NVIDIA NCA Gen AI LLMs practice questions and exam guide

Cosine learning-rate decay is used in LLM training chiefly to:

1. Smoothly reduce the step size so late training refines rather than overshoots
2. Raise the step size as the loss surface flattens near convergence
3. Keep the step size constant for reproducibility across runs
4. Reset the optimizer state at fixed intervals

Answer: A — Smoothly reduce the step size so late training refines rather than overshoots

A) Correct — a gradually annealed step size lets the model settle into a good minimum instead of bouncing around it. C) A cosine schedule is by definition not constant. B) Increasing the rate late in training is the reverse of the intent. D) Periodic optimizer resets describe warm restarts, a distinct variant.

A team wants early stopping on a fine-tuning run. The metric they should monitor is:

1. Training loss, since it most directly reflects optimization progress each epoch
2. Test-set loss, since it estimates true generalization
3. Validation loss, since it estimates generalization without touching test data
4. Gradient norm, since it signals convergence

Answer: C — Validation loss, since it estimates generalization without touching test data

C) Correct — the validation split exists precisely to make stopping decisions without contaminating the final held-out estimate. A) Training loss keeps falling during overfitting and never signals the stopping point. B) Using test loss to choose the stopping epoch leaks the test set and inflates the reported score. D) Gradient norm can shrink for reasons unrelated to generalization.

Which statement about the bias-variance tradeoff is MOST accurate for modern over-parameterized networks?

1. Test error always rises monotonically once a model can interpolate the training set
2. Variance is irrelevant when the parameter count exceeds the sample count
3. Test error can fall again past the interpolation point, a double-descent pattern
4. Bias and variance are entirely independent of model capacity

Answer: C — Test error can fall again past the interpolation point, a double-descent pattern

C) Correct — empirical double descent shows error peaking near the interpolation threshold and then declining as capacity grows further. A) The monotone-rise view is the classical U-curve intuition that over-parameterized models contradict. B) Variance still exists and still matters; it simply behaves non-classically. D) Capacity is precisely what trades one against the other.

Raising the sampling temperature above 1.0 has which effect on the next-token distribution?

1. It truncates the tail below a probability floor
2. It shifts probability mass toward the single most likely candidate token
3. It flattens the distribution, raising the chance of lower-ranked tokens
4. It leaves relative rankings unchanged but rescales the loss

Answer: C — It flattens the distribution, raising the chance of lower-ranked tokens

C) Correct — dividing logits by a temperature above one reduces their spread, making the softmax more uniform. A) Truncation is what top-k and top-p do; temperature reshapes rather than truncates. B) That is the effect of temperature below one. D) Rankings are preserved, but it is the sampling distribution, not the loss, that changes.

Why does increasing context length increase memory consumption super-linearly without optimizations?

1. The tokenizer vocabulary grows with context length
2. Position encodings store the full attention matrix
3. Standard self-attention is O(n^2) in the sequence length
4. Layer normalization scales with the square of depth

Answer: C — Standard self-attention is O(n^2) in the sequence length

A) Vocabulary size is fixed. C) Correct — vanilla attention is quadratic in memory and compute with respect to sequence length. B) Position encodings are small and linear. D) LayerNorm is per-token and linear in width.

BERT-style models use which transformer variant?

1. Encoder-only stack, bidirectional over the full sequence
2. Decoder-only stack with causal masking for generation
3. Encoder-decoder stack with cross-attention for seq2seq
4. Mixture-of-experts routing across many sparse expert FFNs

Answer: A — Encoder-only stack, bidirectional over the full sequence

B) GPT-style. A) Correct — BERT uses an encoder only with bidirectional self-attention, best for classification and embeddings. C) Seq2seq. D) A routing technique, not a directional variant.

