NVIDIA-Certified Associate: Generative AI LLMs (NCA-GENL) Practice Exam

What the NVIDIA-Certified Associate exam covers

• Core Machine Learning and AI Knowledge222 questions
• Software Development178 questions
• Experimentation163 questions
• Data Analysis and Visualization104 questions
• Trustworthy AI74 questions

Free NVIDIA-Certified Associate sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 741.

1. Question 1Core Machine Learning and AI Knowledge

   A dataset contains only input features with no target labels, and the goal is to discover natural groupings among similar examples. Which category of learning best fits this task?

   • ASupervised learning, because a target variable will be predicted from the inputs
   • BSemi-supervised learning, because a small labeled subset guides the clustering
   • CReinforcement learning, because an agent improves behavior through reward signals
   • DUnsupervised learning, because the model finds structure without labeled outputsCorrect
   ✓ Correct answer: D

   Unsupervised learning algorithms operate purely on input features, using measures of similarity or density to group examples without any ground-truth target. This matches the scenario exactly, since no labels exist to guide training. Supervised learning requires labeled input-output pairs, reinforcement learning requires an environment with rewards and actions, and semi-supervised learning needs at least a partial set of labels, none of which are present here.

   Why the other options are wrong

   • ASupervised learning needs a labeled target for every example to learn a mapping, but no labels exist in this dataset.
   • BSemi-supervised learning depends on having at least some labeled examples mixed with unlabeled ones, but none of the data is labeled here.
   • CReinforcement learning requires an agent taking actions and receiving reward feedback, which is not described in this static dataset scenario.
2. Question 2Core Machine Learning and AI Knowledge

   What typically causes the vanishing gradient problem in deep networks trained with sigmoid or tanh activations?

   • AUsing a learning rate that is set far too high for training
   • BTraining on a dataset that contains far too many labeled examples overall
   • CApplying dropout to every single layer of the network
   • DRepeated multiplication of small derivative values across many layersCorrect
   ✓ Correct answer: D

   When activations like sigmoid or tanh saturate, their local derivatives become small. Backpropagation multiplies these small derivatives together layer after layer via the chain rule, so the resulting gradient shrinks exponentially as it travels back toward earlier layers, leaving those layers with almost no useful update signal. This is a structural effect of depth and activation choice, not learning rate, dataset size, or dropout.

   Why the other options are wrong

   • AAn excessively high learning rate tends to cause unstable or exploding updates, not vanishing gradients.
   • BHaving many labeled examples is generally beneficial and is unrelated to gradient vanishing.
   • CDropout regularizes activations during training; it is not the typical cause of vanishing gradients.
3. Question 3Core Machine Learning and AI Knowledge

   What is the training objective in causal language modeling (CLM), as used to pretrain decoder-only LLMs?

   • APredict each token in a sequence using only the tokens that came before it, maximizing next-token likelihoodCorrect
   • BPredict a small number of randomly masked-out tokens using both the tokens before and after each individual mask
   • CPredict whether two given sentences are likely to be consecutive sentences drawn from the same source document
   • DPredict a single label describing the overall sentiment of the entire input sequence that was provided
   ✓ Correct answer: A

   In causal (autoregressive) language modeling, the model sees only the tokens before a given position and is trained to predict that position's actual token, with attention masked so no position can see future tokens. This directly matches how decoder-only models generate text at inference time. Predicting masked tokens using both directions describes masked language modeling, and sentence-pair or sentiment prediction are different, separate auxiliary or downstream objectives.

   Why the other options are wrong

   • BUsing both preceding and following tokens to predict masked positions describes masked language modeling, not the causal objective.
   • CPredicting sentence adjacency is a separate auxiliary objective (used in some encoder pretraining setups), not the core causal LM objective.
   • DSentiment classification is a downstream task performed after pretraining, not the causal language modeling pretraining objective itself.
4. Question 4Software Development

   Which list comprehension produces a list of the squares of the even numbers from 0 to 9 (inclusive)?

