CertArea
All NVIDIA exams

Free NCA-GENL practice test: 30 questions with answers

30 original questions for NVIDIA-Certified Associate: Generative AI LLM, drawn from every topic area. Answer each one to see whether you got it right and why. No sign-up.

Free questions
30
Full bank
100 questions
Real exam
60 min
Passing score
Pass/fail

NCA-GENL exam guide: topics, cost and format

All 30 NCA-GENL questions with answers

The same questions as the test above, grouped by exam topic. Open any question to see the answer and explanation.

Prompt engineering and model adaptation (8 questions)

  1. Question 1

    A developer wants an LLM to return a JSON object with specific keys for every request. Which prompting approach best achieves consistent output?

    1. AAsk for 'any format' and parse whatever comes back with a regex fallback
    2. BSpecify the exact output schema in the prompt and provide an example of valid output
    3. CRemove all instructions from the prompt so the model can respond freely without format constraints
    4. DUse the highest temperature so the model is creative and varied across runs
    Show answer and explanation
    Correct answer: B

    An explicit schema and an example guide consistent structured output. Allowing any format, high temperature and removing instructions reduce consistency.

  2. Question 2

    A team wants to adapt a large LLM to a specialized domain without retraining all weights. Which technique adds small trainable low-rank matrices to the frozen model?

    1. AFull retraining from random initialization
    2. BLoRA
    3. CTokenizer replacement
    4. DGreedy decoding
    Show answer and explanation
    Correct answer: B

    LoRA trains small low-rank adapters while keeping most weights frozen, which lowers cost. Random initialization retrains everything, decoding is inference, and tokenizer replacement changes the input representation.

  3. Question 3

    A chatbot must answer questions about a company's product manuals with citations, and the manuals change monthly. Which approach is most suitable?

    1. ARetrieval-augmented generation over the current manuals
    2. BAsking the model to rely on memory only
    3. CFine-tuning the model on the manuals once and never updating it
    4. DCopying the manuals into the model's tokenizer
    Show answer and explanation
    Correct answer: A

    RAG retrieves current manual passages and can cite sources, which suits monthly changes. A one-time fine-tune becomes stale, memory-only answers can be wrong, and tokenizers do not store knowledge.

  4. Question 4

    A prompt asks the model to reason step by step before giving a final answer for a multi-step math problem. Which technique is this?

    1. AStandardized feature scaling
    2. BChain-of-thought prompting
    3. CEarly stopping on validation loss
    4. DZero-shot entity classification
    Show answer and explanation
    Correct answer: B

    Chain-of-thought prompting elicits intermediate reasoning steps, which often improves multi-step accuracy. The other techniques are not prompting methods for reasoning.

  5. Question 5

    A developer wants the model to respond only with information from a provided document, and to say when the answer is not in the document. Which instruction type should be included?

    1. AAn instruction to ignore the document and answer from general training knowledge instead
    2. BA grounding instruction that limits answers to the provided context and defines the fallback behaviour
    3. CA request for creative answers that extrapolate beyond the document with speculative detail
    4. DNo instruction at all, so the model relies on pretrained parametric memory and its default style with no restriction on sources
    Show answer and explanation
    Correct answer: B

    Grounding instructions restrict answers to the provided context and define fallback behaviour. Creative or ignore instructions, and no instruction, do not enforce grounding.

  6. Question 6

    An application must reduce the risk that a user's input manipulates the model's instructions. Which measure is most relevant?

    1. AAllow user input to override all instructions and system role boundaries
    2. BDisable input length limits on the API gateway so that long prompts skip checks
    3. CConcatenate user input directly with the system prompt without separation
    4. DSeparate system instructions from user content and validate or constrain user input
    Show answer and explanation
    Correct answer: D

    Separating instructions from user content and constraining input reduces prompt injection risk. Mixing them, allowing overrides and removing limits increase the risk.

  7. Question 7

    A team needs an LLM to classify support tickets into five categories. They provide three labelled examples in the prompt for each category. Which approach is this?

    1. AReinforcement learning from rewards
    2. BUnsupervised clustering
    3. CFew-shot prompting
    4. DTokenizer training
    Show answer and explanation
    Correct answer: C

    Providing labelled examples in the prompt is few-shot prompting. Clustering is unsupervised, reinforcement learning uses rewards, and tokenizer training is a separate step.

  8. Question 8

    A company wants a smaller model to imitate the outputs of a larger model to reduce inference cost. Which technique is this?