A developer wants the fastest path from a model name on the Hugging Face Hub to a working sentiment classifier in three lines of Python. The MOST appropriate Transformers API is:

1. The pipeline() factory, which bundles tokenizer and model
2. AutoModel.from_pretrained() followed by manual logit decoding
3. The Trainer class configured with an evaluation-only TrainingArguments
4. A raw torch.load() of the checkpoint plus a hand-written forward pass

Answer: A — The pipeline() factory, which bundles tokenizer and model

A) Correct — pipeline() is the high-level abstraction that wires tokenization, model inference and label mapping together for a task. B) Works but is the lower-level path: you must map logits to labels yourself, so it is not the fastest. C) Trainer is for training/evaluation loops over datasets, not ad-hoc inference. D) torch.load() bypasses the Hub config and tokenizer entirely and is the most manual option.

The PRIMARY job of a collate function in a PyTorch DataLoader for LLM fine-tuning is to:

1. Shuffle the dataset order before each training epoch begins
2. Pad variable-length examples into a batched, masked tensor
3. Compute the loss over the batch after the forward pass
4. Convert raw text into token ids using the tokenizer

Answer: B — Pad variable-length examples into a batched, masked tensor

B) Correct — collation turns a list of per-example dicts into padded tensors with the matching attention mask and label mask. A) Shuffling is the sampler's responsibility. C) Loss is computed by the model or training loop. D) Tokenization normally happens in the dataset map step before collation.

Which practice BEST supports experiment reproducibility during fine-tuning?

1. Record the random seed, data snapshot hash and full config per run
2. Save only the final checkpoint and the best validation score
3. Reuse one shared output directory for every experiment
4. Keep hyperparameters in a notebook cell edited between runs

Answer: A — Record the random seed, data snapshot hash and full config per run

A) Correct — seed, data version and config together let a run be recreated; any one alone is insufficient. B) A score without the recipe cannot be reproduced. C) A shared directory overwrites artifacts and destroys the audit trail. D) Ad-hoc notebook edits leave no record of what produced which result.

Which NVIDIA component is the end-to-end framework for training, customizing, and aligning generative AI models including Megatron-LM and curator tooling?

1. Triton Inference Server
2. NeMo Framework
3. TensorRT-LLM
4. Riva

Answer: B — NeMo Framework

A) Serving. B) Correct — NeMo covers data, pretraining, customization, alignment. C) Inference optimizer. D) Speech.

Which is the BEST high-level architectural separation between NIM, TensorRT-LLM, and Triton?

1. NIM is the packaged microservice, Triton the server, TensorRT-LLM the compiler
2. All three are interchangeable and perform exactly the same function in a stack
3. NIM handles training, TensorRT-LLM handles serving, and Triton handles data prep
4. Triton performs vector search, TensorRT-LLM storage, and NIM business analytics

Answer: A — NIM is the packaged microservice, Triton the server, TensorRT-LLM the compiler

A) Correct — NIM productizes a microservice commonly built over Triton and TensorRT-LLM, which are the server and the LLM inference compiler respectively. B) The three sit at different layers. C) Scrambles the roles. D) None of those functions belong to these components.

Which GPU is positioned by NVIDIA as a cost-effective L40S-class option for enterprise inference workloads with strong FP8 support?

1. L40S
2. Jetson Orin Nano
3. GTX 1080
4. DGX Station

Answer: A — L40S

A) Correct — L40S targets generative AI inference and graphics at scale. B) Edge. C) Consumer. D) System, not a GPU.

LoRA (Low-Rank Adaptation) does which?

1. Updates all base model weights during fine-tuning
2. Distills a large teacher model into a smaller student
3. Learns a new subword tokenizer for the target domain
4. Trains small low-rank matrices inside frozen layers

Answer: D — Trains small low-rank matrices inside frozen layers

A) Full fine-tuning is the alternative LoRA avoids. D) Correct — LoRA freezes the base model and trains low-rank decomposition matrices inserted into attention/FFN layers, typically about 0.1–1% of parameters. Lightweight and composable. B/C) Different techniques.

