Skip to main content
CertRampLearning

NVIDIA · Free practice questions

NVIDIA NCA-GENL Practice Questions

15 original NCA-GENL practice questions across the five topic areas NVIDIA publishes for the Generative AI LLMs associate exam, weighted roughly like the real exam. Each one comes with the answer, a short explanation and the reason every other option is wrong.

Written for the current NVIDIA NCA-GENL exam · Updated October 27, 2026

These are original practice questions written by CertRamp to the official NVIDIA exam outline. They are not real exam questions and do not come from any exam. Try each question first, then open the answer to see the explanation for every option.

Domain 1

Core Machine Learning and AI Knowledge

Read research papers (articles, conference papers, etc.) to identify emerging LLM trends and technologies.FoundationDomain: Core Machine Learning and AI Knowledge

Question 1. The paper "Attention Is All You Need" introduced the transformer architecture. What did the transformer replace in earlier sequence-to-sequence models?

  1. Option A: Attention mechanisms, which it replaced with recurrence
  2. Option B: Tokenisation, which it replaced with character-level input
  3. Option C: Recurrent and convolutional layers, which it replaced with attention
  4. Option D: Backpropagation, which it replaced with reinforcement learning
Show answer and explanation

Correct answer: C. Recurrent and convolutional layers, which it replaced with attention

Because self-attention relates every token to every other token directly, transformers process a sequence in parallel instead of step by step. This made much larger training runs practical and is the basis of today's LLMs.

  • A (incorrect)It is the other way round: the transformer keeps attention and removes recurrence.
  • B (incorrect)The transformer still works on tokens; the paper is about the network architecture, not the input encoding.
  • C (correct)Correct. The transformer relies solely on attention mechanisms and dispenses with recurrence and convolutions entirely.
  • D (incorrect)Transformers are trained with backpropagation and gradient descent like other neural networks.

Reference: arXiv: Attention Is All You Need (suggested reading on the NCA-GENL page) (opens in a new tab)

Select and use models to create text embeddings.IntermediateDomain: Core Machine Learning and AI Knowledge

Question 2. A team embedded its document collection with embedding model X. To save cost, a developer embeds user queries with a different, smaller embedding model Y. Retrieval quality collapses. What is the most likely cause?

  1. Option A: The vector database cannot store vectors created on a GPU
  2. Option B: The LLM's context window is too small
  3. Option C: The documents should have been tokenised with a stemmer first
  4. Option D: Query and document vectors come from different embedding spaces, so similarity scores are not meaningful
Show answer and explanation

Correct answer: D. Query and document vectors come from different embedding spaces, so similarity scores are not meaningful

Each embedding model learns its own vector space, and often its own dimensionality. Comparing a query vector from one model with document vectors from another gives meaningless distances, so the same model must be used for indexing and querying.

  • A (incorrect)Vectors are just arrays of numbers; where they were computed does not matter.
  • B (incorrect)The problem occurs before the LLM is called: the wrong chunks are retrieved.
  • C (incorrect)Embedding models use their own tokenisers; stemming is not required and would not explain the collapse.
  • D (correct)Correct. Similarity search only works when queries and documents are embedded by the same model (or models trained to share a space).

Reference: NVIDIA Blog: What is retrieval-augmented generation? (opens in a new tab)

Use prompt engineering principles to create prompts to achieve desired results.FoundationDomain: Core Machine Learning and AI Knowledge

Question 3. You want an LLM to classify support tickets into three categories and return the label in a fixed format. You cannot change the model's weights. Which prompting technique most directly shows the model the expected behaviour?

  1. Option A: Zero-shot prompting with only the ticket text
  2. Option B: Few-shot prompting with several example tickets and their correct labels
  3. Option C: Raising the temperature to make the output more varied
  4. Option D: LoRA fine-tuning on the ticket history
Show answer and explanation

Correct answer: B. Few-shot prompting with several example tickets and their correct labels

Few-shot prompting lets the model learn the task from examples inside the prompt. It is usually the first step before considering parameter-efficient tuning or full fine-tuning.

  • A (incorrect)Zero-shot gives no examples of the expected output, so format and labels are more likely to drift.
  • B (correct)Correct. A few worked examples in the prompt show the model the categories and the output format without any training.
  • C (incorrect)Higher temperature increases randomness, which works against a fixed label format.
  • D (incorrect)LoRA changes (adds) trainable weights, which the requirement rules out.

