NVIDIA Certified Associate - Generative AI LLMs practice questions

Question 1

An organization has decided to use knowledge distillation in NeMo Framework to obtain a cheaper model from its large LLM, and a developer needs to state the goal of the technique before kicking off the job. Which description matches the NeMo documentation?

Accepted Answer

It transfers knowledge from a large teacher model to a smaller, faster, less resource-intensive student model.

Answer

It replicates the large model across GPUs so each replica serves a slice of incoming production traffic.

Answer

It lowers the teacher's weights to half precision so the same model runs faster at inference time.

Answer

It prunes the teacher's layers until the remaining network fits within a single GPU's memory budget.

Question 2

A developer is debugging why token-usage and message-id information is missing from logs in a LangChain app and needs to know what a Message object actually carries. According to the LangChain messages documentation, which describes a message?

Accepted Answer

A message carries a role, the actual content, and optional metadata such as token usage.

Answer

A message holds only plain text content, with no role label and no associated metadata fields.

Answer

A message is a vector embedding of the conversation used solely for similarity retrieval steps.

Answer

A message is a checkpoint of the full graph state persisted for human-in-the-loop workflows.

Question 3

A developer is comparing GridSampler, RandomSampler, and TPESampler for an Optuna study on RAPIDS and worries that each switch will force a rewrite of the objective function. According to the RAPIDS and Optuna guide, what is true when switching between samplers?

Accepted Answer

The objective function stays the same; samplers are a built-in, swappable Optuna option

Answer

The objective must be rewritten because each sampler expects a different return type value

Answer

The study must be recreated because the sampler is fixed once the first trial has run

Answer

The objective must add a sampler-specific pruning callback for every algorithm being tried

Question 4

A developer is migrating a pandas feature pipeline to cuDF and constructs a Series from a Python list that contains np.nan; by default cuDF turns that entry into <NA> rather than keeping it as a real floating-point NaN. What is the correct way to make the Series constructor keep the value as a float NaN?

Accepted Answer

Pass nan_as_null=False to the Series constructor so the NaN is kept as a float NaN

Answer

Cast the column to int64 first, since cuDF preserves NaN inside integer columns by default

Answer

Call .fillna(np.nan) after construction so the <NA> entries are rewritten as float NaN

Answer

Set the column dtype to category, which forces cuDF to retain the raw NaN float value

Question 5

A team is being tasked with evaluating a RAG assistant and must measure the factual consistency of each generated answer against its retrieved context on a 0-1 scale using a judge LLM. According to the NeMo RAG evaluation documentation, which metric captures this?

Accepted Answer

Faithfulness, the factual consistency of an answer against retrieved context

Answer

Recall@k, the fraction of relevant documents within the top k retrieved results

Answer

nDCG@k, the normalized discounted cumulative gain measuring ranking quality

Answer

Answer Relevancy, which penalizes incomplete or redundant generated responses

Question 6

A developer is optimizing a training loop with mixed precision and must order the per-iteration steps correctly so the performance benefits are captured without losing accuracy. Per the NVIDIA mixed-precision guide, which sequence is correct?

Accepted Answer

Make a half-precision copy for forward/back-propagation, then convert gradients to single precision to update the master.

Answer

Update the half copy in the forward pass, skip back-propagation, then discard the single-precision master each step.

Answer

Convert the weights to INT8 for the forward pass, run back-propagation in FP32, then prune the smallest weights.

Answer

Keep both the forward and backward passes in single precision, then cast the master copy to half precision once.

Question 7

An engineer is choosing between an LLM-specific NIM container with pre-built engines and a Multi-LLM NIM that builds engines dynamically. According to the NVIDIA NIM introduction, which statement correctly describes Multi-LLM NIM's source model flexibility?

Accepted Answer

Multi-LLM NIM supports broad model ranges and builds optimized engines dynamically, accepting models from NGC, Hugging Face, or local storage paths as input sources.

Answer

Multi-LLM NIM accepts models only from a single proprietary registry maintained by NVIDIA, which is what gives it the documented flexibility to build optimized engines dynamically.

Answer

Multi-LLM NIM requires every model to be re-trained on NGC before it can be loaded, which is the constraint that lets it build optimized engines dynamically across configurations.

Answer

Multi-LLM NIM does not support models from Hugging Face under any configuration, which the NIM introduction documents as the source restriction for its dynamic engine builds.

Question 8

An MLOps team is migrating a SteerLM experiment from a small annotated seed dataset to a much larger training corpus and needs to extend the labeled data without sending every example back to human annotators. According to the SteerLM documentation, what does Stage 2 of the pipeline do to support this scale-up?

Accepted Answer

Stage 2 annotates diverse datasets with predicted attribute scores from the Stage 1 reward model in order to expand training data beyond what humans annotated by hand.

Answer

Stage 2 replaces the LLM with a retrieval index and stops training entirely after Stage 1, which the documentation cites as the way SteerLM scales labels beyond the seed set.

