CertifyAIxAce every AI certification.

NCA-GENL sample questions with answers

10 free practice questions for the NVIDIA-Certified Associate: Generative AI LLMs exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Core Machine Learning and AI Knowledge

An e-commerce team plans post-training quantization of both the weights and the activations of its product-description model. A senior engineer specifies static quantization, meaning activation scale factors are computed once and then reused for every future inference request. What must the team supply that weight-only quantization would not have needed?

  1. A.

    A labelled dataset with reference outputs so the quantized model can be retrained to match them

  2. B.

    A representative sample of realistic input data to run through the model and calibrate activation scales

  3. C.

    A second copy of the model in 32-bit precision to serve requests whose activations fall out of range

  4. D.

    Nothing extra, because activation scale factors can always be computed on the fly for each request

Show answer

Answer: B

Activations depend on inputs, so static activation quantization needs calibration data; weights alone do not.

  • A. Calibration uses unlabelled inputs and leaves weights unchanged; retraining against labels is a different method.
  • B. Static activation quantization derives fixed scales from calibration runs over representative inputs.
  • C. No full-precision fallback model is part of post-training quantization; out-of-range values are clipped.
  • D. On-the-fly scale computation describes dynamic quantization, but the stem specifies scales fixed in advance.
Question 2Core Machine Learning and AI Knowledge

A city transport authority clusters GPS stop locations of delivery vans with k-means. Plotting the result, an analyst sees that long, curved groups of stops along ring roads are chopped into several pieces, and parts of different roads are merged into one cluster. Which property of k-means explains this?

  1. A.

    K-means labels every curved group as noise, which leaves the remaining points poorly clustered

  2. B.

    K-means always discovers the true number of clusters, so the roads must genuinely overlap

  3. C.

    K-means assigns points to the nearest centroid, so it works best on compact, roughly spherical clusters

  4. D.

    K-means cannot process geographic coordinates because it only accepts categorical input features

Show answer

Answer: C

Nearest-centroid assignment favours compact, spherical clusters and splits elongated or curved shapes.

  • A. K-means assigns every point to a centroid; labelling points as noise is a feature of density-based methods.
  • B. K-means does not discover the number of clusters; K is supplied in advance by the user.
  • C. Distance-to-centroid assignment favours compact spherical groups, so curved clusters are split and merged.
  • D. Latitude and longitude are numeric features, which k-means handles; input type is not the cause.
Question 3Core Machine Learning and AI Knowledge

A grocery chain builds a model to forecast next week's demand from three years of daily sales, which show a strong upward trend and seasonality. Random 5-fold cross-validation reports excellent accuracy, but forecasts made for the following weeks after deployment are much worse. Which change to the validation approach addresses the likely cause?

  1. A.

    Use stratified folds so that each fold contains the same proportion of high-sales and low-sales days

  2. B.

    Increase the number of random folds from five to ten so that each validation fold is smaller and more precise

  3. C.

    Shuffle the rows more thoroughly before the random split so that the folds are more representative

  4. D.

    Validate with time-ordered splits, training on earlier periods and validating on later ones, as in deployment

Show answer

Answer: D

Random folds let the model train on days after the ones it is validated on; time-aware validation mirrors real forecasting.

  • A. Stratifying by sales level balances folds but still mixes future and past across them.
  • B. More random folds still put later days into training for earlier validation days.
  • C. Shuffling is what destroys the time order; more shuffling cannot fix it.
  • D. Time-ordered validation prevents training on the future and matches how forecasts are actually made.
Question 4Core Machine Learning and AI Knowledge

In a hardware vendor's RAG chatbot, a user asks 'What GPUs does the X500 server support?' and gets a good answer, then asks 'And what about its power requirements?' The second answer is vague and cites unrelated documents. Conversation history is passed to the LLM. What change best fixes retrieval for follow-up questions?

  1. A.

    Raise the generator's maximum output length so that it can include every possible power specification

  2. B.

    Clear the conversation history after every turn so that earlier questions cannot confuse retrieval

  3. C.

    Pass more previous turns to the LLM, because the generator alone can recover from irrelevant retrieved chunks

  4. D.

    Rewrite each follow-up into a standalone query using the conversation history before sending it to retrieval

Show answer

Answer: D

Follow-ups like 'its power requirements' are ambiguous to the retriever; rewriting them into standalone queries restores retrieval accuracy.

  • A. Output length does not affect which documents are retrieved for the follow-up question.
  • B. Clearing history removes the very context needed to interpret 'its' in the follow-up.
  • C. The generator can only use retrieved chunks; more history for it cannot fix wrong retrieval.
  • D. A standalone rewritten query names the X500 explicitly, so retrieval finds its power documentation.
Question 5Core Machine Learning and AI Knowledge

A customer-feedback team has 5,000 labelled comments tagged by topic and wants a lightweight topic classifier. They already run an embedding model for search. A colleague suggests reusing those embeddings. What is the role of embeddings in this design?

  1. A.

    They can only be used for retrieval, so the team needs a separate model to create any features

  2. B.

    They are the classifier's output, giving the probability of each topic for every comment directly

  3. C.

    They serve as numeric input features for a classifier trained on the labelled comments

  4. D.

    They replace the labels, because embedding similarity assigns topics without any training data

Show answer

Answer: C

Embeddings turn text into numeric features, and a simple classifier can be trained on them with the labels.

