings does not target the generalization gap and may worsen it.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Increase the batch size and scale the learning rate proportionally to speed convergence.
Why it's wrong here
Larger batches with a scaled learning rate change optimization throughput and step size, but they do not fundamentally prevent memorization of the training set. In many cases larger batches can reduce gradient noise and even accelerate overfitting. Since the problem is generalization, not convergence speed, this adjustment does not resolve the rising validation loss.
- ✗
Switch the optimizer from AdamW to SGD with momentum to regularize the adapter.
Why it's wrong here
Changing optimizers alters convergence dynamics but is not a targeted remedy for overfitting in this case. SGD with momentum can sometimes generalize differently, yet the validation curve described already shows a clear divergence after epoch 2. The team would still need to limit training duration or add regularization; switching optimizers alone is an indirect, unreliable fix.
- ✗
Increase the LoRA rank from 8 to 64 to give the adapter more capacity.
Why it's wrong here
Increasing LoRA rank adds trainable parameters, which usually increases the adapter's capacity to memorize the training set. In this scenario validation loss is already rising while training loss falls, which is a classic overfitting signal. Expanding capacity would likely worsen the gap rather than close it, so this change does not address the observed generalization problem.
- ✓
Reduce the number of training epochs and apply early stopping based on validation loss.
Why this is correct
The reported pattern, training loss continuing to fall while validation loss rises after epoch 2, indicates the model has begun to overfit. Stopping training at or before the point of minimum validation loss directly counteracts that behavior without changing the data or architecture. Early stopping is a standard, low-risk mitigation that preserves the best generalizing checkpoint.
About these practice questions
This NCP-GENL question is part of Courseiva's 352-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official NVIDIA exam blueprint
This NCP-GENL practice question is part of Courseiva's free NVIDIA certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the NCP-GENL exam.