NCA-GENL Practice Test 20 — 15 Questions

Question 1

A machine learning engineer is preprocessing a dataset for a generative AI model and wants to ensure that the input features have a similar scale. Which technique is most appropriate?

Accepted Answer

Normalize numerical features to have zero mean and unit variance.. Normalizing numerical features to zero mean and unit variance brings them to a similar scale, which aids optimization and model performance. One-hot encoding is for categorical data, outlier removal is for data cleaning, and learning rate is unrelated to feature scaling. Thus, normalization is the correct preprocessing technique.

Answer

Increase the learning rate during training.

Answer

Remove outliers from the dataset.

Answer

Apply one-hot encoding to all categorical features.

Question 2

A media company fine-tunes an NVIDIA Nemotron model on licensed news articles to build a summarization tool. Legal asks how the team can demonstrate that the training data was lawfully acquired and that the model does not reproduce copyrighted passages verbatim. Which combination of practices best addresses both concerns?

Accepted Answer

Maintain a data provenance record for each training source and run memorization tests that check for verbatim reproduction of training passages.. Data provenance records and memorization testing together provide auditable evidence for both legal concerns. Provenance documents origin and licensing, while memorization tests measure whether the model reproduces training passages verbatim. Neither alone is sufficient: provenance without testing leaves reproduction risk unmeasured, and testing without provenance leaves acquisition unverified.

Answer

Apply differential privacy during fine-tuning and assume it eliminates all copyright risk without further testing.

Answer

Add a disclaimer to the tool's user interface stating that summaries may resemble source articles.

Answer

Increase the model's parameter count and retrain on a larger unlicensed web crawl to improve summarization quality.

Question 3

A developer is building a customer-support assistant that must retrieve answers only from an approved internal knowledge base and cite the source document for each reply. They are using NVIDIA NIM microservices for the LLM and an embedding model, and they need the application layer to enforce citation behavior and reject answers that are not grounded in retrieved passages. Which software development approach best enforces this grounding requirement?

Accepted Answer

Post-process the model output with a separate NLI-style entailment check that compares each generated claim against the retrieved passages, and suppress any response whose claims are not entailed.. Grounding must be enforced programmatically in the application layer rather than trusted to the model's instruction-following. After the NIM LLM generates a draft answer from retrieved passages, an entailment or claim-verification step confirms that every statement is supported by those passages; unsupported claims are removed or the whole response is rejected. This yields auditable citations and prevents ungrounded content from reaching the user.

Answer

Fine-tune the NIM-hosted LLM on the internal knowledge base so that all answers are memorized in the model weights and retrieval becomes unnecessary.

Answer

Rely on the system prompt to instruct the model to answer only from context and to include citations, and ship the assistant once spot checks look acceptable.

Answer

Increase the model's temperature parameter so the assistant paraphrases the retrieved passages more freely and avoids repeating source wording verbatim.

Question 4

A developer is building a text summarization assistant that must produce concise, faithful summaries of long support tickets. The team wants to fine-tune a pre-trained large language model on a small labeled dataset of ticket-summary pairs. Which training approach best matches this goal?

Accepted Answer

Supervised fine-tuning (SFT) on the ticket-summary pairs. The team has labeled input-output pairs for a specific generation task, which is exactly what supervised fine-tuning is designed for. SFT updates the pre-trained model's weights to map support tickets to concise summaries, leveraging existing language knowledge while learning the task. Other approaches either ignore the labels, require far more data and compute, or do not modify the model to perform the target task.

Answer

Unsupervised pre-training from scratch on the ticket corpus

Answer

Reinforcement learning from human feedback (RLHF) using pairwise summary preferences

Answer

Retrieval-augmented generation (RAG) with no model weight updates

Question 5

A data scientist is preparing a dataset of 50,000 customer support chat transcripts to fine-tune an LLM for a helpdesk assistant. The raw text contains HTML tags, inconsistent whitespace, and occasional personal information such as email addresses. Which preprocessing step should be performed FIRST to prepare the text for tokenization?

Accepted Answer

Clean the text by stripping HTML, normalizing whitespace, and masking personally identifiable information.. Raw chat transcripts often contain markup and private data that degrade fine-tuning quality and raise privacy concerns. Cleaning the text by removing HTML, normalizing whitespace, and masking emails or phone numbers before tokenization ensures the tokenizer produces meaningful tokens and that sensitive data never reaches the model. This ordering also avoids retokenization work later.

Answer

Apply tokenization using the model's tokenizer.

Answer

Convert all text to lowercase and remove all punctuation.

Answer

Split the transcripts into training and validation sets.

