NCP-GENL Practice Test 15 — 15 Questions

Question 1

An engineer is building a customer-facing FAQ bot using an NVIDIA NIM-hosted Llama 3.1 70B model. The bot must answer ONLY from a supplied product knowledge base and must respond with 'I don't have that information' when the answer is not present. Which prompt engineering approach BEST enforces this constraint?

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

Accepted Answer

Instruct the model to answer only from the provided knowledge base, explicitly state that it must reply with the refusal phrase when the answer is absent, and include the knowledge base inside clearly delimited sections.. Grounded FAQ bots need an explicit behavioral contract: what source to use, what to do when the source lacks the answer, and clear separation of context from user input. Delimiters prevent prompt injection from blending with instructions, and a fixed refusal phrase makes the fallback deterministic and testable. Sampling parameters alone cannot enforce scope constraints.

Answer

Set the top_p value to 0.1 and rely on the model's pretrained knowledge of the product domain instead of supplying the knowledge base in the prompt.

Answer

Append the entire knowledge base to every user message without any instruction about scope or refusal behavior, letting the model infer the rules from context.

Answer

Increase the temperature to 1.0 and add 'Be creative and helpful' to the system prompt so the model can improvise when the knowledge base is incomplete.

Question 2

An engineer is optimizing prompts for an NVIDIA NIM-hosted model used in a multi-turn technical troubleshooting chat. The model forgets earlier constraints, such as the customer's environment and the product version, as the conversation grows. Which prompt engineering technique best preserves these constraints across turns?

Accepted Answer

Maintain a compact running summary of key constraints and inject it into the system prompt each turn.. Multi-turn memory is best preserved by extracting and re-injecting critical constraints rather than replaying all dialogue. A compact running summary placed in the system prompt keeps key facts salient and stable while avoiding context bloat. Temperature and single-turn designs do not address memory, and verbatim history can dilute attention and waste tokens.

Answer

Repeat the full conversation history verbatim in every turn and increase max_tokens.

Answer

Use a single-turn prompt for each message and rely on the model's pretrained knowledge of the product.

Answer

Lower the temperature to 0 and remove the system prompt to avoid conflicting instructions.

Question 3

A data science team is preparing an instruction fine-tuning dataset in NVIDIA NeMo Framework. They notice that after training, the model performs well on the training instructions but poorly on paraphrased versions of the same instructions. They want to improve generalization without increasing dataset size. Which data preparation change is most appropriate?

Accepted Answer

Apply instruction paraphrasing and template diversification so the same intent appears in multiple surface forms during training.. Generalization to paraphrased instructions depends on the model seeing the same intent expressed in varied surface forms. Paraphrasing and template diversification teach the model to respond to meaning rather than exact wording, directly improving performance on unseen paraphrases without enlarging the dataset or changing training hyperparameters.

Answer

Remove all examples where the instruction phrasing is unique, keeping only the most common templates.

Answer

Increase the number of epochs so the model sees each instruction more times and memorizes the exact phrasing.

Answer

Lower the learning rate so the model updates more slowly and avoids overfitting to specific instruction phrasings.

Question 4

An engineer is profiling a CUDA kernel and notices that the achieved occupancy is low, leading to underutilization of the GPU. The kernel uses a large number of registers per thread, limiting the number of resident warps. Which optimization should be attempted first to improve occupancy?

Accepted Answer

Use the __launch_bounds__ qualifier to limit registers per thread.. The low occupancy is caused by high register usage per thread. Using __launch_bounds__ instructs the compiler to limit registers, allowing more warps to be resident and improving occupancy. This directly addresses the bottleneck, whereas other options do not target register pressure and thus would not effectively improve occupancy.

Answer

Enable L1 cache to reduce memory latency.

Answer

Increase the block size to allow more warps per block.

Answer

Convert the kernel to use shared memory for data reuse.

Question 5

An engineer is using TensorRT-LLM to serve a model that occasionally receives prompts far longer than the typical 512 tokens, up to 8K tokens. With the default engine settings, requests near 8K fail with a cache capacity error while short requests succeed. Which configuration change most directly resolves this without rebuilding for a single worst-case shape?

Accepted Answer

Set the maximum input length and maximum sequence length to accommodate 8K tokens and size the KV cache pool accordingly.. The failure is a capacity limit tied to the engine's configured maximum input and sequence lengths and the KV cache pool sized for them. Raising those limits to cover 8K tokens and sizing the pool to match lets long prompts reserve the blocks they need, while paged allocation means short requests still consume only a few blocks, so a single worst-case build shape is unnecessary.

Answer

Enable INT8 KV cache quantization so each token's cached keys and values consume fewer bytes.