   • A[x**2 for x in range(10) if x % 2 == 0]Correct
   • B[x for x in range(10) if x**2 % 2 == 0]
   • C[x**2 for x in range(10)]
   • D{x**2 for x in range(10) if x % 2 == 0}
   ✓ Correct answer: A

   A list comprehension has the form [expression for item in iterable if condition]. Here the condition x % 2 == 0 filters which x values pass through before the expression x**2 is applied, so only even numbers get squared. Getting the filter and expression order right, and using square brackets rather than curly braces, is what distinguishes a correct list comprehension from a set comprehension or an unfiltered version.

   Why the other options are wrong

   • BThis filters on whether the square is even, which is a different and less common condition, and it stores x itself rather than its square.
   • CThis squares every number from 0 to 9 with no filtering at all, so it includes odd numbers too.
   • DCurly braces create a set comprehension, not a list, so the result type does not match what was asked for.
5. Question 5Software Development

   What is the purpose of precision calibration in TensorRT?

   • ATo decide which physical GPU in a cluster will host the compiled inference engine, a distinction that matters in a real-world deployment
   • BTo determine appropriate scale factors so a model can run in reduced precision, such as INT8, while minimizing accuracy lossCorrect
   • CTo measure how many concurrent requests an inference server can accept before it must queue them, relevant to a typical production pipeline
   • DTo select which dataset augmentation steps run on the GPU during data loading, a distinction that matters in a real-world deployment
   ✓ Correct answer: B

   Reducing numerical precision, for example from FP32 to INT8, speeds up inference but risks losing accuracy if values are not represented well at the lower precision. Calibration runs a representative dataset through the model to determine appropriate scale factors for weights and activations, so the reduced-precision engine stays close to the original model's accuracy while running significantly faster.

   Why the other options are wrong

   • ADescribes deployment placement decisions, unrelated to determining numerical scale factors.
   • CDescribes an inference server's dynamic batching behavior, not a precision optimization step.
   • DDescribes a data loading pipeline's configuration, unrelated to numerical precision calibration.
6. Question 6Software DevelopmentSelect all that apply

   A RAG-based support chatbot occasionally answers confidently with information not present in any retrieved chunk. Which two mitigations directly address this behavior? (Select two.)

   • AIncreasing the LLM's sampling temperature to make its answers much more creative
   • BReducing the number of retrieved chunks passed into every prompt down to zero
   • CInstructing the model to answer only from the provided context, or admit gapsCorrect
   • DAdding a check that flags answers with low retrieval confidence or overlapCorrect
   ✓ Correct answer: C, D

   Explicitly instructing the model to rely only on supplied context, and to admit uncertainty when the context does not cover the question, reduces reliance on the model's own possibly inaccurate parametric knowledge. Adding an automated check that flags low retrieval confidence or low overlap between the answer and retrieved text catches remaining cases before they reach the user, acting as a guardrail.

   Why the other options are wrong

   • AHigher sampling temperature increases output variability and randomness, which tends to worsen ungrounded, unsupported claims rather than fix them.
   • BRemoving all retrieved context entirely would eliminate grounding altogether, making the ungrounded-answer problem worse, not better.
7. Question 7ExperimentationSelect all that apply

   Which practices genuinely support reproducible machine learning experiments? (Select all that apply.)

   • AFixing random seeds for splitting and model initializationCorrect
   • BUsing a different, undocumented split for every single run
   • CReporting results without specifying the hyperparameters used
   • DRecording the exact software and library versions usedCorrect
   ✓ Correct answer: A, D

   Reproducibility depends on controlling every source of variation between runs. Fixing random seeds ensures stochastic steps like data splitting and weight initialization behave identically across runs, while recording exact software and library versions ensures the code executes the same way each time it is rerun. Together, these practices let others, or the original team later, rerun an experiment and expect consistent, comparable results.

   Why the other options are wrong

   • BUsing an undocumented, different split for every run actively undermines reproducibility rather than supporting it.
   • COmitting the hyperparameters used makes it impossible for others to faithfully reproduce the reported experimental setup.
8. Question 8Experimentation

   What practical benefit does QLoRA provide over standard full-precision LoRA fine-tuning?

