# Free NCA-GENL practice questions

> 10 free NCA-GENL practice questions with the correct answer and an explanation.

Page: https://devopscodelab.com/exams/nvidia-nca-genl/free-practice-questions/
Practice exams: https://devopscodelab.com/exams/nvidia-nca-genl/

### Question 1 of 10

A team notices that their transformer, trained on sequences of 2,048 tokens, degrades sharply when given 8,192-token prompts. They want a positional scheme that encodes relative position by rotating query and key vectors and that is widely used to extend context length with scaling tricks. Which scheme fits?

- A. Learned absolute position embeddings added to the token embeddings
- B. Sinusoidal absolute encodings added once at the input layer
- C. Rotary position embeddings (RoPE)
- D. Segment embeddings that mark sentence A and sentence B

**Answer:** C

**Explanation:** RoPE rotates query and key vectors by position-dependent angles so attention scores depend on relative distance, and techniques such as position interpolation build on it to stretch context. Learned absolute embeddings cannot represent positions beyond the trained table, sinusoidal encodings are absolute and extrapolate poorly in practice, and segment embeddings only distinguish sentence pairs in BERT-style inputs.

### Question 2 of 10

During fine-tuning, training loss keeps falling while validation loss starts rising after the second epoch. Which action is the MOST appropriate first response?

- A. Stop at the checkpoint with the lowest validation loss and consider stronger regularization
- B. Raise the learning rate so the optimizer escapes the region where validation loss is rising
- C. Remove the validation set and train for more epochs until training loss reaches zero
- D. Increase the model size, because a larger network has more capacity and will close the gap between the two curves

**Answer:** A

**Explanation:** Diverging training and validation loss is the classic sign of overfitting, so early stopping at the best validation checkpoint plus regularization (dropout, weight decay, more data) is the standard response. A higher learning rate does not cure memorization, dropping validation removes the signal you need, and a larger model usually overfits a small dataset even faster.

### Question 3 of 10

A team pretrains a decoder model and observes that most of its attention heads in deep layers focus heavily on the first token of every sequence regardless of content. Which explanation is MOST widely accepted?

- A. The first token acts as an attention sink where heads park weight when nothing else is relevant
- B. The causal mask is broken, so every later token is forced to look only at the first position
- C. Positional encodings give position zero an infinite score that the softmax cannot normalize
- D. The first token always carries the most semantic information about the entire input sequence that follows

**Answer:** A

**Explanation:** Because softmax must distribute probability, heads with nothing useful to attend to often dump weight on an early token, a phenomenon called an attention sink that streaming methods preserve deliberately. A broken mask would cause other errors, the first token is not inherently the most informative, and positional encodings do not produce infinite scores.

### Question 4 of 10

A company wants to customize a foundation model with LoRA through a managed microservice API and then evaluate the result with standard benchmarks and LLM-as-a-judge, all on its own Kubernetes cluster. Which NVIDIA components are designed for these two tasks?

- A. NeMo Curator for tuning and NeMo Guardrails for evaluation
- B. NeMo Customizer for tuning and NeMo Evaluator for evaluation
- C. NeMo Retriever for tuning and RAPIDS cuML for evaluation
- D. TensorRT-LLM for tuning and Triton metrics for evaluation

**Answer:** B

**Explanation:** NeMo Customizer exposes fine-tuning jobs such as LoRA as a microservice, and NeMo Evaluator runs academic benchmarks, custom metrics and LLM-as-a-judge evaluations. Curator prepares data and Guardrails controls runtime behavior, TensorRT-LLM optimizes inference, Triton metrics report serving statistics, NeMo Retriever handles embedding and reranking, and cuML is a classical ML library.

### Question 5 of 10

A team runs Hugging Face Transformers with PyTorch and trains a 13B model on 8 GPUs. They want each GPU to hold only a shard of parameters, gradients and optimizer states with minimal code changes. Which approach fits?

