Free NVIDIA Generative AI LLM Associate Sample Questions

Correct answer: D — Convert the embeddings back to text for review

• A. Dimensionality reduction like t-SNE is commonly applied to embeddings for visualization in 2D or 3D space, making it a valid use case.
• B. Clustering algorithms like k-means are frequently used on embeddings to group similar texts or assess semantic similarity, making it a typical use case.
• C. Training a classifier on embeddings is a standard practice for supervised learning tasks, as embeddings capture semantic meaning.
• D. Converting embeddings back to text is not a typical use case because embeddings are numerical vectors that encode compressed semantic information and cannot be losslessly reversed to the original text.

Correct answers: B, D — To encode both the user query and the document chunks; To compute similarity scores between query and documents

• A. Incorrect. Generating the final response text is the role of the language model after relevant context has been retrieved. Text embeddings are used for retrieval and semantic representation, not for generation.
• B. Correct. Text embeddings encode both the user query and document chunks into numerical vectors, enabling efficient semantic comparison and retrieval of relevant documents.
• C. Incorrect. Text embeddings are vector representations for semantic search and retrieval; they do not compress or reduce the size of the underlying language model.
• D. Correct. Text embeddings enable the computation of similarity scores (e.g., cosine similarity) between query and document embeddings to identify the most relevant documents for retrieval.
• E. Incorrect. Text embeddings are not used for fine-tuning; fine-tuning is a separate training process that occurs before deployment, not during inference.

Correct answer: B — Tagger

• A. Incorrect. The tokenizer splits text into tokens such as words or punctuation, but does not assign part-of-speech tags.
• B. Correct. The tagger assigns part-of-speech tags (e.g., noun, verb, adjective) to tokens. It is a standard spaCy pipeline component.
• C. Incorrect. The parser analyzes syntactic structure (dependency parsing). It is a different component from the tagger.
• D. Incorrect. NER recognizes named entities (e.g., persons, organizations). This task is distinct from part-of-speech tagging.

Correct answer: D — Continuously training the language model on domain-specific data to improve performance.

• A. Generating a final answer by conditioning the LLM on the retrieved information is a core step in a RAG pipeline. It is the generation phase that combines the query with external knowledge to produce a response.
• B. Transforming the user's natural language query into a dense vector is part of the embedding or retrieval step in RAG. It enables semantic similarity search against stored document embeddings.
• C. Retrieving top-k relevant document chunks from the vector store is the retrieval step in a RAG workflow. It provides the context for the LLM to generate an answer.
• D. Continuously training the language model on domain-specific data describes fine-tuning or continued pretraining, which is not a typical step in RAG. RAG relies on retrieving external context at inference time rather than retraining the model.

Correct answer: C — nDCG (normalized Discounted Cumulative Gain)

• A. Incorrect. BLEU score is primarily used for evaluating machine translation by measuring n-gram overlap with reference translations. It is not suitable for evaluating retrieval ranking in a RAG system.
• B. Incorrect. ROUGE-L measures longest common subsequence overlap and is often used for summarization or text generation quality. It is not designed for assessing retrieval performance; it is more applicable to evaluating the generation component of RAG.
• C. Correct. nDCG is a ranking metric commonly used to evaluate information retrieval systems, including the retriever in a RAG pipeline. It rewards relevant documents appearing higher in the ranked list, making it well-suited for assessing retriever performance.
• D. Incorrect. F1 score is used for classification and token-level overlap tasks, such as QA extraction. It does not directly measure ranking quality of retrieved documents in a RAG pipeline.

Correct answer: B — token.tag_

• A. Incorrect. `token.pos_` returns the coarse-grained universal part-of-speech label (e.g., 'NOUN', 'VERB'), not the fine-grained tag like 'NN' or 'VB'.
• B. Correct. `token.tag_` returns the fine-grained part-of-speech tag from the tagger, such as 'NN' for noun or 'VB' for verb, used for Penn Treebank-style tags.
• C. Incorrect. `token.dep_` returns the syntactic dependency relation of the token (e.g., 'nsubj', 'dobj'), describing sentence structure, not the POS tag.
• D. Incorrect. `token.lemma_` returns the base form of the word (e.g., 'run' for 'running'), which is unrelated to the part-of-speech tag.

