NCA-GENL Practice Test 18 — 15 Questions

Question 1

A data scientist is preprocessing a text corpus to train a large language model. They want to convert each word into a dense vector representation that captures semantic relationships before feeding it into the transformer. Which technique should they use?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Word embeddings (e.g., Word2Vec, GloVe). Word embeddings such as Word2Vec or GloVe convert words into dense vectors that encode semantic relationships, which is essential for language models to understand meaning. Tokenization is only a splitting step, while one-hot and TF-IDF yield sparse vectors lacking semantic depth. Thus, embeddings are the correct choice for capturing semantics before transformer processing.

Answer

TF-IDF vectorization

Answer

Tokenization

Answer

One-hot encoding

Question 2

A global e-commerce company is deploying an LLM-based chatbot to handle customer inquiries. To ensure Trustworthy AI, they must implement a mechanism that allows users to understand why the chatbot provided a specific response, especially for decisions like refund approvals. Which approach best addresses this requirement?

Accepted Answer

Integrate a feature that provides natural language explanations of the chatbot's reasoning, citing the relevant policy or data used.. To meet the explainability requirement, the chatbot should provide natural language explanations that cite the policies or data behind its decisions. This allows users to understand the rationale, especially for consequential actions like refund approvals. Logging, smaller models, or feedback buttons do not directly deliver understandable explanations for specific responses.

Answer

Use a smaller, more interpretable model instead of a large LLM.

Answer

Log all chatbot interactions and make the logs available to users upon request.

Answer

Implement a feedback button that lets users rate the helpfulness of each response.

Question 3

An AI platform team is preparing an LLM for a public-facing legal information assistant. During evaluation, they observe that the model gives systematically different quality answers depending on the dialect used in the prompt. Which action most directly addresses this Trustworthy AI concern?

Accepted Answer

Curate and augment training and evaluation data to include balanced dialect representation, then re-measure quality per dialect.. Dialect-dependent quality differences are a fairness and bias issue rooted in data representation. The durable fix is to rebalance training and evaluation data across dialects and then verify with per-dialect metrics. Prompt instructions, decoding parameters, and output normalization do not alter the learned statistical disparities, so they cannot reliably eliminate the gap or demonstrate improvement to reviewers.

Answer

Restrict the assistant to answering only in a single standardized dialect.

Answer

Lower the model temperature to make responses more deterministic across all users.

Answer

Add a system prompt instructing the model to treat all users equally.

Question 4

A machine learning engineer is training a deep neural network and notices that the training loss decreases but the validation loss starts to increase after several epochs. Which two techniques are most appropriate to mitigate this issue? (Choose two.)

Accepted Answer

Use early stopping based on validation loss.. When validation loss increases while training loss decreases, the model is overfitting. L2 regularization penalizes complexity, and early stopping halts training at the optimal point. Both techniques reduce overfitting and improve generalization. Other options either worsen overfitting or are unrelated to the problem.

Answer

Increase the number of training epochs.

Answer

Increase the learning rate.

Answer

Add more layers to the neural network.

Question 5

A financial services company deploys an NVIDIA NIM inference microservice for an LLM that drafts internal investment summaries. The security team wants to ensure that the model does not reveal sensitive account numbers that appear in its training data. Which NVIDIA NeMo Guardrails mechanism should be configured to detect and block such disclosures at runtime?

Accepted Answer

An output rail that applies a custom action to scan the model response for sensitive patterns and block or mask them.. Output rails are the correct guardrail type because they inspect the model's generated response before it is returned to the user. By attaching a custom action that scans for account-number patterns, the system can block or mask the disclosure. Input, dialog, and retrieval rails operate at different stages and cannot reliably prevent sensitive data from appearing in the final output.

Answer

A retrieval rail that filters documents from the vector database before they are passed to the model.

Answer

A dialog rail that uses a canonical form to redirect the conversation when a sensitive pattern is detected.

Answer

An input rail that validates user prompts against a blocklist of forbidden terms.

Question 6

A media company uses an LLM to generate summaries of user-submitted articles. Legal counsel requires that the system detect and refuse requests that attempt to extract verbatim copyrighted passages longer than a defined threshold. Which capability should the team implement?