- A. PyTorch DistributedDataParallel with a full model replica on every GPU
- B. PyTorch Fully Sharded Data Parallel (FSDP) or DeepSpeed ZeRO stage 3
- C. Running eight independent training jobs and averaging the final weights
- D. Single-GPU training with gradient accumulation over many micro-batches

**Answer:** B

**Explanation:** FSDP and ZeRO stage 3 shard parameters, gradients and optimizer states across data-parallel ranks, cutting per-GPU memory. DDP replicates the full model, single-GPU accumulation cannot fit a 13B model's training state, and averaging independently trained models is not equivalent to distributed training.

### Question 6 of 10

A team compiles a 70B model with TensorRT-LLM for multi-GPU serving and must cut per-token latency on a single node with eight GPUs connected by NVLink. Which parallelism choice is MOST suitable for latency?

- A. Tensor parallelism across the eight GPUs
- B. Pipeline parallelism with eight stages, one per GPU
- C. Data parallelism with a full model replica per GPU
- D. Expert parallelism, since every 70B model uses experts

**Answer:** A

**Explanation:** Tensor parallelism splits each layer's matrix operations across GPUs so every token's computation is shared, which lowers latency when fast NVLink interconnect is available. Pipeline parallelism adds stage-to-stage bubbles that favor throughput over latency, a full 70B replica in FP16 does not fit on a single typical GPU, and expert parallelism applies only to mixture-of-experts models.

### Question 7 of 10

A team runs an ablation on a RAG pipeline: removing the reranker drops answer accuracy by 1 point but cuts latency by 40 percent. The product requirement is a p95 latency budget the current pipeline exceeds. What is the BEST next experiment?

- A. Rerank fewer candidates or use a smaller reranker, then measure both accuracy and p95 latency
- B. Keep the reranker unchanged, because accuracy always outweighs latency requirements
- C. Report the ablation without changes, because a 1-point drop proves rerankers are never useful
- D. Remove both the reranker and the retriever, since latency improved when components were removed

**Answer:** A

**Explanation:** Ablations reveal trade-offs; a middle option such as reranking fewer candidates or a lighter model can meet the latency budget while keeping most of the accuracy, measured on both metrics. Ignoring the stated latency requirement, removing retrieval entirely, or drawing a universal conclusion from one ablation are all poor experimental decisions.

### Question 8 of 10

A summarization model is compared against human-written reference summaries, and the team wants a metric based on n-gram overlap that emphasizes recall of reference content. Which metric is MOST appropriate?

- A. Exact match
- B. ROUGE
- C. Perplexity
- D. BLEU

**Answer:** B

**Explanation:** ROUGE measures recall-oriented n-gram and longest-common-subsequence overlap with references and is the standard for summarization. BLEU is precision-oriented and designed for translation, perplexity measures how well a model predicts text rather than comparing to references, and exact match is too strict for free-form summaries.

### Question 9 of 10

A pretraining corpus built from web crawls contains many near-identical pages that differ only in boilerplate. Which NeMo Curator capability targets this problem efficiently on GPUs?

- A. PII redaction that masks names and email addresses
- B. Exact deduplication by hashing each full document
- C. Language identification with a fastText classifier
- D. Fuzzy deduplication using MinHash and locality-sensitive hashing

**Answer:** D

**Explanation:** Fuzzy deduplication computes MinHash signatures and uses LSH buckets to find documents that are similar but not identical. Exact hashing misses pages that differ by a few characters, language identification sorts documents by language, and PII redaction protects privacy rather than removing duplicates.

### Question 10 of 10

A RAG assistant sometimes states facts that are not in any retrieved document. Which mitigation MOST directly addresses this?

- A. Use a larger chunk size so fewer documents are retrieved
- B. Remove the system prompt so the model has fewer constraints
- C. Require citations and add a grounding check on the output
- D. Increase temperature so the model explores more phrasings

**Answer:** C

**Explanation:** Asking for citations and verifying claims against retrieved context (for example with a fact-checking output rail) reduces ungrounded statements. Higher temperature tends to increase hallucination, removing instructions weakens control, and chunk size alone does not stop the model inventing facts.

Ready for a timed run? Take the free 15-question NCA-GENL mini exam: https://devopscodelab.com/app/exams/nvidia-nca-genl/mini/

DevOpsCode Lab is not affiliated with NVIDIA. NCA-GENL and related names are trademarks of their owners. We write every question ourselves; these are not real exam questions.
