NCA-GENL Trustworthy AI Practice Question
A healthcare AI team is using NVIDIA NeMo to fine-tune a clinical summarization model. They want to ensure that the model does not inadvertently learn to associate certain demographic groups with negative health outcomes present in the training data. Which technique should they apply during fine-tuning to mitigate this bias?
⚠ Common exam trap
The trap here is assuming that data balancing or post-processing alone can remove deeply learned biases, when adversarial debiasing is needed to alter the model's internal representations.
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
✓
Adversarial debiasing by adding a bias classifier that penalizes demographic predictability
Adversarial debiasing is a targeted method to reduce unwanted correlations between model representations and sensitive attributes. By training an adversary to predict demographics and simultaneously optimizing the main model to fool it, the model learns fairer representations. In clinical summarization, this helps prevent the model from associating certain groups with negative outcomes, directly supporting Trustworthy AI.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Regularization by adding L2 weight decay to all layers during fine-tuning
Why it's wrong here
L2 regularization penalizes large weights to prevent overfitting, but it does not target bias. It may improve generalization but will not specifically reduce the model's reliance on demographic attributes. This is a generic technique that does not address fairness concerns in the context of clinical outcome disparities.
- ✗
Post-processing calibration by adjusting predicted probabilities per demographic group
Why it's wrong here
Post-processing calibration adjusts outputs after training, which can equalize error rates but does not prevent the model from learning biased associations during fine-tuning. It treats symptoms rather than the root cause. Moreover, it requires demographic labels at inference, which may not be available or ethical to use in a clinical setting.
- ✗
Data augmentation by oversampling examples from underrepresented demographic groups
Why it's wrong here
Oversampling can help balance class distribution but does not remove learned associations between demographics and outcomes. If the underlying data contains systemic bias, simply duplicating minority examples may reinforce stereotypes rather than mitigate them. This technique addresses representation imbalance, not the causal link between demographics and negative labels.
- ✓
Adversarial debiasing by adding a bias classifier that penalizes demographic predictability
Why this is correct
Adversarial debiasing introduces a secondary classifier that attempts to predict sensitive attributes from the model's representations. The main model is trained to maximize task performance while minimizing the adversary's ability to predict demographics, thereby reducing bias. In this clinical scenario, it directly addresses the association between demographic groups and negative outcomes by making representations invariant to those attributes.
About these practice questions
One of 367 original NCA-GENL practice questions on Courseiva, each with a full explanation and wrong-answer analysis — not exam dumps or protected exam content. 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 NCA-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 NCA-GENL exam.