KTO (Kahneman-Tversky Optimization) is best described as a preference-tuning method that:

1. Requires strictly paired chosen/rejected preference data
2. Replaces the language modeling head with a regression head
3. Uses unpaired good/bad labels via a prospect-theory loss
4. Trains only the embedding layer and freezes the rest

Answer: C — Uses unpaired good/bad labels via a prospect-theory loss

C) KTO applies a prospect-theory-inspired utility loss to unpaired desirable/undesirable examples. A) Pairing is exactly the requirement KTO drops. B) The LM head is retained. D) KTO is not restricted to the embedding layer.

Catastrophic forgetting during fine-tuning is best mitigated by:

1. Using a very high learning rate on a narrow dataset
2. Skipping evaluation entirely and shipping the model fast
3. Mixing in general data, using PEFT and low learning rates
4. Training for a single epoch on a very small dataset

Answer: C — Mixing in general data, using PEFT and low learning rates

C) Replaying general-domain data, constraining the update with PEFT or small learning rates, and regularizing toward the base model all limit drift. A) A large learning rate on narrow data accelerates forgetting. B) Skipping evaluation hides the problem rather than mitigating it. D) Shrinking the run limits learning without protecting prior capabilities.

Which practice MOST clearly constitutes data leakage in an ML experiment?

1. Applying the same random seed to all training runs
2. Reporting both the validation and the test metrics in the final writeup
3. Fitting the feature scaler on the full dataset before splitting
4. Using stratified sampling when creating the data splits

Answer: C — Fitting the feature scaler on the full dataset before splitting

C) Correct — scaler statistics computed over all rows carry test-set information into training, inflating measured performance. A) A shared seed aids reproducibility and leaks nothing. B) Reporting both sets is transparent as long as decisions used validation only. D) Stratification preserves class balance and is standard good practice.

Two models report perplexity of 12.4 and 15.8 on different corpora tokenized with different tokenizers. The MOST defensible conclusion is:

1. The first model generalizes better across all domains
2. The second model has 27% more hallucinations
3. The comparison is invalid without a shared corpus and tokenizer
4. Both models are equivalent, since the observed gap is under five points

Answer: C — The comparison is invalid without a shared corpus and tokenizer

C) Correct — perplexity depends on the evaluation text and the token segmentation, so cross-setup numbers are not comparable. A) Different corpora make any generalization claim unsupported. B) Perplexity carries no information about hallucination rate. D) Declaring equivalence from an arbitrary threshold on incomparable numbers is unfounded.

Which practice contributes MOST to reproducibility of an LLM fine-tuning experiment?

1. Using the largest available batch size for stability
2. Saving only the final model checkpoint to shared team storage
3. Documenting the final evaluation score on the project wiki page
4. Recording seeds, library versions, data snapshot, and full config

Answer: D — Recording seeds, library versions, data snapshot, and full config

D) Correct — the full provenance of code, data, and configuration is what lets another team regenerate the result. B) A checkpoint alone cannot be re-derived or audited. C) A recorded score is a claim, not a reproduction path. A) Batch size is one hyperparameter and does not by itself confer reproducibility.

Which vector database is best described as an open-source, cloud-native, columnar vector database originally developed by Zilliz?

1. Pinecone
2. Qdrant
3. Milvus
4. Weaviate

Answer: C — Milvus

A) Pinecone is closed-source SaaS. B) Qdrant is Rust-based open source. C) Correct — Milvus is the Zilliz-originated OSS vector DB. D) Weaviate is a different OSS vector DB.

NeMo Curator is purpose-built for:

1. Online model serving with autoscaling inference endpoints
2. Large-scale data curation: dedup, filtering, and scoring
3. Vector database management for embedding storage and search
4. Voice synthesis and speech-to-text for conversational agents

Answer: B — Large-scale data curation: dedup, filtering, and scoring

B) Correct — NeMo Curator is GPU-accelerated pipeline tooling for preparing pretraining corpora. A) Serving is Triton and NIM. C) Retrieval storage is a separate layer. D) Speech is Riva.