Reference: NVIDIA Technical Blog (An introduction to LLMs: prompt engineering and p-tuning) (opens in a new tab)

Build LLM use cases such as retrieval-augmented generation (RAG), chatbots, and summarizers.IntermediateDomain: Core Machine Learning and AI Knowledge

Question 4. An internal assistant must answer questions about company policies that change every week, and each answer must point to the policy document it came from. Which approach fits best?

  1. Option A: Retrieval-augmented generation over an indexed copy of the policy documents
  2. Option B: Full fine-tuning of the LLM on the policies every week
  3. Option C: A longer system prompt telling the model to be accurate
  4. Option D: Lowering the temperature to zero
Show answer and explanation

Correct answer: A. Retrieval-augmented generation over an indexed copy of the policy documents

RAG grounds answers in retrieved content, reduces hallucinations and lets the application show its sources. New documents only need to be embedded and added to the index.

  • A (correct)Correct. RAG retrieves the current documents at query time, so updates take effect without retraining and the sources can be cited.
  • B (incorrect)Weekly retraining is slow and expensive, and a fine-tuned model still cannot reliably cite which document an answer came from.
  • C (incorrect)Instructions alone do not give the model knowledge it does not have.
  • D (incorrect)Deterministic decoding makes output repeatable but does not add current facts.

Reference: NVIDIA Blog: What is retrieval-augmented generation? (opens in a new tab)

Familiarity with the fundamentals of machine learning (e.g., feature engineering, model comparison, cross validation).FoundationDomain: Core Machine Learning and AI Knowledge

Question 5. You have a small labelled dataset and want a reliable estimate of how two classifiers will perform on unseen data. What does k-fold cross-validation give you?

  1. Option A: A larger dataset, because each record is duplicated k times
  2. Option B: An average score across k train/validation splits, so every record is used for validation exactly once
  3. Option C: A guarantee that the model will not overfit
  4. Option D: The optimal learning rate for each model
Show answer and explanation

Correct answer: B. An average score across k train/validation splits, so every record is used for validation exactly once

Averaging over k folds gives a more stable performance estimate than one train/test split, which matters most with small datasets. It is a standard way to compare models fairly.

  • A (incorrect)Cross-validation does not create new data; it re-uses the existing data in different splits.
  • B (correct)Correct. The data is split into k folds; each fold serves once as the validation set while the others are used for training.
  • C (incorrect)Cross-validation helps detect overfitting; it does not prevent it.
  • D (incorrect)It can be used inside hyperparameter search, but by itself it only estimates performance.

Reference: NVIDIA: NCA-GENL Generative AI LLMs certification exam (opens in a new tab)

Domain 2

Data Analysis

Compare models using statistical performance metrics, such as loss functions or proportion of explained variance.FoundationDomain: Data Analysis

Question 6. A regression model predicting GPU inference latency has an R² of 0.82 on the test set. What does this mean?

  1. Option A: 82% of predictions are exactly correct
  2. Option B: The model explains about 82% of the variance in latency
  3. Option C: The mean squared error is 0.18
  4. Option D: The model is 82% likely to generalise
Show answer and explanation

Correct answer: B. The model explains about 82% of the variance in latency

R² compares the model's errors with those of simply predicting the mean. A value of 1 is a perfect fit, 0 is no better than the mean, and it can be negative for a model that is worse than the mean.

  • A (incorrect)R² is not an accuracy rate; regression predictions are rarely exactly correct.
  • B (correct)Correct. R² (the coefficient of determination) is the proportion of variance in the target explained by the model.
  • C (incorrect)R² is relative to the variance of the target; it cannot be converted to MSE without knowing that variance.
  • D (incorrect)R² describes fit on the evaluated data, not a probability of generalisation.

Reference: NVIDIA: NCA-GENL Generative AI LLMs certification exam (opens in a new tab)

Identify relationships and trends or any factors that could affect the results of research.IntermediateDomain: Data Analysis

Question 7. You plot training and validation loss per epoch for a fine-tuning run. Training loss keeps falling, but validation loss starts rising after epoch 3. What does this trend indicate?