Correct answer: B — Returns the indices that would sort an array

• A. Incorrect. `argsort` does not return the sorted values themselves; it returns index positions. The sorted array can be obtained using `np.sort()` or `sorted()`.
• B. Correct. `numpy.argsort` returns the indices that would arrange the array in sorted order. For example, with input `[3, 1, 2]`, it returns `[1, 2, 0]` because those indices would sort the array. These indices can then be used to reorder the original array or related arrays consistently.
• C. Incorrect. `argsort` does not directly compute ranks or ordinal positions of elements. While it can be used to derive ranks, its primary purpose is to return the permutation of indices needed to sort the array.
• D. Incorrect. `argsort` does not modify the original array in place and does not return `None`. In-place sorting is handled by `ndarray.sort()` or `np.sort()` with the `out` parameter.

Correct answers: A, C — Ambiguous phrasing often causes off-target model responses.; Hallucinations can still arise despite well-designed prompts.

• A. Correct. Ambiguous phrasing leaves the model uncertain about the intended task, often resulting in vague or irrelevant responses. Clear and specific prompts are essential for guiding the model toward desired outputs.
• B. Incorrect. Models do not inherently understand the user's context; they rely solely on the information provided in the prompt. Explicit context and well-structured prompts are necessary for appropriate responses.
• C. Correct. Hallucinations, where the model generates factually incorrect information, are a known limitation of LLMs and can occur even with carefully designed prompts. Prompt engineering reduces but does not eliminate this risk.
• D. Incorrect. Longer prompts do not guarantee better performance; they can introduce noise or conflicting instructions. Effective prompts are clear, concise, and well-structured rather than simply long.
• E. Incorrect. LLMs are non-deterministic by default; outputs can vary due to sampling parameters like temperature. Identical outputs are not guaranteed unless decoding is made fully deterministic.

Correct answer: B — Prompting the model to explain its reasoning steps before answering.

• A. Incorrect. While asking for only the final answer may yield concise responses, it does not improve reasoning quality or reliability. It hides the model's thought process, making it harder to assess correctness.
• B. Correct. Prompting the model to explain its reasoning steps before answering encourages structured thinking, increases interpretability, and often leads to more reliable outputs by reducing careless errors. This is a common zero-shot prompt engineering technique.
• C. Incorrect. Providing many training examples is part of fine-tuning or few-shot learning, which are different from inference-time prompt engineering. While it can improve performance, it is not a prompt engineering technique per se.
• D. Incorrect. A prompt consisting solely of punctuation characters lacks meaningful task instruction or context, and will not produce a useful or reliable response from the model.

Correct answer: B — ROUGE

• A. Incorrect. BLEU (Bilingual Evaluation Understudy) is primarily used for evaluating machine translation quality by comparing n-gram overlaps with reference translations. It is not specifically designed for abstractive summarization.
• B. Correct. ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is specifically designed for evaluating text summarization, including abstractive summarization. It measures n-gram, word sequence, and word pair overlaps between candidate and reference summaries.
• C. Incorrect. Perplexity measures how well a probabilistic language model predicts a sample. It is a language modeling metric and does not directly assess summarization quality, such as content overlap or faithfulness.
• D. Incorrect. Accuracy is a classification metric that measures the proportion of correct predictions. It is not suitable for evaluating free-form generated text like abstractive summaries.

Correct answer: B — The model likely overfits the specific benchmark distribution.