Answer

Enable chunked context prefill so long prompts are processed in multiple smaller prefill passes.

Answer

Reduce the maximum batch size to 1 so all KV cache blocks are available to a single long request.

Question 6

An engineer is optimizing an LLM for inference with NVIDIA TensorRT-LLM and wants to reduce both memory footprint and latency without retraining the model. Which two techniques should they apply? (Choose two.)

Accepted Answer

Quantize model weights to INT8 or FP8 using post-training quantization with calibration.. Post-training quantization reduces weight memory and enables faster low-precision kernels, while paged KV cache manages generation-time memory efficiently and improves batching. Together they lower footprint and latency without retraining. The rejected options either require retraining, increase model size, or break the architecture, so they do not fit the no-retraining constraint or the optimization goals.

Answer

Disable attention and rely solely on feed-forward layers to speed up decoding.

Answer

Increase the number of transformer layers to improve parallelism across the GPU.

Answer

Retrain the model with quantization-aware training to learn optimal scaling factors.

Question 7

A team fine-tunes an NVIDIA NeMo model to classify support tickets into five categories. In production, the model sometimes outputs free-form explanations instead of a single category label, breaking the downstream parser. Which prompt engineering change MOST reliably constrains the output format?

Accepted Answer

Add three few-shot examples that each end with the exact category label, and instruct the model to output only the label with no additional text.. Strict output contracts are best enforced by showing the exact desired format through few-shot examples and explicitly prohibiting any additional text. Demonstrations act as in-context conditioning that shapes the model's continuation pattern, while the instruction closes the loophole of adding commentary. Sampling parameters control randomness, not structure.

Answer

Add a polite request asking the model to 'try to keep answers short and to the point' while leaving the output format unspecified.

Answer

Increase max_tokens so the model has more room to explain its reasoning before stating the final category label.

Answer

Lower the temperature to zero and trust that deterministic sampling will force the model to emit only a category label.

Question 8

A healthcare AI team is evaluating a fine-tuned GPT-based model for clinical note summarization using NVIDIA NeMo. They need to assess whether the model's summaries contain no fabricated medical facts. Which evaluation approach is most appropriate?

Accepted Answer

Use a factual consistency metric such as SummaC or Q2. The critical requirement is detecting fabricated medical facts, which demands checking entailment between the summary and the source clinical note. Factual consistency metrics such as SummaC or Q2 are specifically designed for this purpose, unlike similarity-based metrics (BERTScore, ROUGE-L) that compare to references and may miss hallucinations. Perplexity assesses fluency, not factuality. Therefore, a factual consistency metric is the correct evaluation approach.

Answer

Measure BERTScore against reference summaries

Answer

Compute perplexity of the summaries

Answer

Calculate ROUGE-L scores

Question 9

You are responsible for the reliability of an LLM inference service running on NVIDIA Triton Inference Server across a fleet of A100 GPUs. The service is deployed with dynamic batching enabled, but during peak hours you observe that end-to-end latency for some requests exceeds the SLO while GPU utilization remains moderate. You suspect that the dynamic batching configuration is causing the issue. Which Triton configuration parameter should you adjust to directly limit the maximum time a request waits in the scheduler queue before being batched?

Accepted Answer

Set the `max_queue_delay_microseconds` parameter to a lower value.. The `max_queue_delay_microseconds` parameter in Triton's dynamic batching configuration sets the maximum time a request can wait in the scheduler queue before being dispatched for inference. Lowering this value reduces the worst-case latency introduced by batching, directly addressing the SLO violations observed during peak hours. While other parameters like `max_batch_size` and `instance_group` affect performance, only `max_queue_delay_microseconds` explicitly bounds the queue wait time, making it the correct adjustment for this scenario.

Answer

Increase the `instance_group` count to add more model instances.

Answer

Set the `max_batch_size` parameter to a lower value.

Answer

Enable the `priority_levels` parameter to prioritize certain requests.

Question 10

An engineer is using an NVIDIA NIM for a Mixtral model to extract structured data from invoices. The model occasionally returns fields with the wrong data type, such as a numeric amount as a string. The team wants a prompt engineering fix that does not require changing the model or adding a separate parser. Which approach is most effective?

Accepted Answer

Describe the desired output schema in the system prompt, including field names and expected data types, and provide one fully formatted example.. An explicit schema with field names and data types, paired with a complete example, gives the model an unambiguous template. The example shows numeric values without quotes, so the model imitates the correct types. Self-checking, sampling parameters, and natural language output do not provide the concrete structural guidance needed to eliminate type errors at the prompt level.

Answer

Instruct the model to output the data in a natural language paragraph and then extract the fields manually.