  1. Option A: The learning rate is too low
  2. Option B: The model is underfitting
  3. Option C: The validation set is too large
  4. Option D: The model has started to overfit the training data
Show answer and explanation

Correct answer: D. The model has started to overfit the training data

Plotting both curves is the quickest way to spot overfitting. Typical responses are early stopping at the best validation checkpoint, more or more varied data, or stronger regularisation.

  • A (incorrect)A too-low learning rate typically makes both losses fall slowly, not diverge.
  • B (incorrect)Underfitting shows as high training loss; here the training loss keeps improving.
  • C (incorrect)A larger validation set gives a more reliable estimate; it does not cause loss to rise.
  • D (correct)Correct. Diverging curves mean the model is memorising training data instead of generalising.

Reference: NVIDIA: NCA-GENL Generative AI LLMs certification exam (opens in a new tab)

Domain 3

Experimentation

Assist in preparing (e.g., scraping, tokenization) large datasets for pretraining, fine-tuning, and RLHF.IntermediateDomain: Experimentation

Question 8. Why do most LLMs use subword tokenisation (for example byte-pair encoding) rather than a fixed vocabulary of whole words?

  1. Option A: It guarantees that every word becomes exactly one token
  2. Option B: It removes the need for an embedding layer
  3. Option C: Rare or unseen words can be represented from smaller known pieces, keeping the vocabulary a manageable size
  4. Option D: It makes token counts identical across all languages
Show answer and explanation

Correct answer: C. Rare or unseen words can be represented from smaller known pieces, keeping the vocabulary a manageable size

Subword tokenisers balance vocabulary size against sequence length. Because token counts drive context limits and cost, the tokeniser matters when you prepare datasets and estimate training or inference budgets.

  • A (incorrect)The opposite: rare words are split into several subword tokens.
  • B (incorrect)Tokens are still mapped to embedding vectors; subword tokenisation only changes what a token is.
  • C (correct)Correct. Subword vocabularies avoid out-of-vocabulary failures while keeping common words as single tokens.
  • D (incorrect)Token counts vary considerably between languages and scripts with the same tokeniser.

Reference: NVIDIA: NCA-GENL Generative AI LLMs certification exam (opens in a new tab)

Awareness of, and/or participation in, data collection from human subjects (e.g., RLHF).IntermediateDomain: Experimentation

Question 9. In reinforcement learning from human feedback (RLHF), what data do human annotators typically provide to train the reward model?

  1. Option A: Rankings or pairwise comparisons of several model responses to the same prompt
  2. Option B: The model's weights adjusted by hand
  3. Option C: Tokenised pretraining text scraped from the web
  4. Option D: Labelled validation loss curves
Show answer and explanation

Correct answer: A. Rankings or pairwise comparisons of several model responses to the same prompt

RLHF trains a reward model on human preferences and then optimises the LLM against it with reinforcement learning, commonly PPO with a KL penalty that keeps it close to the original model. Collecting this data involves human subjects, which raises consent and quality considerations.

  • A (correct)Correct. Relative judgements are more consistent between annotators than absolute scores and are turned into a scalar reward.
  • B (incorrect)Humans never edit weights directly; the reward model and the RL step do the learning.
  • C (incorrect)That is pretraining data, not human preference data.
  • D (incorrect)Loss curves are training diagnostics, not feedback on response quality.

Reference: NVIDIA: NCA-GENL Generative AI LLMs certification exam (opens in a new tab)

Assist in model training and training optimization under the supervision of a senior team member.AdvancedDomain: Experimentation

Question 10. Your team must adapt a large pretrained LLM to a specialist domain on limited GPU memory and wants no extra inference latency after deployment. Which approach fits best?

  1. Option A: Full fine-tuning of all weights with the Adam optimiser
  2. Option B: Training a new model from scratch on the domain data
  3. Option C: Adding adapter layers that run in sequence at inference
  4. Option D: LoRA, which freezes the pretrained weights and trains small low-rank matrices
Show answer and explanation

Correct answer: D. LoRA, which freezes the pretrained weights and trains small low-rank matrices

LoRA is a parameter-efficient fine-tuning (PEFT) method. Its paper reports about 10,000 times fewer trainable parameters and about three times less GPU memory than full fine-tuning of GPT-3 175B, with no additional inference latency.