• A. Incorrect. While insufficient training data can limit generalization, the model's strong results on GLUE suggest it was trained on a substantial and relevant dataset for that benchmark. The issue is more likely related to overfitting to the benchmark distribution rather than data quantity. The team should focus on domain mismatch and transferability.
• B. Correct. The model may have overfit to the GLUE benchmark's specific distribution, meaning it performs well on data similar to GLUE but fails to generalize to out-of-distribution data like the internal customer service dataset. This is a common limitation in benchmark-focused models, where models are heavily tuned to the benchmark's characteristics, leading to poor domain transfer. The team should investigate whether the paper's methodology involved benchmark-specific tuning or narrow data coverage.
• C. Incorrect. An outdated tokenization method could affect performance, especially for domain-specific text, but it is unlikely to cause such a stark discrepancy between GLUE and a different dataset unless the tokenization is fundamentally incompatible with the internal data. Overfitting to the benchmark distribution is a more likely explanation for the observed mismatch.
• D. Incorrect. Modern LLMs, especially those achieving state-of-the-art results on GLUE, almost certainly incorporate attention mechanisms. This option is implausible given the context, as attention-based architectures are foundational for such models. The mismatch is better explained by overfitting to the benchmark distribution.

Correct answer: A — Strong negative for X-Y; no correlation for X-Z

• A. Correct. A correlation coefficient of -0.95 indicates a very strong negative linear relationship between X and Y, meaning as one variable increases, the other tends to decrease. A coefficient of 0.10 is very close to zero, indicating little to no linear relationship between X and Z.
• B. Incorrect. The correlation between X and Y is strong negative, not weak positive. The correlation for X and Z is near zero (no correlation), not strong negative.
• C. Incorrect. The sign and magnitude for X-Y are wrong: -0.95 is strongly negative, not positive. For X-Z, 0.10 is weakly positive, not weakly negative.
• D. Incorrect. X-Y does show a strong linear relationship, but it is negative, not just 'strong.' X-Z does not show a strong linear relationship because 0.10 is very weak and close to zero.

Correct answer: D — Neither; use the F1-score for comparison, as accuracy is misleading for imbalanced data.

• A. Incorrect. Higher accuracy alone is not reliable for imbalanced datasets; a model can achieve high accuracy by predicting the majority class, hiding poor performance on the minority class.
• B. Incorrect. While higher precision indicates fewer false positives, it does not consider false negatives. For imbalanced data, recall is equally important, and a single metric like precision is insufficient to compare models.
• C. Incorrect. Adjusting the threshold can change precision and recall, but there is no guarantee that Model X will outperform Model Y after tuning. The outcome depends on the trade-off and specific use case.
• D. Correct. The F1-score balances precision and recall, making it a more reliable metric for imbalanced datasets. Accuracy can be misleading due to class imbalance, so using F1-score provides a better basis for model comparison.

Correct answers: A, C, F — One-hot encoding of categorical variables to create binary columns; Normalizing numeric features to a common scale such as min-max scaling; Creating polynomial features from numeric features to model interactions

• A. Correct. One-hot encoding converts categorical variables into binary columns, enabling machine learning models to process them. This is a classic feature engineering technique that transforms input representation to improve learning.
• B. Incorrect. Dropout is a regularization method used during neural network training to reduce overfitting. It does not modify input data features, so it is not a feature engineering technique.
• C. Correct. Normalizing numeric features rescales inputs to a common range (e.g., min-max scaling), which is a standard preprocessing and feature engineering step. It improves model training stability and performance.
• D. Incorrect. Gradient descent is an optimization algorithm for updating model parameters during training. It affects how the model learns, not how features are constructed or transformed, so it is not feature engineering.
• E. Incorrect. Selecting the learning rate is hyperparameter tuning, not feature engineering. It controls the training process rather than the input features themselves.
• F. Correct. Creating polynomial features expands the feature set to include interactions and nonlinear relationships. This is a direct example of feature engineering because it constructs new inputs from existing ones.

Correct answers: A, C, F — GPU utilization on the serving nodes; Current queue length for incoming inference requests; Disk I/O wait time observed on the serving node

• A. Correct. GPU utilization directly indicates how effectively compute resources are being used for inference. High utilization suggests efficient use, while low utilization may point to underutilization or bottlenecks. Monitoring this metric helps assess serving performance and scalability.
• B. Incorrect. While user count can indicate demand, it is an application-level metric that does not directly measure model serving performance or scalability. It may correlate with load but is less actionable than inference-specific metrics.
• C. Correct. Queue length for incoming inference requests is a strong indicator of serving scalability and latency under load. A growing queue often signals that the model or serving system cannot keep up with demand, prompting scaling actions.
• D. Incorrect. Model accuracy on validation data measures model quality, not serving performance or scalability. It is important for evaluation but does not indicate whether the deployment can handle traffic reliably or efficiently.
• E. Incorrect. Time since the most recent model update relates to model freshness or versioning, not serving throughput or reliability. It may be relevant for operational tracking but is not a primary metric for scalability or performance.
• F. Correct. Disk I/O wait time on the serving node is an infrastructure performance metric that can reveal storage bottlenecks affecting inference latency and reliability. High wait times can slow down the serving system even if GPU utilization is low.