Answer

Ask the model to double-check its output for type correctness before returning it.

Answer

Lower the temperature to 0.0 and increase the repetition penalty to discourage type mistakes.

Question 11

A company is running a large language model inference service on NVIDIA GPUs. They observe that GPU memory is nearly full, limiting the batch size and thus throughput. The model weights are stored in FP16, and the KV cache consumes a significant portion of memory. Which technique can reduce memory usage while maintaining model accuracy and enabling larger batch sizes?

Accepted Answer

Quantize the KV cache to INT8.. The memory bottleneck is largely due to the KV cache. Quantizing it to INT8 halves its memory footprint, freeing space for larger batch sizes and improving throughput. This technique is supported in frameworks like TensorRT-LLM and maintains accuracy with proper calibration. Other options either do not apply to inference or increase memory usage.

Answer

Increase the number of GPUs to distribute the model.

Answer

Convert the model to FP32 to improve numerical stability.

Answer

Use gradient checkpointing during inference.

Question 12

A developer has a fine-tuned Llama-family model in Hugging Face format and wants to run it with NVIDIA TensorRT-LLM on an H100. Which artifact must be produced before the runtime can execute the model?

Accepted Answer

A TensorRT engine built from the model's network definition for the target GPU architecture.. TensorRT-LLM executes a serialized TensorRT engine that is compiled for a specific GPU architecture, precision, and parallelism configuration. After converting the Hugging Face checkpoint into the TensorRT-LLM checkpoint format, the developer must build the engine for the H100. Only that engine can be loaded by the runtime, so it is the required artifact before any inference can occur.

Answer

An ONNX graph exported with dynamic axes and a matching runtime configuration JSON.

Answer

A PyTorch TorchScript trace of the full forward pass saved as a .pt file.

Answer

A quantized GGUF file generated with a community conversion script.

Question 13

You are preparing a large instruction-tuning dataset with NVIDIA NeMo Curator. The dataset contains many near-duplicate instruction-response pairs that differ only in punctuation or minor wording. Which NeMo Curator stage should you apply to remove these near-duplicates before fine-tuning?

Accepted Answer

Fuzzy deduplication using MinHash locality-sensitive hashing.. Near-duplicate instruction-response pairs that differ only slightly require fuzzy matching, not exact hashing or heuristic quality filters. MinHash locality-sensitive hashing estimates Jaccard similarity between records and clusters near-identical ones, letting you retain a single representative. Applying fuzzy deduplication in NeMo Curator reduces redundancy, prevents the model from overfitting repeated patterns, and improves generalization on the instruction-tuning task.

Answer

Exact substring deduplication to remove overlapping text spans.

Answer

Heuristic filtering to remove short or malformed records.

Answer

Exact duplicate removal using a hash of the full instruction-response pair.

Question 14

A team is designing prompts for an NVIDIA NIM-hosted LLM that must produce concise, citation-backed answers from retrieved documents. They want to improve factual grounding and reduce unsupported claims. Which two prompt engineering practices best support this goal? (Choose two.)

Accepted Answer

Instruct the model to cite the specific document ID or snippet for each claim it makes.. Factual grounding in retrieval-augmented generation improves when the model must cite sources and when it is allowed to refuse when evidence is missing. Citations create traceability, and a refusal fallback prevents guessing. Allowing general knowledge, raising temperature, or stripping identifiers weakens provenance and increases unsupported claims.

Answer

Remove document identifiers from the context to simplify the prompt and reduce tokens.

Answer

Increase temperature to encourage the model to synthesize multiple documents creatively.

Answer

Allow the model to answer from general knowledge when retrieved documents are incomplete.

Question 15

A team is building a NeMo-based LLM pipeline and must tokenize a corpus that mixes English, Japanese, and Python source code. They plan to train a custom tokenizer with NVIDIA NeMo. Which tokenizer configuration best supports all three content types without excessive sequence length?

Accepted Answer

Byte-level Byte-Pair Encoding with a vocabulary of 64,000 to 128,000 tokens trained on a balanced sample of all three content types.. Byte-level BPE with a large vocabulary trained on a balanced multilingual and code sample handles arbitrary scripts and symbols without unknown tokens. It compresses common English words, Japanese subwords, and Python identifiers into single tokens, controlling sequence length. Training on all three distributions aligns merge statistics with actual usage, unlike English-only or Japanese-only training, and the 64k-128k range provides enough capacity for multilingual plus code coverage.

Answer

SentencePiece unigram tokenizer trained exclusively on Japanese text.

Answer

WordPiece tokenization trained only on the English portion of the corpus.

Answer

Byte-Pair Encoding with a vocabulary limited to 8,000 tokens.