Answer

Stage 2 freezes the reward model and pretrains the base LLM from scratch on raw web crawl data, presented in the documentation as SteerLM's way to scale labeled data.

Answer

Stage 2 reinforces good behaviors through positive/negative feedback using PPO-style rollouts, which the SteerLM documentation lists as its way to scale annotated datasets.

Question 9

A developer is migrating a groupby-heavy pandas script to cuDF and finds that a groupby().apply() that previously returned a pandas Series now yields a single-column DataFrame, and a call to describe() inside the callable raises an error. According to the cuDF GroupBy documentation, how should the developer interpret these differences?

Accepted Answer

Results may not match pandas exactly, so post-processing is needed and some apply cases such as describe are unsupported.

Answer

cuDF guarantees byte-for-byte parity with pandas apply, so any difference must indicate a corrupted GPU driver installation.

Answer

The Series-to-DataFrame change means the data silently moved to CPU and must be reloaded with from_arrow().

Answer

describe() inside apply is supported only when the groups are sorted, so the developer must pre-sort the keys.

Question 10

An alignment team is evaluating SteerLM to give end users inference-time control over attributes like helpfulness and creativity, instead of baking in predetermined preferences as RLHF does. According to the NeMo documentation, what is the FIRST of SteerLM's four steps?

Accepted Answer

Train an attribute prediction model on human-annotated data to score quality

Answer

Perform attribute-conditioned SFT so the LLM conditions on the chosen attribute set

Answer

Bootstrap by sampling diverse responses conditioned on the maximum quality value

Answer

Annotate diverse datasets by predicting their per-attribute scores with the model

Question 11

A team is designing a baseline image classifier and a baseline sequential model and is reviewing the TensorRT architecture overview for the right family of networks. Which pairing of architecture and traditional use case is consistent with the documentation?

Accepted Answer

CNNs use spatial filters for image processing and RNNs maintain hidden states for sequential data, both largely superseded by Transformer architectures.

Answer

CNNs are positioned as recurrent state machines for hidden-state sequence handling, while RNNs are described as convolutional filters for image processing.

Answer

CNNs and RNNs are described as identical architectures with interchangeable use cases, both equally suited to image classification and sequence modelling.

Answer

Both CNNs and RNNs are described as fully deprecated and explicitly no longer supported by TensorRT for any image or sequence task.

Question 12

An MLOps team is standing up NeMo model training and wants multi-GPU, mixed-precision, and checkpointing without writing PyTorch training boilerplate, so engineers can focus on the ASR/NLP/TTS domain logic. According to the NeMo Framework documentation, what enables this?

Accepted Answer

NeMo leverages PyTorch Lightning, which decouples the AI code from the PyTorch training code.

Answer

NeMo replaces PyTorch with TensorRT-LLM so all training runs through the inference engine instead.

Answer

NeMo uses the Triton model repository to schedule distributed training jobs across the GPU fleet.

Answer

NeMo relies on trtexec to orchestrate multi-GPU training and to emit checkpoints during each epoch.

Question 13

A team is evaluating an LLM on machine-translation and summarization outputs and needs reference-based metrics that compare generated text against human-written references. According to the NVIDIA LLM evaluation blog, which TWO metrics fit this requirement? (Choose TWO.)

Accepted Answer

BLEU, which evaluates machine-translation quality against reference translations

Accepted Answer

ROUGE, which measures overlap against human-generated reference summaries

Answer

Perplexity, which quantifies a model's uncertainty when predicting word sequences

Answer

Faithfulness, which scores answer consistency against the retrieved context

Question 14

A data team is being built to analyze a content library where each item includes a text description and several reference images, and they need to visualize attribute differences across images for the same item. Which VLM capability supports this comparison-style analysis?

Accepted Answer

Multi-image processing that lets VLMs accept several images simultaneously and is documented as enabling comparison and contextual learning across multiple reference images.

Answer

Single-image classification heads that only accept one input image at a time, which the prompt-engineering guide cites as the path to multi-image cross-comparison analysis.

Answer

Text-only summarization that ignores the images entirely, which the prompt-engineering guide identifies as the recommended method for multi-image comparison and contextual learning.

Answer

OCR-only pipelines that read text from images but cannot compare visual content, which the prompt-engineering guide labels as the technique for multi-image attribute comparison.

Question 15

A developer needs to activate several pre-built NeMo Guardrails from the catalog and combine multiple approaches for defense in depth. According to the catalog documentation, where is each guardrail added in the configuration?

Accepted Answer

Into the input, output, or retrieval section of the config.yml configuration

Answer

Into the dialog section of the Colang flow files used for multi-turn logic

Answer

Into the SteerLM attribute prediction model section of the alignment config file

Answer

Into the relevant_chunks context variable consumed by the fact-checking rail

NVIDIA Certified Associate - Generative AI LLMs

Sample questions

Sources

Similar exams