Correct answer: C — To track intent, slots, and conversation context across turns.

• A. Incorrect. While logging user messages can be useful for analytics, auditing, or debugging, that is not the primary purpose of dialog state. Dialog state is about managing the conversation's flow and context, not just logging.
• B. Incorrect. Token limit management is a separate concern handled by prompt construction, context trimming, or memory management. Dialog state is about preserving meaningful conversational information, not primarily about controlling token usage.
• C. Correct. Dialog state tracks intent, slots, and conversation context across turns so the assistant can maintain continuity in the conversation. This helps the system understand what the user has already said, what information is missing (e.g., in task-oriented systems), and how to respond consistently and coherently.
• D. Incorrect. The LLM's pre-trained knowledge is stored in its model parameters, not in dialog state. Dialog state captures temporary, conversation-specific information rather than the model's learned knowledge.

Correct answer: A — sns.heatmap()

• A. Correct. sns.heatmap() is the standard Seaborn function for creating heatmaps, often used to visualize correlation matrices. It displays the correlation values with color gradients, making patterns easy to identify.
• B. Incorrect. sns.corrplot() is not a valid function in Seaborn. Seaborn does not provide a function with this name; correlation visualizations are achieved using heatmap() or pairplot().
• C. Incorrect. sns.clustermap() creates a clustered heatmap with hierarchical clustering applied to rows and columns. While it can also display correlation matrices, it adds clustering, which is not necessary for a simple heatmap visualization.
• D. Incorrect. sns.pairplot() generates a matrix of scatterplots to show pairwise relationships between variables. It does not produce a heatmap of correlation values.

Correct answer: D — To compute stat transforms

• A. Incorrect. Saving a plot to a file is a common function of plotting libraries (e.g., plt.savefig() in Matplotlib).
• B. Incorrect. Creating a new plot canvas initializes the figure (e.g., plt.figure()), which is essential for drawing visualizations.
• C. Incorrect. Adding a legend is a standard step to label plot elements (e.g., plt.legend()).
• D. Correct. Computing statistical transformations is not a primary role of plotting libraries; it is typically performed by data analysis libraries such as Pandas, NumPy, or SciPy.

Correct answer: B — Heatmap

• A. Incorrect. A pie chart is used to show the proportion of categories within a whole, not correlations between variables. It does not effectively represent pairwise correlation patterns across a dataset.
• B. Correct. A heatmap is ideal for visualizing correlation matrices, as it uses color intensity to represent the strength and direction of correlations between multiple variables. It makes it easy to compare many variable relationships at once and scales well to high-dimensional datasets.
• C. Incorrect. A line chart is best for displaying trends over time or another ordered sequence, not for showing correlations among multiple variables. It is not suitable for visualizing a full correlation matrix.
• D. Incorrect. A 3D scatter plot can show relationships among three variables, but it becomes hard to interpret and does not scale well to multiple variables. For correlation across many variables, a heatmap is typically more appropriate.

Correct answer: C — Streamlit

• A. Incorrect. Matplotlib is a plotting library for creating static, animated, or interactive visualizations, but it is not designed for building full interactive web dashboards with filtering and linked charts; it requires additional frameworks to achieve that.
• B. Incorrect. Seaborn is a high-level statistical visualization library built on Matplotlib, primarily for attractive static plots, and it does not support real-time filtering, linked charts, or dashboard creation.
• C. Correct. Streamlit is a Python framework specifically designed for quickly building interactive data apps and dashboards. It supports widgets for filtering time ranges and can display multiple coordinated charts, making it ideal for real-time monitoring of deployed model metrics.
• D. Incorrect. PIL (Python Imaging Library) is used for image processing and manipulation, not for data visualization or dashboard creation; it lacks support for interactive charts and dashboard features.