Question 6

An enterprise is deploying an LLM-based HR assistant that answers questions about leave policies. The team wants to ensure the assistant cites the current policy document rather than relying on the model's parametric memory, which may be outdated. Which approach best supports trustworthy, verifiable answers?

Accepted Answer

Use retrieval-augmented generation to fetch relevant passages from the current policy document and instruct the model to cite them in its answer.. Retrieval-augmented generation grounds answers in the current policy document and enables citations, so employees can verify the source. It also decouples knowledge updates from model retraining, letting the team refresh the index when policies change. Fine-tuning, temperature tuning, and system prompts do not provide source-backed, current answers on their own.

Answer

Lower the model's temperature to zero and rely on the model's internal knowledge of HR policies.

Answer

Fine-tune the model weekly on the latest policy PDF so the knowledge is embedded in the weights.

Answer

Add a system prompt instructing the model to always answer truthfully about leave policies.

Question 7

A media company uses an LLM to generate article drafts. Legal requires that the system never reproduce long verbatim passages from copyrighted training sources. Which mitigation most directly reduces this risk at generation time?

Accepted Answer

Apply a decoding-time constraint that blocks or rewrites outputs containing long n-gram matches against a reference corpus of copyrighted text.. Copyright risk from verbatim reproduction is best addressed by detecting long overlapping sequences between generated output and known source text, then blocking or rewriting those spans before delivery. This is a generation-time control that directly measures the prohibited behavior. Temperature changes, prompt instructions, and stylistic fine-tuning do not verify overlap and therefore cannot guarantee that infringing passages are stopped.

Answer

Fine-tune the model on a corpus of original company articles so that its outputs resemble proprietary writing rather than external sources.

Answer

Lower the sampling temperature to make the model more deterministic and consistent in its phrasing.

Answer

Add a system prompt instructing the model to always paraphrase and never quote more than a few consecutive words from any source.

Question 8

A developer is building a document summarization service using an NVIDIA NIM microservice for Llama-3. The service must process batches of 20 documents at once to maximize throughput. The NIM container is already running with default settings. Which API parameter should the developer configure to enable efficient batched inference?

Accepted Answer

Configure the `max_batch_size` parameter when starting the NIM container, and send multiple prompts in a single request using the `prompt` field as an array.. To efficiently process a batch of documents with an NVIDIA NIM microservice, the developer must ensure the server is configured with an appropriate `max_batch_size` at launch. Then, the client should send a single request containing an array of prompts. This leverages the NIM's dynamic batching capabilities, which group concurrent requests or batched prompts to maximize GPU utilization and throughput. Other options either misuse parameters or misunderstand the batching mechanism.

Answer

Set the `batch_size` field in the request payload to 20, allowing the NIM to process all documents in a single forward pass.

Answer

Enable the `tensor_parallel` option in the request headers and set it to 20, so the NIM distributes the batch across multiple GPUs.

Answer

Use the `stream` parameter set to `true` and send each document as a separate request; the NIM will automatically coalesce them into a batch on the server side.

Question 9

A machine learning engineer is evaluating a language model's performance on a text summarization task. The model achieves a BLEU score of 0.45 and a ROUGE-L score of 0.62 on the test set. The engineer wants to understand how well the model captures the overall meaning of the source documents. Which evaluation metric should they prioritize?

Accepted Answer

ROUGE-L score, because it measures longest common subsequence and recall of reference content.. ROUGE-L is designed for summarization evaluation and measures recall of the longest common subsequence, which correlates with how much essential content from the reference is included. BLEU focuses on precision and is better for translation. Perplexity and accuracy do not assess summary quality against references. Therefore, ROUGE-L is the most appropriate metric for capturing overall meaning.

Answer

Perplexity, because it measures how well the model predicts the next token.

Answer

Accuracy, because it measures the percentage of correctly predicted tokens.

Answer

BLEU score, because it measures n-gram overlap with reference summaries.

Question 10

A team fine-tunes a Llama-3 8B model with NVIDIA NeMo Framework and must ship an inference artifact that a C++ service can load without a Python runtime. They want maximum throughput on Hopper GPUs and plan to serve many concurrent requests with in-flight batching. Which artifact and runtime pairing best satisfies these constraints?

Accepted Answer

Export a TensorRT-LLM engine and run it through the TensorRT-LLM C++ runtime with in-flight batching enabled.. The requirement combination of no Python runtime, Hopper-tuned kernels, and in-flight batching points to a TensorRT-LLM engine executed by its C++ runtime. Checkpoint formats, ONNX, and the transformers pipeline each fail at least one of the three constraints, most commonly the C++ deployment or the continuous batching capability.

