NCA-GENL Practice Questions: Free Practice Test With Explanations

1. A data scientist is explaining why a deep neural network uses nonlinear activation functions (such as ReLU) between its layers. Which statement best describes the primary reason they are necessary?

   • A. They guarantee that the loss function becomes convex, ensuring a single global minimum.
   • B. They normalize the input data so that every feature has zero mean and unit variance.
   • C. They reduce the total number of parameters the network must store in GPU memory.
   • D. They let the network learn nonlinear relationships between inputs.

   Answer: D. Activation functions introduce nonlinearity, which lets a deep network approximate complex nonlinear functions; without them, any number of stacked linear layers reduces to one linear layer. They do not reduce parameter count, do not make the loss convex (deep nets are highly non-convex), and input normalization is a separate preprocessing step, not the role of activations.

   Exam domain: Core Machine Learning and AI Knowledge

2. During training of a feedforward network, an engineer observes that gradients in the earliest layers shrink toward zero as they propagate backward, slowing learning in those layers. What is this phenomenon commonly called, and which choice is a typical mitigation?

   • A. Overfitting; mitigate by increasing the number of training epochs.
   • B. Data leakage; mitigate by shuffling the validation set.
   • C. Catastrophic forgetting; mitigate by lowering the batch size.
   • D. The vanishing gradient problem; use ReLU and residual connections.

   Answer: D. When gradients shrink as they flow backward through many layers, early-layer weights barely update; this is the vanishing gradient problem. ReLU-style activations and residual/skip connections (as in ResNets and Transformers) help gradients flow. Overfitting, catastrophic forgetting, and data leakage are unrelated phenomena with different causes.

   Exam domain: Core Machine Learning and AI Knowledge

3. In the self-attention mechanism of a Transformer, each token's representation is computed using three learned projections of the input. What are these three components called?

   • A. Filter, Stride, and Padding
   • B. Gate, Cell, and Hidden state
   • C. Encoder, Decoder, and Bottleneck
   • D. Query, Key, and Value

   Answer: D. Self-attention projects each token into Query, Key, and Value vectors; attention scores come from Query-Key dot products and are used to weight the Values. Encoder/decoder describe overall architecture blocks, gate/cell/hidden belong to LSTMs, and filter/stride/padding are convolution concepts.

   Exam domain: Core Machine Learning and AI Knowledge

4. Why do Transformer models add positional encodings (or positional embeddings) to their input token embeddings?

   • A. To apply dropout selectively only to rare tokens.
   • B. To compress the vocabulary so fewer embedding rows are needed.
   • C. To convert floating-point embeddings into integer token IDs for faster lookup.
   • D. Because self-attention is order-agnostic and needs position information.

   Answer: D. Self-attention treats input as an unordered set, so without positional signals the model cannot distinguish word order; positional encodings inject that order information. They do not compress vocabulary, do not convert embeddings to IDs (the reverse of how lookup works), and are unrelated to selective dropout.

   Exam domain: Core Machine Learning and AI Knowledge

5. A developer notices that the string "unbelievable" is split into the sub-pieces "un", "believ", and "able" before being fed to an LLM. Which step of the LLM input pipeline is producing these sub-pieces?

   • A. Subword tokenization
   • B. Attention computation
   • C. Positional encoding
   • D. Embedding lookup

   Answer: A. Modern LLMs use subword tokenization (e.g., Byte-Pair Encoding or WordPiece) that breaks rare or compound words into smaller, reusable sub-pieces so the vocabulary stays manageable while still covering unseen words. Embedding lookup happens after tokenization (it converts each token ID into a vector), positional encoding adds order information, and attention is part of the transformer layers, none of which split text into sub-pieces.

   Exam domain: Core Machine Learning and AI Knowledge

6. A team feeds an LLM a 20,000-token document, but the model is configured with an 8,192-token context window. What is the most likely consequence?

   • A. The model automatically increases its context window to fit the document.
   • B. The model raises the temperature to compensate for the longer input.
   • C. Tokens beyond the context window limit are not attended to, so content past the limit is truncated or dropped.
   • D. The document is converted to a single embedding vector that preserves all information losslessly.

   Answer: C. The context window is a hard upper bound on how many tokens the model can process at once; input exceeding it is truncated (or must be chunked), so content past the limit is effectively ignored. A model cannot grow its own context window at runtime, a single embedding cannot losslessly hold 20,000 tokens, and temperature controls sampling randomness, not input length.

   Exam domain: Core Machine Learning and AI Knowledge

7. In a semantic search system, why are the words "automobile" and "car" expected to have embedding vectors that are close together in vector space?

   • A. Because they contain a similar number of characters
   • B. Because they were assigned consecutive token IDs in the vocabulary
   • C. Because embeddings encode meaning, so similar contexts map nearby
   • D. Because the tokenizer merged them into the same token

   Answer: C. Embeddings are learned representations that place semantically similar items near one another, so synonyms used in similar contexts end up close in vector space, which is what enables semantic (meaning-based) search. Character count, token-ID ordering, and tokenizer merges are surface-level or unrelated artifacts that do not capture meaning.

   Exam domain: Core Machine Learning and AI Knowledge

8. A prompt instructs an LLM to "Classify the sentiment of this review as positive or negative" and provides three labeled example reviews before the new review to classify. Which prompting technique is being used?

   • A. Zero-shot prompting
   • B. Chain-of-thought prompting
   • C. Retrieval-augmented generation
   • D. Few-shot prompting

   Answer: D. Providing several in-context labeled examples before the target input is few-shot prompting, which helps the model infer the desired format and behavior. Zero-shot gives no examples, chain-of-thought asks the model to reason step by step, and RAG injects retrieved external documents rather than demonstration examples.

   Exam domain: Software Development

9. A team is assembling a large web-scraped text corpus to use for continuous pre-training of an LLM. During an initial scan they find that many documents appear multiple times verbatim. Which preprocessing step most directly addresses this issue and why is it important?

   • A. Deduplication, since repeats cause memorization and waste compute
   • B. Tokenization, because duplicate text must be converted to token IDs before training
   • C. Stop-word removal, because duplicate documents are usually filler text
   • D. Normalization, because duplicates indicate inconsistent character encodings

   Answer: A. Deduplication removes exact or near-duplicate documents, which reduces memorization, improves generalization, and avoids spending compute repeatedly on the same content. Tokenization converts text to IDs but does not detect duplicates; normalization handles encoding/casing; stop-word removal is rarely used for modern LLM corpora and does not address repetition.

   Exam domain: Data Analysis and Visualization

10. While cleaning a multilingual text dataset, you notice the same character (for example, an accented letter) is represented with different byte sequences across documents. Which preprocessing operation best resolves this inconsistency before tokenization?

    • A. Lemmatization of every word form
    • B. Unicode normalization (NFC/NFKC)
    • C. Random shuffling of all records
    • D. One-hot encoding of each token

    Answer: B. Unicode normalization (such as NFC or NFKC) maps visually identical characters with different byte representations to a canonical form, ensuring consistent tokenization. Lemmatization reduces words to base forms (a different concern), shuffling changes order not encoding, and one-hot encoding is for categorical features, not text encoding consistency.

    Exam domain: Data Analysis and Visualization

Want more? Finish the quiz above to unlock 10 more free questions, or open the full NCA-GENL practice app.