Correct answer: C — Report the issue to your supervisor and discuss handling methods.

• A. Incorrect. Removing all rows with missing values can lead to significant data loss and bias, especially if the missingness is not random. This should not be the default action without proper justification and guidance from a supervisor.
• B. Incorrect. While imputing missing values can be a valid technique, doing it unilaterally without consulting a senior team member is not appropriate in a supervised analysis setting. The choice of imputation method depends on the data distribution and analysis goals; collaboration is key.
• C. Correct. In a supervised setting, missing data issues should be escalated to a supervisor so the team can decide on an appropriate handling method. This ensures the approach is validated, aligned with best practices, and consistent with analysis objectives.
• D. Incorrect. Ignoring missing values can lead to biased or inaccurate results, as missing data can significantly impact statistical analyses and model performance. Missing data must be assessed and handled deliberately rather than assumed negligible.

Correct answers: A, F — Interpreting a strong correlation as direct evidence of a causal relationship.; Relying solely on p-values without assessing practical significance.

• A. Correct. Interpreting a strong correlation as direct evidence of a causal relationship is a common pitfall because correlation does not imply causation. A third variable or reverse causality could explain the observed relationship.
• B. Incorrect. Using a large sample size to increase statistical power is a best practice, not a pitfall. It helps detect true effects and reduces the risk of Type II errors.
• C. Incorrect. Counterbalancing experimental conditions to reduce order effects is a valid experimental design technique, not a pitfall when interpreting results.
• D. Incorrect. Randomly assigning participants to experimental conditions eliminates selection bias and is a sound experimental design principle, not an error in interpretation.
• E. Incorrect. Conducting a power analysis prior to the experiment is a best practice to ensure adequate sample size and reliable conclusions, not a pitfall.
• F. Correct. Relying exclusively on p-values without evaluating practical importance (e.g., effect size) is a pitfall because statistical significance does not always equate to meaningful or impactful results. It can lead to overemphasizing significance while ignoring real-world relevance.

Correct answer: A — Extending the test may be warranted, but early stopping based on a trend can inflate false positives.

• A. Correct. Extending the test may gather more data to reach statistical significance, especially if the effect takes time to emerge. However, deciding to stop early based on a short-term trend (e.g., the last 3 days) can inflate false positive rates due to peeking at interim results. Any extension should be planned to avoid biases.
• B. Incorrect. A p-value above 0.05 means we fail to reject the null hypothesis, but we do not accept it as true. The null hypothesis is never proven; we only have insufficient evidence to reject it.
• C. Incorrect. A three-day trend is too short to override the overall experiment result. Such short-term patterns can easily arise from random variation or external factors and are not sufficient evidence to conclude the new flow is better.
• D. Incorrect. While an inadequate sample size can contribute to a non-significant result, the given information does not confirm the sample size was insufficient. The team should evaluate statistical power, variance, and whether the test duration captured sufficient user behavior before deciding to restart.

Correct answer: B — To track dialogue history for coherent conversations.

• A. Storing authentication tokens is a security task, not the primary role of conversation memory, which is about preserving dialogue context across turns.
• B. Conversation memory tracks dialogue history, enabling the chatbot to maintain context and respond coherently over multiple turns, which is its primary purpose.
• C. Caching API responses is a performance optimization, unrelated to the conversational context that memory provides.
• D. Logging errors is for debugging and monitoring, not for preserving the flow and meaning of a conversation.

Correct answer: B — Enforcing rules on the conversational flow and content.

• A. Dialog rails in NeMo Guardrails are not related to physical infrastructure. They are software constructs for managing conversational behavior, not hardware or deployment environments.
• B. Dialog rails are used to enforce rules on conversational flow and content, ensuring the LLM's responses stay within predefined boundaries, such as avoiding harmful or off-topic content. This is central to NeMo Guardrails for guiding and restricting dialogue behavior.
• C. While logging may be useful for auditing and compliance, that is not the primary purpose of a dialog rail. Dialog rails are about governing the conversation itself, not recording it.
• D. User interface management is handled by the application layer or frontend components, not by dialog rails. NeMo Guardrails operates on conversation logic and safety constraints rather than display elements.