Answer

Convert the checkpoint to ONNX and execute it with the ONNX Runtime CUDA execution provider.

Answer

Save the fine-tuned checkpoint in NeMo .nemo format and load it with the NeMo inference toolkit in a Python worker.

Answer

Merge the LoRA weights into the base model and serve the merged Hugging Face checkpoint with the standard transformers pipeline.

Question 11

A developer is building a generative AI application that uses an NVIDIA NIM microservice for a Llama 3 model. They need to persist the model's responses and associated metadata for later auditing. Which approach best integrates NIM with an external datastore?

Accepted Answer

Call the NIM's inference endpoint from the application code, then write the response and metadata to the datastore using the application's own database client.. NIM microservices are designed to be stateless inference endpoints. The application that calls the NIM API is responsible for handling the response and any associated metadata. To persist this information, the application should use its own database client to write to the external datastore. This separation of concerns ensures scalability and allows the developer to choose the appropriate datastore and schema.

Answer

Use the NIM's built-in PostgreSQL connector by setting the NIM_PG_CONNECTION environment variable.

Answer

Mount a shared volume into the NIM container and have the NIM write response files that the application later reads.

Answer

Enable the NIM's audit logging feature and configure it to forward logs directly to the datastore.

Question 12

A developer is building a customer support chatbot using NVIDIA NIM microservices. The application must reliably return structured JSON containing 'intent' and 'confidence' fields for downstream ticket routing. Which approach should the developer use to constrain the model's output format?

Accepted Answer

Use NIM's guided decoding with a JSON schema supplied in the request.. Reliable structured output in NIM comes from guided decoding against a JSON schema, which constrains the decoder so every response validates. Temperature, token limits, and prompt wording influence style or length but cannot guarantee the required 'intent' and 'confidence' fields, making them unsuitable for automated ticket routing.

Answer

Set the 'temperature' parameter to 0 in the NIM API request.

Answer

Add a system prompt instructing the model to 'always respond in JSON'.

Answer

Increase 'max_tokens' so the model has room to complete the JSON.

Question 13

A researcher is fine-tuning a large language model on a downstream task with a small dataset. They notice that the model achieves high training accuracy but poor validation accuracy. Which regularization technique is most appropriate to address this issue?

Accepted Answer

Apply dropout to the transformer layers during fine-tuning.. Overfitting on a small dataset is best addressed by regularization methods like dropout, which introduce noise during training and force the model to learn more generalizable patterns. Dropout is particularly effective in transformers and is easy to apply during fine-tuning. Increasing learning rate, reducing data, or removing early stopping would all worsen the problem.

Answer

Increase the learning rate to speed up convergence.

Answer

Remove early stopping and train for more epochs.

Answer

Reduce the size of the training dataset further.

Question 14

A developer is using the NVIDIA Triton Inference Server to deploy a TensorRT-LLM optimized model. They want to send a request with multiple prompts to be processed in a single inference call. Which Triton feature should they use?

Accepted Answer

Request batching. Triton Inference Server allows clients to send a single request containing multiple input elements, which are then processed together as a batch. This is often referred to as request batching or client-side batching. For TensorRT-LLM models, the model's input tensor can have a batch dimension, so the client can populate it with multiple prompts. This reduces the number of network round trips and can improve GPU utilization.

Answer

Ensemble models

Answer

Sequence batching

Answer

Dynamic batching

Question 15

A team is building an internal document assistant and wants the model to answer only from an approved corpus of HR policy PDFs. They will deploy the model with NVIDIA NIM and control grounding at generation time by injecting retrieved passages into the prompt. Which parameter combination in the NIM chat completions request best enforces this grounding while keeping responses deterministic for audit logs?

Accepted Answer

Set temperature to 0 and provide the retrieved passages inside the system and user messages as the only context, instructing the model to refuse when the answer is not present.. Grounding in a NIM-hosted model is achieved by supplying the approved passages as prompt context and constraining behavior through the system message, while temperature zero gives repeatable outputs for audit. The other choices either increase randomness, assume server-side PDF parsing that does not exist in the chat completions API, or apply sampling penalties that do not limit the model to the HR corpus.

Answer

Set temperature to 1.5 and rely on the model's pretrained HR knowledge, then post-filter answers with a regex for policy numbers.

Answer

Set top_p to 0.1 and pass the PDFs as base64 file attachments in the request body so NIM parses them automatically.

Answer

Set presence_penalty to 2.0 and include only the document titles so the model is discouraged from inventing policy details.