Byte Pair Encoding (BPE) tokenization works by:

1. Splitting text on whitespace and punctuation boundaries only
2. Mapping every Unicode codepoint to its own unique token id
3. Merging the most frequent adjacent symbol pair repeatedly
4. Assigning one token per word from a fixed word dictionary

Answer: C — Merging the most frequent adjacent symbol pair repeatedly

A) That is a baseline tokenizer, not BPE. B) That is byte- or character-level coding only. C) Correct — BPE greedily merges the most frequent adjacent pair over and over to grow a subword vocabulary. D) Word-level tokenization is brittle to out-of-vocabulary terms.

Red-teaming an LLM application typically involves:

1. Running only automated unit tests against the application code
2. Adversarial probing for injections, jailbreaks, and data leaks
3. Recoloring and restyling the chat interface before it launches
4. Replacing the deployed model with a smaller, cheaper variant

Answer: B — Adversarial probing for injections, jailbreaks, and data leaks

B) Correct — manual and automated probing surfaces injections, jailbreaks, privacy leaks, and tool misuse before release. A) Unit tests do not explore adversarial behavior. C) Interface styling is cosmetic. D) Swapping models is a cost decision, not a security exercise.

Which statement about LLM safety and capability tradeoffs is most accurate?

1. Stronger alignment always drives model helpfulness down to zero
2. Safety concerns can be ignored entirely once a model is large
3. Safety measures carry no measurable cost to a model's capability
4. Safety and helpfulness are co-designed; both refusal errors matter

Answer: D — Safety and helpfulness are co-designed; both refusal errors matter

D) Correct — over-refusal and under-refusal are both real failures, so current practice measures and balances them together. A) An unsupported extreme. B) Scale alone does not deliver safety. C) There are genuine, measurable tradeoffs.

Watermarking generated text aims to:

1. Embed statistical signals in output that detectors can spot
2. Reduce inference latency by shortening the generated sequence
3. Improve perplexity scores measured on held-out evaluation text
4. Encrypt the prompt so intermediaries cannot read user input

Answer: A — Embed statistical signals in output that detectors can spot

A) Correct — watermarking biases token selection so a downstream detector can flag text as machine-generated. B) It does not target latency. C) It does not improve modeling quality. D) Encryption protects confidentiality, a different goal.

What does the temperature parameter do to a model's output distribution, and what do low and high values produce?

Temperature divides the logits before softmax. Values below 1 sharpen the distribution toward the argmax, giving deterministic, conservative text suited to extraction and code. Values above 1 flatten it, raising diversity and the chance of incoherence. Temperature approaching zero is effectively greedy decoding.

Transformer architecture — core innovation?

Self-attention mechanism. Processes sequence in parallel, captures long-range dependencies.

Rank FP32, BF16, FP16, FP8, and INT8 by memory footprint and state the main tradeoff of moving down.

FP32 uses 4 bytes per value, BF16 and FP16 use 2, FP8 and INT8 use 1. Lower precision cuts memory and bandwidth roughly proportionally and enables faster tensor-core paths, at the cost of numerical range or resolution. BF16 keeps FP32's exponent range so it rarely overflows in training, while FP16 has more mantissa bits but a narrow range that often needs loss scaling.

What do neural scaling laws say about the relationship between model loss, parameters, data, and compute?

Test loss falls as a smooth power law in model parameters, training tokens, and compute, provided the other factors are not bottlenecks. The practical implication is compute-optimal training: for a fixed compute budget there is a joint parameter/token allocation that beats making the model larger while starving it of data, and returns are predictable rather than sudden.

You change the base model, the prompt, and the retrieval chunk size at once and quality improves. What is wrong and what should you do?

You have confounded three factors, so you cannot attribute the gain, cannot revert the harmful one, and cannot tell whether two changes partially cancelled. Run an ablation: change one factor at a time against a fixed evaluation set with fixed seeds and decoding parameters, report the effect of each, and only then combine. Also check that the improvement exceeds run-to-run variance before calling it real.

What are the format facts for the NCA-GENL exam: question count, time limit, and how long the credential stays valid?

50-60 multiple-choice questions (single- and multi-select) in 60 minutes, delivered online proctored with no prerequisites. The certification is valid for 2 years.