Correct answer: B — Use a cluster of NVIDIA A100 GPUs in a DGX system for maximum throughput.

• A. Incorrect. While the NVIDIA H100 is powerful, a single GPU cannot match the throughput and scalability of a multi-GPU cluster for large-scale training. Distributed training across multiple GPUs is essential for maximizing throughput.
• B. Correct. A cluster of NVIDIA A100 GPUs in a DGX system provides high-performance AI training with maximum throughput, scalability, and optimized inter-GPU communication (e.g., NVLink). This setup is ideal for large-scale generative AI models.
• C. Incorrect. The NVIDIA L40S is optimized for inference and cost-effective workloads, not for high-throughput training. It lacks the compute power and scalability of DGX-class systems.
• D. Incorrect. Multiple RTX 4090 GPUs offer strong consumer-grade performance but lack enterprise features like NVLink and efficient scaling. They are not designed for maximum throughput in large-scale generative AI training.

Correct answer: B — Implement data validation on incoming data.

• A. Incorrect. Alerts on model inference latency monitor serving performance, but they do not directly address the functioning of data collection or software processes. This is an operational metric for model serving, not a data pipeline monitoring measure.
• B. Correct. Implementing data validation on incoming data ensures the data collection pipeline produces valid, expected, and schema-compliant inputs. This directly monitors data quality and integrity, which is critical for reliable ML workflow outputs.
• C. Incorrect. Increasing the frequency of model retraining is a model lifecycle strategy to adapt to drift, not a monitoring activity. It does not monitor the functioning of data collection or other software processes.
• D. Incorrect. A canary release is a deployment technique to reduce risk when introducing a new model. It helps observe behavior after release but is not a monitoring activity for existing data collection or software processes.

Correct answer: B — To compute the dot product or matrix multiplication.

• A. Incorrect. The element-wise product of two arrays is performed using `np.multiply()` or the `*` operator, not `np.dot()`.
• B. Correct. `np.dot()` computes the dot product of two vectors and performs matrix multiplication for arrays with compatible shapes.
• C. Incorrect. Creating a diagonal matrix from a vector is done using `np.diag()`, not `np.dot()`.
• D. Incorrect. The determinant of a square matrix is calculated using `np.linalg.det()`, not `np.dot()`.

Correct answer: B — It enables efficient similarity search via ANN algorithms.

• A. Incorrect. Vector databases are not designed for full ACID transactions across multiple embedding vectors. They prioritize fast similarity search over transactional consistency, and any transactional support is typically limited.
• B. Correct. Vector databases are optimized for approximate nearest neighbor (ANN) search, enabling efficient similarity search in high-dimensional vector spaces. This is their core strength, used for tasks like semantic search and recommendation systems.
• C. Incorrect. Vector databases do not provide automatic schema inference for unstructured text data. They store embeddings and metadata for similarity search; schema inference is more typical of document databases or NLP libraries.
• D. Incorrect. While vector databases aim to scale well, they do not guarantee linear scalability without trade-offs in query performance. Performance depends on indexing method, hardware, and dataset size.

Correct answers: A, C — Storing confidential data like API keys securely.; Configuring behavior across development and production.

• A. Correct. Environment variables are often used to store sensitive data like API keys, tokens, and passwords without hardcoding them directly into scripts. This helps reduce the risk of exposing confidential data in source code repositories and can be secured via system-level permissions.
• B. Incorrect. Compiling source code into executable binaries is a build process task handled by tools like gcc or make, not a typical use of environment variables in scripting.
• C. Correct. Environment variables are widely used to configure application behavior across different environments, such as development, testing, and production. This makes scripts portable and easier to manage without changing the code (e.g., DEBUG=true vs. DEBUG=false).
• D. Incorrect. Managing disk partitions and mounting drives is a system administration task handled by commands like fdisk or mount, not environment variables.
• E. Incorrect. Installing software packages is done via package managers like apt, yum, or pip. While environment variables may influence package manager behavior, they are not a common scripting use for this purpose.
• F. Incorrect. Creating and managing user accounts is an administrative task handled by commands like useradd or usermod, not environment variables.

