Data Analysis and Visualization — NVIDIA NCA Gen AI LLMs practice questions

WordPiece differs from classic BPE MOST notably in that WordPiece:

1. Operates strictly on whitespace-delimited whole words only
2. Requires a fixed vocabulary of exactly 30,000 entries
3. Selects merges by likelihood gain rather than raw pair frequency
4. Cannot represent characters outside the training corpus at all

Answer: C — Selects merges by likelihood gain rather than raw pair frequency

C) Correct — WordPiece picks the merge that most improves corpus likelihood, whereas BPE greedily merges the most frequent pair. A) Both algorithms produce subword units below the word level. B) 30k is a common choice, not a requirement of the algorithm. D) Subword schemes fall back to characters or byte units for unseen input.

HNSW is BEST described as:

1. A gradient-based training algorithm for encoder models
2. A graph-based approximate nearest-neighbor index
3. A subword tokenizer applied before embedding text
4. An embedding model producing dense semantic vectors

Answer: B — A graph-based approximate nearest-neighbor index

A) Not a training method. B) Correct — HNSW (Hierarchical Navigable Small World) builds a navigable multi-layer graph for sub-linear approximate nearest-neighbor search. C/D) Different components of the pipeline.

A retrieval system stores chunk metadata alongside vectors. Which feature-engineering choice MOST improves filtered retrieval quality?

1. Storing the raw chunk text a second time inside the vector payload
2. Replacing the document title with its numeric hash value
3. Adding normalized fields such as document date and access tier
4. Concatenating every metadata field into one free-text blob

Answer: C — Adding normalized fields such as document date and access tier

C) Correct — clean, typed attributes let the query engine apply pre-filters that cut the candidate set to what is relevant and permitted. A) Duplicating text adds storage without adding filterable structure. B) A hash destroys the semantics needed for filtering or display. D) A single blob defeats structured filtering entirely.

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.

NeMo Curator is purpose-built to:

1. Curate and clean large text corpora at scale for LLM training
2. Route inference requests across a fleet of GPU replicas
3. Convert PyTorch checkpoints into TensorRT engines
4. Track experiment metrics and model lineage for audits

Answer: A — Curate and clean large text corpora at scale for LLM training

A) Correct — Curator handles GPU-accelerated download, filtering, deduplication and cleaning of pretraining-scale text. B) Request routing is a serving concern. C) Engine building is TensorRT-LLM's role. D) Experiment tracking is a separate MLOps function.

Which scenario suggests RAG rather than fine-tuning is the better first lever?

1. The model must emit a new structured output format
2. The model must internalize a distinct writing style
3. The model needs a new persona and consistent tone
4. Answers must cite a knowledge base that updates daily

Answer: D — Answers must cite a knowledge base that updates daily

D) Volatile, citable facts are what retrieval handles well, since the index can be refreshed without touching weights. A) Consistent output formatting is learned efficiently by SFT. B) Style is a weights-level behavior. C) Persona and tone likewise belong in tuning or the system prompt.

Enlarging the tokenizer vocabulary from 32k to 128k for a fixed model size MOST directly trades:

1. Slower attention against a smaller KV cache per token
2. Lower training throughput against improved numerical stability
3. A larger embedding matrix against shorter tokenized sequences
4. Reduced OOV risk against a mandatory increase in model depth

Answer: C — A larger embedding matrix against shorter tokenized sequences

C) Correct — more tokens means bigger input and output embedding tables but fewer tokens per document, easing context pressure. A) Attention cost depends on sequence length, and per-token KV size is unaffected by vocabulary. B) Vocabulary size is not a numerical-stability lever. D) Depth is an independent architectural choice.

Choosing cosine similarity over dot product for retrieval is most appropriate when:

1. Embedding magnitudes vary and direction carries meaning
2. You want scoring to reward larger vector magnitudes
3. You require strict Euclidean-distance geometry guarantees
4. You need exact integer arithmetic inside the index

Answer: A — Embedding magnitudes vary and direction carries meaning

A) Cosine normalizes length away, so only the angle between vectors contributes to the score. B) That is the case for keeping dot product. C) Metric guarantees point to L2 distance instead. D) Integer arithmetic is a quantization concern, unrelated to metric choice.

Agentic RAG differs from naive RAG primarily by:

1. Dropping the LLM and returning only raw retrieval hits
2. Using a single fixed prompt template with no tool use
3. Always retrieving exactly one document for each query
4. Letting the LLM plan queries, tools, and retrieval rounds

Answer: D — Letting the LLM plan queries, tools, and retrieval rounds

D) Agentic RAG puts the model in control of planning, tool selection, query reformulation, and repeated retrieval with self-checking before it answers. A) Removing generation defeats the purpose. B) A fixed single-shot template is naive RAG. C) A fixed top-1 fetch is a further simplification.

Cosine similarity is preferred over raw Euclidean distance for text embeddings MAINLY because it:

1. Compares direction and is insensitive to vector magnitude
2. Is always faster to compute on GPU than a distance metric
3. Produces values bounded between zero and one for any vectors
4. Requires no normalization step at any point in the pipeline

Answer: A — Compares direction and is insensitive to vector magnitude

A) Correct — semantic content is carried by direction, and magnitude often reflects length or frequency artifacts. B) Both are cheap dot-product-family operations. C) Cosine ranges from -1 to 1 in general. D) Normalizing is the usual way cosine is implemented efficiently.

HNSW (Hierarchical Navigable Small World)?

Graph-based ANN: build navigable hierarchy of neighbors. Fast, accurate, memory-heavy.

Why do modern LLMs use subword tokens instead of words or characters?

A word vocabulary cannot cover unseen words, misspellings, or code and explodes in size. Character tokens have a tiny vocabulary but make sequences very long and force the model to relearn spelling. Subwords keep the vocabulary bounded, represent any string without out-of-vocabulary failures, and keep sequences short enough for attention.

Token vs Subword vs Word tokenization?

Token: smallest unit (often subword). Subword (BPE, SentencePiece): balance vocab size + coverage. Word: brittle to OOV.

Why RAG instead of fine-tuning?

Dynamic data, citation/auditability, smaller context per query, cheaper to update than re-training.