  • A. Embedding documentation lists classification and clustering uses as well as retrieval.
  • B. An embedding is a representation of the text, not a probability distribution over the topics.
  • C. Embeddings supply numeric features, and the labelled comments train a classifier on top of them.
  • D. Similarity alone does not know the team's topic names, so the labels are still needed to train the mapping.
Question 6Core Machine Learning and AI Knowledge

A consultancy's research digest distinguishes three kinds of AI scaling. One entry describes a newer approach in which a model spends more computation at inference time, working through several steps or passes before answering, to improve accuracy on complex problems. Which scaling law does that entry describe?

  1. A.

    Test-time scaling, which applies more compute during inference to improve reasoning accuracy

  2. B.

    Pretraining scaling, which increases dataset size, parameter count and compute during pre-training

  3. C.

    Post-training scaling, which specialises a pre-trained model for a domain through further training

  4. D.

    Data-parallel scaling, which splits one training batch across many GPUs to shorten training time

Show answer

Answer: A

Spending more compute at inference to reason through problems is test-time scaling.

  • A. More compute at inference, used to reason through steps before answering, is how the explainer defines test-time scaling.
  • B. Pretraining scaling concerns data, parameters and compute during pre-training, not extra inference-time work.
  • C. Post-training scaling specialises a trained model through further training rather than extra inference compute.
  • D. Data parallelism is a way to distribute training across GPUs, not a law about inference-time reasoning.
Question 7Core Machine Learning and AI Knowledge

A fishing-equipment retailer's search uses static word vectors, where each word has one fixed vector. Queries about 'bass' return guitar-amplifier listings alongside fishing lures, because the same word has two meanings. Which embedding approach addresses this ambiguity?

  1. A.

    Removing the word 'bass' from all queries, so that other words determine the search results instead

  2. B.

    Static word vectors with a larger dimension, which store both meanings of the word in one fixed vector

  3. C.

    Contextual embeddings from a transformer model, which represent a word differently depending on its context

  4. D.

    Sorting the vocabulary alphabetically before training, so that different senses receive separate indexes

Show answer

Answer: C

Transformer-based contextual embeddings give 'bass' different representations in fishing and music contexts.

  • A. Removing the key word from queries discards the user's main intent and degrades relevance.
  • B. A static vector is still one fixed vector per word, however large, so both senses stay merged.
  • C. Contextual models compute a word's representation from its neighbours, separating the fishing and music senses.
  • D. Vocabulary ordering has no effect on how word meanings are represented in the vectors.
Question 8Core Machine Learning and AI Knowledge

A legal-AI team notices that doubling the length of the documents it feeds into a BERT-style encoder much more than doubles the processing time and memory per document. A reading group is asked to connect this to the architecture described in the Transformer literature. Which explanation is correct?

  1. A.

    The tokenizer runs once per token pair, so tokenization is what grows quadratically with length

  2. B.

    Self-attention compares every token with every other token, so time and memory grow quadratically with sequence length

  3. C.

    Positional encodings are recomputed for every layer, so cost grows exponentially with document length

  4. D.

    Longer documents need more output classes, so the classification head dominates the processing cost

Show answer

Answer: B

Full self-attention scales quadratically with sequence length, which is why long inputs are disproportionately expensive.

  • A. Tokenization scales roughly linearly with text length and is not performed once per pair of tokens.
  • B. Pairwise attention scores grow with the square of the sequence length, which matches the disproportionate slowdown.
  • C. Positional encodings add one fixed-size vector per position and do not create exponential growth in cost.
  • D. The classification head depends on the number of labels, not on how long the input document is.
Question 9Core Machine Learning and AI Knowledge

A copywriting tool is configured with temperature 0.9 so that it produces varied taglines, yet every request with the same prompt returns exactly the same tagline. A review of the generation settings shows top-k is set to 1. What explains the behaviour?

  1. A.

    Top-k of 1 disables the stop sequence, so the model always generates the same length of text

  2. B.

    The prompt is cached, so the model is never called again after the first request

  3. C.

    Top-k of 1 always selects the single most probable token, so temperature has no room to vary the choice

  4. D.

    Temperature above 0.8 is ignored by most models, so the effective setting is zero

Show answer

Answer: C

With top-k set to 1, only the most probable token is ever a candidate, making output deterministic regardless of temperature.

  • A. Top-k does not affect stop sequences, and the symptom is identical text, not only identical length.
  • B. A cache is speculative here, while the settings already show top-k of 1 forcing greedy decoding.
  • C. With only one candidate token per step, reshaping probabilities with temperature cannot change the selection.
  • D. No rule makes models ignore temperatures above 0.8; the setting is applied as configured.
Question 10Core Machine Learning and AI Knowledge

A webmail provider wants a fast, simple baseline to classify incoming emails as spam or not spam from their text. It has a large labelled corpus and needs something that trains quickly and scales well before trying deep learning. Which traditional analysis fits best?

  1. A.

    Cluster the emails with k-means into two groups and call the larger cluster the non-spam class

  2. B.

    Fit a linear regression to the raw email text and treat predictions above 0.5 as spam

  3. C.

    Vectorise the emails with term counts or TF-IDF and train a multinomial naive Bayes classifier

  4. D.

    Compute the average word length of each email and set a threshold that separates spam

Show answer

Answer: C

Multinomial naive Bayes on count or TF-IDF features is the classic, fast text-classification baseline.

  • A. Clustering ignores the available labels, and no rule makes the larger cluster the legitimate mail.
  • B. Linear regression needs numeric features and predicts values rather than calibrated class probabilities.
  • C. Count or TF-IDF features with multinomial naive Bayes is the documented fast, scalable text-classification baseline.
  • D. A single average-word-length feature discards the vocabulary information that distinguishes spam.

Keep going with 559 more NCA-GENL questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all NCA-GENL papers
NCA-GENL sample questions with answers (10 free) · CertifyAIx