    1. AHyperparameter search
    2. BEarly stopping
    3. CKnowledge distillation
    4. DData augmentation with flips
    Show answer and explanation
    Correct answer: C

    Distillation trains a smaller student model to mimic a larger teacher, reducing cost. The other techniques address training search, image augmentation and stopping rules.

Core machine learning and LLM concepts (8 questions)

  1. Question 9

    A language model is trained to predict the next token in a sequence. What is the main training objective used for most decoder-only LLMs?

    1. APredicting the image class of each token
    2. BMaximizing the length of the context window
    3. CNext-token prediction with cross-entropy loss
    4. DMinimizing the number of tokens in the output
    Show answer and explanation
    Correct answer: C

    Decoder-only LLMs are trained to predict the next token, usually with cross-entropy loss over the vocabulary. The other objectives do not describe standard LLM pretraining.

  2. Question 10

    In a transformer, which mechanism allows each token to weigh the relevance of every other token in the sequence when building its representation?

    1. ASelf-attention
    2. BMax pooling over channels
    3. CBatch normalization only
    4. DDropout on the output layer
    Show answer and explanation
    Correct answer: A

    Self-attention computes weighted relationships between all tokens, which is the core of the transformer. Pooling, normalization and dropout do not compute token-to-token relevance.

  3. Question 11

    Why do transformer models add positional encodings to token embeddings?

    1. ABecause attention alone does not know the order of tokens
    2. BTo replace the need for tokenization by encoding raw byte sequences
    3. CTo reduce the model's memory footprint during inference and training
    4. DTo increase the vocabulary size with one extra token per position index
    Show answer and explanation
    Correct answer: A

    Self-attention is permutation-invariant, so positional information must be injected. Positional encodings do not change vocabulary size, reduce memory, or replace tokenization.

  4. Question 12

    A tokenizer splits the word 'unhappiness' into several subword pieces. What is the main benefit of subword tokenization for LLMs?

    1. AIt handles rare and unseen words by composing known pieces, with a manageable vocabulary size
    2. BIt guarantees that the model will not make spelling mistakes, since every word is spelled out letter by letter
    3. CIt removes the need for embeddings by feeding one-hot IDs straight into attention
    4. DIt makes every word a single token regardless of length by giving each whole word its own embedding table entry
    Show answer and explanation
    Correct answer: A

    Subword tokenization balances vocabulary size and coverage, so unseen words can be built from known pieces. Whole-word tokens, removing embeddings and spelling guarantees are incorrect.

  5. Question 13

    An LLM generates text one token at a time. Which decoding strategy selects the single most probable token at each step?

    1. ANucleus sampling with p equal to 1
    2. BRandom sampling from a uniform distribution
    3. CBeam sampling with a width of 1000
    4. DGreedy decoding
    Show answer and explanation
    Correct answer: D

    Greedy decoding picks the highest-probability token at each step. The other strategies either sample or search multiple candidates.

  6. Question 14

    Which parameter increases the randomness of sampled text from an LLM?

    1. AA smaller context window
    2. BA shorter maximum output length
    3. CA lower top-k value
    4. DA higher temperature
    Show answer and explanation
    Correct answer: D

    Higher temperature flattens the probability distribution and increases diversity. Lower top-k restricts choices, and length or context size do not directly control randomness.

  7. Question 15

    A model's context window is 8,000 tokens. A user pastes a document of 20,000 tokens. What is the most likely consequence?

    1. APart of the document is truncated or must be chunked, because it exceeds what the model can attend to at once
    2. BThe model automatically increases its context window by allocating extra positional embeddings and extending the attention mask
    3. CThe model reads the whole document with no loss of information because its context window is effectively unlimited at inference
    4. DThe model converts the text into images by rendering each page as a pixel grid for a vision encoder
    Show answer and explanation
    Correct answer: A

    Inputs beyond the context window must be truncated, summarized or chunked. The model cannot attend beyond its window, and it does not change its context size on its own.

  8. Question 16

    A team compares a dense model and a mixture-of-experts model. What is the main idea behind mixture-of-experts?

    1. AExperts are separate databases queried by the user through a retrieval API, with each expert holding a fixed partition
    2. BMixture-of-experts removes the need for training by reusing frozen weights from a checkpoint
    3. COnly a subset of expert sub-networks is activated for each token, which reduces compute per token
    4. DAll experts are always activated for every token, so the gating network's routing output is computed but ignored
    Show answer and explanation
    Correct answer: C

    Mixture-of-experts routes each token to a few experts, so only part of the parameters are used per token. Activating all experts for every token, separate databases and no training are incorrect.