Accepted Answer

Output filtering that compares generated text against source documents and blocks responses exceeding the verbatim length threshold.. The requirement is a measurable output constraint: no verbatim spans above a defined length. Only output filtering that compares generated text against the source and blocks violations enforces that threshold deterministically. Input keyword blocking, larger context windows, and stylistic fine-tuning may influence behavior but cannot guarantee or verify the specific legal limit, leaving the organization exposed.

Answer

Input filtering that rejects any prompt containing words like summarize or excerpt.

Answer

Fine-tuning the model on public-domain summaries so it learns to paraphrase.

Answer

Increasing the model's context window so it can consider the entire source document when summarizing.

Question 7

A data scientist is preparing a labeled dataset of 50,000 customer support tickets for supervised fine-tuning of an LLM. Each ticket must be assigned exactly one of eight department labels. Which loss function is most appropriate for training this classification head?

Accepted Answer

Categorical cross-entropy loss. Because every support ticket carries exactly one of eight mutually exclusive department labels, the task is single-label multi-class classification. Categorical cross-entropy, paired with a softmax output layer, directly maximizes the probability of the correct department while normalizing across all eight classes, giving the strongest and most stable training signal for this scenario.

Answer

Mean squared error loss

Answer

Contrastive loss

Answer

Binary cross-entropy loss

Question 8

A team is deploying a large language model for real-time text generation. They observe that the model sometimes produces repetitive and dull outputs, especially when generating longer sequences. They want to encourage more diverse and creative text without significantly degrading coherence. Which decoding strategy should they consider?

Accepted Answer

Top-k sampling. Top-k sampling introduces controlled randomness by sampling from the top k tokens, which increases diversity and reduces repetition. It avoids the pitfalls of greedy and beam search, which are deterministic and prone to dull outputs, while preventing the incoherence that can come from sampling the entire distribution.

Answer

Temperature scaling with a low temperature

Answer

Beam search

Answer

Greedy search

Question 9

An AI team is deploying an LLM-based coding assistant. They observe that the model sometimes generates insecure code snippets, such as hardcoded credentials or SQL injection vulnerabilities. To mitigate this without retraining the model, which approach aligns with NVIDIA's Trustworthy AI recommendations?

Accepted Answer

Implement a post-processing output rail that scans generated code for known vulnerability patterns and blocks or flags them.. A post-processing output rail is the most direct mitigation because it inspects the generated code before it reaches the user, using static analysis or pattern matching to catch vulnerabilities. It does not require retraining and can be updated as new vulnerability patterns emerge. The other options either increase risk, require retraining, or do not target the security of the output.

Answer

Restrict the model's context window to limit the amount of code it can generate, reducing the chance of vulnerabilities.

Answer

Fine-tune the model on a dataset of secure code snippets to teach it to avoid vulnerabilities.

Answer

Increase the model's temperature to encourage more diverse code suggestions, reducing the chance of insecure patterns.

Question 10

A healthcare startup is fine-tuning an NVIDIA Llama 2 model on patient records to build a clinical summarization assistant. Before training, the team wants to ensure that individually identifiable information cannot be reconstructed from the model. Which data preparation step best supports this Trustworthy AI goal?

Accepted Answer

Apply differential privacy during fine-tuning by adding calibrated noise to the training process.. Differential privacy is the only option that offers a formal, quantifiable guarantee against re-identification. By injecting calibrated noise during fine-tuning, it limits how much any single patient record can influence the model, making reconstruction attacks provably harder. The other options are either informal heuristics or only remove direct identifiers without addressing model memorization.

Answer

Increase the learning rate to make the model generalize better and forget specific patient details.

Answer

Train the model for fewer epochs to reduce the amount of data it memorizes.

Answer

Use k-anonymity to replace patient names with pseudonyms before fine-tuning.

Question 11

A media company is deploying an LLM that writes first drafts of news briefs. The editorial board wants safeguards that reduce the risk of the model emitting defamatory or unverified claims about named individuals before a human editor reviews the draft. Which two measures best address this risk? (Choose two.)