Correct answer: C — To grant individuals authority over their data usage.

• A. Incorrect. While intellectual property regulations are important, data consent is not primarily about compliance with IP laws. It focuses on individual rights over personal data.
• B. Incorrect. Improving model accuracy and performance can be a benefit of using data, but it is not the primary purpose of obtaining consent. Consent is fundamentally about respecting privacy rights and giving people control over their personal information.
• C. Correct. Data consent grants individuals the authority to control how their personal data is collected, used, and shared. This is central to balancing privacy with responsible AI/data practices and aligns with regulations like GDPR.
• D. Incorrect. Reducing data processing costs may be a side effect of using less data, but it is not the purpose of consent. Consent is about ethical and legal control over personal data, not cost optimization.

Correct answers: A, B, C — General Data Protection Regulation (GDPR); Health Insurance Portability and Accountability Act (HIPAA); California Consumer Privacy Act (CCPA)

• A. Correct. GDPR is a comprehensive data privacy regulation in the EU that emphasizes user consent, data protection, and individual rights over personal data.
• B. Correct. HIPAA is a US regulation that protects sensitive patient health information and requires patient consent for disclosure, balancing privacy and consent in healthcare.
• C. Correct. CCPA is a California state law that grants consumers rights over their personal data, including the right to know, delete, and opt-out of data sales, emphasizing consent mechanisms.
• D. Incorrect. PCI DSS focuses on securing payment card data and transactions, not on broader data privacy or consent.
• E. Incorrect. SOX is a US law aimed at improving corporate governance and financial transparency, not data privacy or consent.
• F. Incorrect. FISMA is a US law that mandates federal agencies to implement information security programs, but it does not focus on data privacy or individual consent.

Correct answer: D — NVIDIA Triton Inference Server

• A. NVIDIA TensorRT is a high-performance inference optimization SDK that accelerates deep learning model deployment. It improves performance and efficiency but does not directly address trustworthiness features such as security, verification, or monitoring.
• B. NVIDIA OptiX is a ray tracing and AI-accelerated graphics API used for rendering applications. It is unrelated to AI model deployment or trustworthiness.
• C. NVIDIA cuQuantum is a library for accelerating quantum computing simulations and has no direct role in improving AI trustworthiness.
• D. NVIDIA Triton Inference Server is designed to deploy and manage AI models at scale, supporting features like model versioning, monitoring, logging, and ensemble workflows. These capabilities enhance operational control, reproducibility, and observability, which are key aspects of trustworthy AI.

Correct answer: A — NVIDIA FLARE (Federated Learning)

• A. Correct. NVIDIA FLARE is a federated learning framework that enables multiple parties to collaboratively train AI models without sharing raw data, thereby preserving privacy and enhancing trustworthiness. It directly addresses privacy concerns in distributed learning scenarios.
• B. Incorrect. NVIDIA DGX is a high-performance AI supercomputer for training and inference. While it powers AI workloads, it does not inherently provide privacy-preserving collaborative training. Federated learning capabilities are not a feature of DGX hardware.
• C. Incorrect. NVIDIA NGC is a catalog of GPU-optimized software, containers, and models. It supports AI development but lacks built-in federated learning or privacy-preserving mechanisms. It is a resource hub, not a privacy framework.
• D. Incorrect. NVIDIA Base Command Platform is a management and orchestration tool for AI workflows across hybrid and multi-cloud environments. It does not offer federated learning capabilities or specialize in privacy-preserving collaboration.

Correct answer: C — Collect more images from the underrepresented group to balance the data.

• A. Incorrect. Increasing model complexity does not address dataset imbalance and may amplify overfitting to existing biases.
• B. Incorrect. Excluding data from the affected group removes important examples, worsening bias and reducing model representativeness.
• C. Correct. Collecting more data from the underrepresented group helps balance the dataset, improving exposure and reducing bias at the source.
• D. Incorrect. Adjusting thresholds is a post-hoc measure that does not fix the root cause of bias from imbalanced training data.