Data preparation and evaluation (7 questions)

  1. Question 17

    A dataset for training an LLM contains many near-duplicate documents. What problem can this create?

    1. AFaster convergence in early epochs, with no effect on evaluation scores or loss curves.
    2. BOverrepresentation of some content and inflated evaluation results if duplicates leak across splits
    3. CAutomatic removal of bias, since repeated text is averaged out during gradient steps.
    4. DSmaller vocabulary size, because the tokenizer merges frequent duplicate strings into shorter subword units.
    Show answer and explanation
    Correct answer: B

    Duplicates overweight some content and can leak into evaluation sets, inflating scores. Faster convergence and bias removal are not outcomes, and vocabulary size is set by the tokenizer.

  2. Question 18

    A team wants to measure whether an LLM's summaries are factually consistent with the source documents. Which evaluation approach is most appropriate?

    1. ACompare the summary's font size, line spacing, and heading styles against the source layout.
    2. BCount the number of paragraphs and bullet points in the summary and the source documents.
    3. CCheck claims against the source documents with human or automated fact checks on a sample
    4. DMeasure the number of words in the summary and compare it against a fixed word limit.
    Show answer and explanation
    Correct answer: C

    Factual consistency requires comparing claims to the sources, through human or automated checks. Word counts, font size and paragraph counts do not measure faithfulness.

  3. Question 19

    A team evaluates an LLM on a benchmark that the model may have seen during training. Which concern should they raise?

    1. ABenchmark contamination, which can inflate the measured performance
    2. BLatency of the benchmark server during long evaluation runs.
    3. CBenchmark formatting, like prompt templates and parsing rules, raises the score.
    4. DThe number of colours and fonts in the final report template.
    Show answer and explanation
    Correct answer: A

    Contamination means test items appeared in training, inflating scores. Formatting, server latency and report colours do not cause inflated accuracy.

  4. Question 20

    A team splits a dataset into training, validation, and test sets. What is the purpose of the validation set?

    1. ATo tune hyperparameters and select models without touching the test set
    2. BTo store the model's learned weights and optimizer state after each epoch as a checkpoint.
    3. CTo train the model on additional examples to expand the effective size of the training set.
    4. DTo replace the test set at the end with a freshly drawn sample from production logs.
    Show answer and explanation
    Correct answer: A

    The validation set guides tuning and model selection, keeping the test set for final evaluation. It is not extra training data or a replacement for the test set, and it does not store weights.

  5. Question 21

    A team wants to know whether a new model version is better than the current one for customer support. Which comparison design is most reliable?

    1. AComparing the team's recollections of recent customer conversations from memory, without scoring.
    2. BRunning the new model alone on a few hand-picked prompts and reporting its score.
    3. CAn A/B evaluation on the same representative set of support tasks, scored with agreed criteria
    4. DChoosing the newer version because its release date and version number are more recent.
    Show answer and explanation
    Correct answer: C

    A controlled A/B comparison on the same representative tasks with agreed criteria gives a fair decision. Memory, single-model reports and recency alone do not support a fair comparison.

  6. Question 22

    Which TWO are common sources of error in LLM evaluation? (Choose two.)

    1. AA clearly defined task with a written scoring rubric
    2. BOverlap between training data and the test set
    3. CJudging subjective quality with an unclear rubric
    4. DRecording every result with the model version
    5. EA large and diverse test set
    Show answer and explanation
    Correct answer: B, C

    Train-test overlap and unclear rubrics inflate or distort results. Clear rubrics, diverse tests and versioned records improve evaluation.

  7. Question 23

    A dataset of support chats includes customer names and account numbers. Before using it for fine-tuning, what is the most appropriate step?

    1. AKeep customer names and account numbers because they help the model learn
    2. BPublish the raw chats to improve model transparency
    3. CDetect and remove or mask personal data according to policy
    4. DTrain on the raw data because it is internal and already secured
    Show answer and explanation
    Correct answer: C

    Personal data should be detected and masked or removed under policy before training. Internal data still has privacy obligations, and publishing or keeping personal details is not appropriate.