  • A (incorrect)Updating every weight needs the most GPU memory, because gradients and optimiser states are kept for all parameters.
  • B (incorrect)Pretraining from scratch needs far more data and compute than adaptation.
  • C (incorrect)Classic adapter layers add computation to the forward pass, which adds inference latency.
  • D (correct)Correct. LoRA trains a tiny fraction of the parameters, and the learned matrices can be merged into the weights so inference is no slower.

Reference: arXiv, LoRA: Low-Rank Adaptation of Large Language Models (suggested reading on the NCA-GENL page) (opens in a new tab)

Domain 4

Software Development

Identify system data, hardware, or software components required to meet user needs.FoundationDomain: Software Development

Question 11. A team must serve a PyTorch model, an ONNX model and a TensorRT-LLM engine from one inference service, exposed over HTTP and gRPC. Which NVIDIA component is designed for this?

  1. Option A: RAPIDS cuDF
  2. Option B: Triton Inference Server
  3. Option C: NeMo Guardrails
  4. Option D: spaCy
Show answer and explanation

Correct answer: B. Triton Inference Server

Triton loads models from a model repository and supports backends such as PyTorch, ONNX Runtime, TensorRT, TensorRT-LLM and Python. One server can host many models behind standard HTTP and gRPC APIs.

  • A (incorrect)cuDF is a GPU DataFrame library for data processing, not a model server.
  • B (correct)Correct. Triton serves models from a model repository through multiple framework backends and exposes HTTP and gRPC endpoints.
  • C (incorrect)NeMo Guardrails adds safety and topic controls to LLM applications; it does not host models.
  • D (incorrect)spaCy is a Python NLP library, not an inference server.

Reference: NVIDIA Docs: Triton Inference Server quickstart (opens in a new tab)

Assist in the deployment and evaluations of model scalability, performance, and reliability under the supervision of a senior team member.AdvancedDomain: Software Development

Question 12. A model's weights fit comfortably in GPU memory, but the service runs out of memory when many users send long conversations at the same time. Which part of LLM inference grows with batch size and sequence length and most likely explains this?

  1. Option A: The tokeniser vocabulary
  2. Option B: The model weights
  3. Option C: The key-value (KV) cache
  4. Option D: The softmax output of the final token only
Show answer and explanation

Correct answer: C. The key-value (KV) cache

KV caching avoids recomputing attention for earlier tokens during decoding, but the cache scales with layers, hidden size, batch size and sequence length. Capacity planning must include it, not just the weights.

  • A (incorrect)The vocabulary is fixed in size and does not grow with traffic.
  • B (incorrect)Weights are loaded once and shared across requests; their size does not depend on the number of users.
  • C (correct)Correct. The keys and values of previous tokens are cached for every layer and every sequence, so memory grows with batch size and context length.
  • D (incorrect)A single output distribution per step is small compared with the cached attention state.

Reference: NVIDIA Technical Blog (Mastering LLM techniques: inference optimization) (opens in a new tab)

Identify system data, hardware, or software components required to meet user needs.IntermediateDomain: Software Development

Question 13. The weights of a 7-billion-parameter model need about 14 GB of GPU memory in FP16. Roughly how much memory do the weights need after quantisation to 8-bit integers (ignoring KV cache and runtime overhead)?

  1. Option A: About 28 GB
  2. Option B: About 14 GB
  3. Option C: About 1.75 GB
  4. Option D: About 7 GB
Show answer and explanation

Correct answer: D. About 7 GB

Weight memory is roughly parameters × bytes per parameter: FP16 uses 2 bytes, INT8 uses 1. Quantisation lets larger models fit on the same hardware, usually with a small accuracy trade-off that should be measured.

  • A (incorrect)That would be 4 bytes per parameter (FP32), which is larger, not smaller.
  • B (incorrect)That is the FP16 size; quantisation reduces the bytes per parameter.
  • C (incorrect)That would correspond to 2 bits per parameter, not 8.
  • D (correct)Correct. 8 bits is 1 byte per parameter, so 7 billion parameters need about 7 GB, half the FP16 size.

Reference: NVIDIA Technical Blog (Mastering LLM techniques: inference optimization) (opens in a new tab)

Familiarity with the capabilities of Python natural language packages (spaCy, NumPy, vector databases, etc.).FoundationDomain: Software Development

Question 14. Before indexing documents for a RAG system, you want to extract organisation and place names from the text with a lightweight, traditional NLP pipeline in Python. Which package is designed for this?