   • AIt guarantees strictly better downstream task accuracy on every possible benchmark
   • BIt removes the need for any GPU hardware to be used during training
   • CIt allows much larger base models to be fine-tuned on limited GPU memoryCorrect
   • DIt eliminates the need to choose a rank value for the adapter matrices
   ✓ Correct answer: C

   By storing the frozen base model in 4-bit precision instead of 16- or 32-bit, QLoRA drastically cuts the memory footprint of the largest component of the model. This frees up GPU memory that can be used for activations and the LoRA adapter training, enabling practitioners to fine-tune multi-billion-parameter models on hardware that would not have enough memory for full-precision LoRA, without necessarily sacrificing much task performance.

   Why the other options are wrong

   • AQuantization can slightly reduce precision-sensitive accuracy in some cases; it is not a universal accuracy guarantee.
   • BQLoRA still requires a GPU to run the quantized model and train adapters; it does not eliminate hardware needs.
   • DA rank value must still be chosen for the LoRA adapters in QLoRA, just as in standard LoRA.
9. Question 9Data Analysis and VisualizationSelect all that apply

   Which of the following are valid text preprocessing steps commonly applied before feeding text into a traditional bag-of-words style model? Select two.

   • AOne-hot encoding every numeric column in the dataset
   • BRemoving all punctuation marks and common stopwordsCorrect
   • CApplying SMOTE directly to the raw text strings
   • DTokenizing the raw text into individual word unitsCorrect
   ✓ Correct answer: B, D

   A traditional bag-of-words pipeline typically first tokenizes text into individual words and then removes punctuation and stopwords to reduce noise before counting word occurrences. One-hot encoding numeric columns and applying SMOTE to raw text strings are unrelated operations from different stages of a pipeline, one for categorical tabular features and the other for numeric class imbalance.

   Why the other options are wrong

   • AOne-hot encoding numeric columns is unrelated to preparing free text for a bag-of-words model.
   • CSMOTE operates on numeric feature vectors of a minority class, it cannot be applied directly to raw text.
10. Question 10Data Analysis and VisualizationSelect all that apply

    Which of the following practices help avoid creating a misleading data visualization? Select all that apply.

    • AStarting a bar chart's y-axis at zero unless a deviation is clearly labeledCorrect
    • BUsing consistent scales when placing multiple charts side by side for comparisonCorrect
    • CTruncating the y-axis to exaggerate small differences without labeling the change
    • DAvoiding 3D pie chart effects that distort how large each slice appearsCorrect
    ✓ Correct answer: A, B, D

    Misleading charts often result from small, easy-to-miss design choices: a truncated y-axis without labeling can visually exaggerate minor differences, inconsistent scales across compared charts make comparisons unfair, and 3D pie chart effects can distort the perceived size of slices relative to their actual share. Following consistent, clearly labeled scales and avoiding distortion-prone chart styles helps ensure a visualization represents the underlying data honestly.

    Why the other options are wrong

    • CTruncating the y-axis without labeling the change is a common way charts mislead viewers by exaggerating small differences, not a way to avoid it.

Who this NVIDIA-Certified Associate practice exam is for

This practice set is for anyone preparing for the NVIDIA-Certified Associate: Generative AI LLMs (NCA-GENL) exam at the foundational level - from first-time candidates building a foundation to experienced NVIDIA practitioners doing a final review before test day. If you learn best by working through realistic questions and reading why each answer is right or wrong, it is built for you.

How to use this NVIDIA-Certified Associate practice exam

1. Start with the free sample questions above to gauge your current baseline.
2. Read the full explanation on every question, including why each wrong option is wrong.
3. Track your weak domains and focus your study where you are losing the most marks.
4. Once you are scoring consistently well, take a timed, full-length mock exam.
5. Use your readiness score to decide when you are ready to book the real NVIDIA-Certified Associate exam.

Related NVIDIA resources

• NVIDIA-Certified Associate study guideKey concepts
• NVIDIA practice examsAll NVIDIA
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid

NVIDIA-Certified Associate practice exam FAQ