Free CertNexus® GenAIBIZ™ Practice Questions

Sample Practice Questions

Try these 10 sample questions from the CertNexus® GenAIBIZ™ practice bank. Each includes the answer and explanation. Continue with the interactive practice above to review the full question set.

1. 1A retailer trains a churn predictor using only historical customer records paired with known churn labels. Which learning approach describes training on this labeled dataset?

   1. A.Unsupervised learning
   2. B.Supervised learningCorrect answer: B
   3. C.Reinforcement learning
   4. D.Semi-supervised learning

   Explanation: NIST states that in supervised learning, "labeled training data is given as input to a training algorithm". The retailer uses labeled examples throughout; unsupervised learning uses unlabeled data, semi-supervised learning combines labeled and predominantly unlabeled samples, and reinforcement learning uses interaction and rewards.

   Reference: Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations — Section 2.1.1, Stages of Learning, p. 5

2. 2A company's question-answering assistant searches an external policy collection for relevant passages and supplies them to a language model with the user's question. Which approach does this system use?

   1. A.Retrieval-augmented generationCorrect answer: A
   2. B.Supervised fine-tuning
   3. C.Self-supervised pre-training
   4. D.Reinforcement learning from feedback

   Explanation: NIST explains that a retrieval-augmented system "identifies relevant information within the knowledge base and provides it to the GenAI model in context". The assistant retrieves policy passages for use during response generation.

   Reference: Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations — Glossary, retrieval-augmented generation, p. 112

3. 3A generated report confidently cites a study that does not exist. Which generative AI shortcoming is illustrated by the invented citation?

   1. A.ConfabulationCorrect answer: A
   2. B.Data drift
   3. C.Overfitting
   4. D.Distribution shift

   Explanation: NIST says that "GAI outputs may also include confabulated logic or citations". An invented study presented as a real citation is a confabulation; the other terms describe changes in data distributions or a model fitting its training data too closely.

   Reference: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — Section 2.2, Confabulation, p. 6

4. 4An organization uses the NIST AI Risk Management Framework while planning an AI project. Under which framework function should it define the business value and context of the proposed use?

   1. A.Govern
   2. B.MapCorrect answer: B
   3. C.Measure
   4. D.Manage

   Explanation: NIST's MAP 1.4 states that "The business value or context of business use has been clearly defined". The proposed-use context therefore belongs to the Map function.

   Reference: Artificial Intelligence Risk Management Framework (AI RMF 1.0) — Section 5.2, MAP 1.4

5. 5A business wants a model to estimate next month's revenue as a continuous monetary value. Which type of predictive task is being specified?

   1. A.Classification
   2. B.Clustering
   3. C.RegressionCorrect answer: C
   4. D.Dimensionality reduction

   Explanation: NIST distinguishes "continuous random variables as in a regression algorithm" from binary or multi-class classification. A continuous revenue estimate is a regression response; clustering groups observations, while dimensionality reduction reduces features.

   Reference: A Technical Framework to Map Functional Recovery Performance Objectives to Prescriptive Seismic Design Provisions for Buildings — Section 3.5.3, Training, p. 26

6. 6A manager asks a text model to summarize a report and includes exactly one completed example of the desired summary in the prompt before the new report. What is the most specific name for this prompting setup?

   1. A.One-shot promptingCorrect answer: A
   2. B.Zero-shot prompting
   3. C.Three-shot prompting
   4. D.Ten-shot prompting

   Explanation: The NIST-hosted research describes a summarization prompt containing "the one-shot example, which includes the context documents ([Document 1]), the question, the evidence paragraph" before the target example. The manager similarly supplies one completed demonstration; the other terms specify different numbers of demonstrations.

   Reference: Using Neural Reranking and GPT-3 for Social Media Disaster Content Summarization — Aggregation, prompt components

7. 7A company repeatedly uses a generative model to process customer requests. Which phase accounts for the computing resources used each time the deployed model produces a response?

   1. A.InferenceCorrect answer: A
   2. B.Pre-training
   3. C.Fine-tuning
   4. D.Model evaluation

   Explanation: NIST distinguishes "operating (running inference on) GAI systems" from training and maintaining them. Producing responses for deployed requests is inference and consumes resources.

   Reference: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — Section 2.5, Environmental Impacts, p. 8

8. 8An AI-assisted screening project flags customer messages for review. The project owner wants the proportion of flagged messages that truly needed review. Which metric measures that proportion?

   1. A.PrecisionCorrect answer: A
   2. B.Recall
   3. C.Overall accuracy
   4. D.F1 score

   Explanation: NIST says "precision quantifies the ratio of predicted positives that were correct". The flagged messages are predicted positives, and the metric asks what fraction were actually correct.

   Reference: A Technical Framework to Map Functional Recovery Performance Objectives to Prescriptive Seismic Design Provisions for Buildings — Section 3.5.1, Model Performance Metrics, p. 24

9. 9After selecting and training a prediction model, a team wants a final estimate of performance on examples held aside throughout model fitting and selection. Which dataset should it use?

   1. A.The training dataset
   2. B.The validation dataset
   3. C.The final test datasetCorrect answer: C
   4. D.The calibration dataset

   Explanation: NIST says "the remaining testing data partition is used to quantify the decision tree’s performance" and describes testing on separate data to assess generality. A final test set is held apart for this assessment; training, validation and calibration sets serve model fitting, selection, or calibration.

   Reference: A Technical Framework to Map Functional Recovery Performance Objectives to Prescriptive Seismic Design Provisions for Buildings — Section 3.5.4, Testing, p. 27

10. 10A team wants to submit a long report and its instructions together to a language model in one request. Which model specification should it check to determine how much context can be included at once?

    1. A.The context-window sizeCorrect answer: A
    2. B.The training epoch count
    3. C.The model's learning rate
    4. D.The training batch size

    Explanation: NIST explains that the context window provides information for processing a prompt and "sets limits to the size of the user prompt". The team must budget the instructions and report against that limit.

    Reference: Robust AI Security and Alignment: A Sisyphean Endeavor? — Section 3.2, Considerations for real AI Systems with finite context windows, p. 8