  1. Option A: spaCy, using its named entity recognition component
  2. Option B: NumPy
  3. Option C: Matplotlib
  4. Option D: A vector database such as Milvus
Show answer and explanation

Correct answer: A. spaCy, using its named entity recognition component

Traditional NLP tools such as spaCy are still useful in LLM pipelines, for example to extract metadata, clean text or split sentences before chunking and embedding.

  • A (correct)Correct. spaCy pipelines include tokenisation, part-of-speech tagging and named entity recognition.
  • B (incorrect)NumPy provides numerical arrays and maths, not language processing.
  • C (incorrect)Matplotlib creates charts and plots; it does not analyse text.
  • D (incorrect)A vector database stores and searches embeddings; it does not extract entities from text.

Reference: spaCy usage documentation (suggested reading on the NCA-GENL page) (opens in a new tab)

Domain 5

Trustworthy AI

Describe how to use NVIDIA and other technologies to improve AI trustworthiness.IntermediateDomain: Trustworthy AI

Question 15. A customer-service chatbot must refuse off-topic requests and check its responses before they reach the user, without retraining the underlying LLM. Which NVIDIA technology is designed for this?

  1. Option A: TensorRT-LLM
  2. Option B: RAPIDS cuML
  3. Option C: NeMo Guardrails
  4. Option D: Triton Inference Server
Show answer and explanation

Correct answer: C. NeMo Guardrails

Guardrails sit between the user, the application and the LLM, so policies can be changed without retraining the model. They complement other trustworthy AI measures such as bias evaluation, privacy controls and human oversight.

  • A (incorrect)TensorRT-LLM optimises inference performance; it does not enforce conversation policies.
  • B (incorrect)cuML provides GPU-accelerated classical machine learning algorithms, not chatbot safety controls.
  • C (correct)Correct. NeMo Guardrails adds programmable rails, such as input, output, dialog and topical rails, around an LLM application.
  • D (incorrect)Triton serves models efficiently but does not decide which topics a chatbot may discuss.

Reference: NVIDIA Docs: NeMo Guardrails (opens in a new tab)

Exam information last verified: October 27, 2026

Exam details can change. Always confirm them with the official provider before you book.

Want more practice?

15 questions are a sample. The exam is 50–60 items.

The full CertRamp NVIDIA NCA-GENL practice exams have 6 tests that grow with you, from a diagnostic start to an exam-level Test 4 and a challenge set beyond the exam, with an explanation for every option.

Who writes the exams

Written by a practitioner, not a question mill

Joshua Ravnjak, founder of CertRamp Learning

Joshua Ravnjak

Senior Consultant, digital transformation and AI

Founder, CertRamp Learning

921 learners and 61 reviews on Udemy (as of November 3, 2026)

I have more than 12 years of professional experience and work as a senior consultant in an IT management consultancy, focusing on digital transformation, AI and requirements engineering in the insurance industry. I have earned many certifications myself, so I know what exam preparation feels like. That is why every CertRamp practice exam follows the same progression: an easy start to see where you stand, then real exam level, then beyond it. You walk into the exam with a buffer and a good feeling.

Selected certifications

  • IREB® CPRE Foundation Level
  • ITIL® 4 Foundation
  • Professional Scrum Master I (PSM I)
  • Professional Scrum Product Owner I (PSPO I)
  • Certified SAFe® 6 Scrum Master
  • Certified AI-Empowered SAFe® 6 Product Owner/Product Manager

Plus further certifications from Microsoft and Anthropic (Claude).

ITIL® is a registered trademark of PeopleCert. IREB® is a registered trademark of IREB e.V. SAFe® is a registered trademark of Scaled Agile, Inc. PSM and PSPO are trademarks of Scrum.org.

Also in print

Cover: CPRE Foundation Level Prüfungsvorbereitung 2026 by Joshua Ravnjak

CPRE Foundation Level Prüfungsvorbereitung 2026

6 Übungsprüfungen mit 270 Fragen und ausführlichen Erklärungen

Paperback, German, October 2026

My CPRE Foundation Level exam-prep book (in German) follows the same progressive approach.

View on Amazon (opens in a new tab)