Accepted Answer

Add an output moderation rail that flags or blocks sentences making unverified assertions about named people before the draft is shown to the editor.. The risk is that unverified or defamatory assertions reach the editor before review. An output moderation rail provides a runtime filter that flags or blocks such sentences, while retrieval-augmented grounding requires person-specific claims to be supported by source documents. Together they reduce fabrication at generation and catch problematic statements before display. Context size, sampling settings, and logging changes do not constrain claim veracity.

Answer

Lower the model's top-p sampling value to make the generated text more deterministic.

Answer

Increase the model's context window so it can ingest the entire archive of past articles for every draft.

Answer

Disable logging of prompts and outputs to protect the privacy of the individuals mentioned in drafts.

Question 12

A retail company's LLM-based product recommendation assistant begins suggesting discontinued items and outdated pricing roughly six weeks after launch, even though the model weights have not changed. The team confirms the training data and prompts are unchanged. Which phenomenon best explains the degraded output quality?

Accepted Answer

Concept drift in the business environment. Because the model, prompts, and training data are static while recommendations degrade over weeks, the change must come from the environment the model describes. Discontinued items and updated prices alter the correct mapping from customer query to product, which is concept drift. The other choices require retraining, fixed preprocessing faults, or time-accumulating numeric errors, none of which fit an unchanged model that initially performed well.

Answer

Quantization error accumulation over time

Answer

Catastrophic forgetting during inference

Answer

Tokenization mismatch in the embedding layer

Question 13

A financial services firm is deploying an NVIDIA NIM microservice for an internal LLM assistant that summarizes confidential client portfolios. The security team wants to enforce that every prompt and completion is screened for PII and prompt-injection attempts before reaching the model. Which two NVIDIA components are purpose-built for this enforcement layer? (Choose two.)

Accepted Answer

NVIDIA Morpheus. NeMo Guardrails provides the programmable input/output rail layer that inspects and blocks prompts and completions, while Morpheus supplies accelerated cybersecurity inspection of the data stream for sensitive information. Together they create defense in depth around the NIM microservice without modifying the model itself. Triton, Nsight Systems, and TensorRT-LLM address serving, profiling, and optimization respectively, none of which enforce content policy or detect PII.

Answer

NVIDIA TensorRT-LLM

Answer

NVIDIA Nsight Systems

Answer

NVIDIA Triton Inference Server

Question 14

A developer is evaluating a fine-tuned LLM with NVIDIA NeMo and observes that evaluation loss keeps decreasing while downstream task accuracy plateaus and then declines. Which action should the developer take to address this?

Accepted Answer

Apply early stopping based on the downstream validation metric and consider regularization such as dropout or weight decay.. A widening gap between decreasing evaluation loss and declining downstream accuracy is the classic signature of overfitting. Early stopping on the task metric preserves the best generalizing checkpoint, and regularization techniques such as dropout or weight decay constrain the model so optimization progress translates into real task gains.

Answer

Extend training for more epochs so the loss can converge further.

Answer

Increase the learning rate to escape the plateau.

Answer

Switch the evaluation metric to training loss for consistency.

Question 15

A healthcare analytics team uses an LLM to summarize patient notes for clinician review. The team observes that summaries for patients from one demographic group systematically omit certain chronic conditions that appear in the source notes. Which action most directly addresses this Trustworthy AI failure?

Accepted Answer

Measure summarization completeness per demographic subgroup and retrain or adjust the pipeline until omission rates are comparable across groups.. A subgroup-specific pattern of omitted chronic conditions is a fairness defect that must be quantified before it can be fixed. Measuring completeness per demographic group establishes whether the disparity is real and whether interventions work. Disclaimers, longer summaries, and larger models are generic changes that do not target the measured gap and cannot demonstrate that equitable performance has been achieved in this clinical setting.

Answer

Switch to a larger foundation model with a longer context window so that the entire patient record fits in a single prompt.

Answer

Add a disclaimer to every generated summary stating that the output may be incomplete and must be verified by a clinician.

Answer

Increase the maximum summary length so that the model has more room to include every condition mentioned in the source note.