Deployment and trustworthy AI (7 questions)

  1. Question 24

    A company wants to track how an LLM application performs after launch, including quality, latency, and cost. Which practice is most appropriate?

    1. AMonitoring dashboards with defined metrics, sampling of outputs for review, and alerts
    2. BTurning off most metrics to reduce logging overhead on production servers
    3. CRunning a manual review of the application once a year during the annual planning cycle and budgeting
    4. DRelying on user complaints and support tickets to surface problems quickly
    Show answer and explanation
    Correct answer: A

    Defined metrics, output sampling and alerts provide ongoing visibility. Annual checks, complaint-only monitoring and disabled metrics hide problems.

  2. Question 25

    A team needs to decide whether an LLM can be used for a regulated financial advice tool. Which step is most important before launch?

    1. AA risk assessment and human oversight plan that defines what the model may and may not decide
    2. BGive customers the model weights without access controls, usage limits, or audit logging
    3. CDrop human review, because the model's benchmark accuracy is high on test sets
    4. DLaunch immediately to the public and fix safety issues in later patches
    Show answer and explanation
    Correct answer: A

    Risk assessments and human oversight define safe use in regulated settings. Immediate launch, removing review and releasing weights without controls create serious risks.

  3. Question 26

    A company wants to deploy an LLM service with a predictable cost per request and the ability to scale during peaks. Which combination best supports this?

    1. ARunning the model on every employee's laptop and personal workstation for demos
    2. BA single fixed GPU server with a static batch size and no request queue
    3. CAutoscaling of GPU inference replicas with a request queue and a cost per request dashboard
    4. DDisabling monitoring and cost logging for budget reasons during peak traffic
    Show answer and explanation
    Correct answer: C

    Autoscaling with queues and cost dashboards gives controlled scaling and visibility. A single fixed server cannot handle peaks, laptops are unmanaged, and disabling monitoring hides cost.

  4. Question 27

    A team wants to reduce the risk of a deployed LLM generating unsafe output. Which layer of protection is most effective when placed at the output stage?

    1. AA higher learning rate applied at inference to sharpen the next-token distribution
    2. BA content safety classifier that checks responses before they are returned
    3. CA longer prompt that repeats the question verbatim twice
    4. DA larger vocabulary size in the tokenizer, which increases the embedding table
    Show answer and explanation
    Correct answer: B

    A safety classifier at output can block or flag unsafe responses before they reach users. Learning rates, longer prompts and vocabulary size do not provide output safety checks.

  5. Question 28

    Which TWO techniques can lower the cost of serving an LLM? (Choose two.)

    1. AServing every request on the largest model available
    2. BQuantizing weights to lower precision
    3. CUsing KV caching for repeated context
    4. DDisabling request batching
    5. EKeeping replicas running at peak capacity around the clock
    Show answer and explanation
    Correct answer: B, C

    Quantization and KV caching reduce memory and compute per request. Routing everything to the largest model, always-on peak capacity and disabling batching all raise cost.

  6. Question 29

    Which THREE are components of a trustworthy AI program for LLM applications? (Choose three.)

    1. ARemoving all human oversight to increase speed
    2. BHiding model limitations from users
    3. CDocumented intended use and known limitations
    4. DRed teaming to find harmful outputs before launch
    5. EOngoing monitoring with user feedback loops
    Show answer and explanation
    Correct answer: C, D, E

    Documented use, red teaming and monitoring with feedback build trust. Removing oversight and hiding limitations undermine it.

  7. Question 30

    A company needs to deploy an LLM on-premises for data privacy reasons, while still updating the model when a new version is approved. Which practice is most appropriate?

    1. AUse a controlled model registry and a documented, tested update process on the on-premises infrastructure
    2. BDownload new model checkpoints from random websites and community mirrors without checksum verification into the serving cluster
    3. CKeep the deployed model version frozen after go-live and leave it unchanged for the full product lifetime so that quality fixes are deferred
    4. DUpdate the model directly on the production server without testing, rollback plans, or a change window during peak business hours
    Show answer and explanation
    Correct answer: A

    A controlled registry and tested update process keep on-premises deployments safe and current. Random downloads, untested production updates and never updating all create risk.

Want more NCA-GENL questions?

The full bank has 100 questions with answers and explanations, as a PDF, an online practice exam, or both.

Full question bank · 100 questions
$75$104

Both formats together. The best value.

Instant access after payment, with 90 days of updates. Checkout is on our store, AIExamCerts.