1Z0-1127-25 LLM Fundamentals Practice Test 7 — 15 Questions

Question 1

What is the key advantage of multi-head attention over single-head attention in transformer models?

Accepted Answer

It allows the model to focus on different parts of the sequence simultaneously from different representation subspaces. Multi-head attention allows the model to attend to information from different representation subspaces at different positions, capturing a richer understanding.

Answer

It eliminates the need for positional encoding

Answer

It reduces the total number of parameters

Answer

It makes the model non-autoregressive

Question 2

A team is deploying an LLM-based application that must adhere to strict data residency requirements. All processing must occur within a specific OCI region. Which OCI service should they use to host and serve the LLM?

Accepted Answer

OCI Data Science with a model deployment endpoint. OCI Data Science with a model deployment endpoint is the correct choice because it provides a managed infrastructure for hosting and serving LLMs within a specific OCI region, ensuring all processing and data remain within that region to meet strict data residency requirements. The model deployment endpoint runs on dedicated compute resources in the chosen region, allowing low-latency inference while adhering to regional data boundaries.

Answer

OCI Streaming

Answer

OCI Functions

Answer

OCI Object Storage

Question 3

Which of the following is a decoder-only model architecture?

Accepted Answer

GPT-3. GPT is a decoder-only model. BERT is encoder-only. T5 is encoder-decoder.

Answer

T5

Answer

BART

Answer

BERT

Question 4

A data scientist is evaluating an LLM for a summarization task. They have a set of human-written reference summaries. Which THREE metrics are commonly used to evaluate summarization quality? (Choose three.)

Accepted Answer

BLEU. ROUGE, BLEU, and BERTScore are all used for summarization evaluation. Perplexity measures model confidence, and cosine similarity is for embedding comparison.

Answer

Cosine similarity

Answer

Perplexity

Question 5

An OCI customer is deploying a chatbot using a pre-trained LLM. They are concerned about the model generating biased or harmful content. Which TWO strategies should they implement as part of their responsible AI approach? (Choose two.)

Accepted Answer

Set up a human-in-the-loop review for sensitive queries. Option C is correct because a human-in-the-loop review process routes sensitive or high-risk queries to human reviewers who can validate, correct, or escalate the model's responses, which directly mitigates the risk of biased or harmful content reaching users. Option E is correct because a content filtering layer (for example, using OCI Language moderation or a guardrails service) inspects prompts and completions and blocks or redacts harmful outputs before they are returned, providing a deterministic safety control independent of the model's behavior. Option A is not appropriate here because the scenario uses a pre-trained LLM and retraining from scratch is costly, slow, and unnecessary for a responsible-AI guardrail. Option B is incorrect because enlarging the context window only gives the model more input tokens and does not by itself reduce bias or harmful generation. Option D is incorrect because top-k sampling with k=1 is greedy decoding, which affects determinism and diversity but provides no safety or bias mitigation.

Answer

Train the model from scratch on a curated dataset

Answer

Increase the context window to include more examples

Answer

Use top-k sampling with k=1

Question 6

Which component of the Transformer architecture allows the model to weigh the importance of different words in a sequence when processing a given word?

Accepted Answer

Self-attention mechanism. The self-attention mechanism is the core component of the Transformer architecture that computes attention scores between every pair of words in the input sequence. These scores determine how much each word should influence the representation of the current word, allowing the model to dynamically weigh the importance of different words regardless of their positional distance. This mechanism is what enables the Transformer to capture long-range dependencies and contextual relationships in parallel.

Answer

Feed-forward neural network

Answer

Positional encoding

Answer

Layer normalization

Question 7

A data scientist is evaluating two LLMs for a summarization task. Model X scores 45 on ROUGE-L, while Model Y scores 42. However, in human evaluation, Model Y is preferred 60% of the time. What is the most likely explanation?

Accepted Answer

ROUGE-L measures lexical overlap, which may not capture the semantic quality that humans value. ROUGE-L measures the longest common subsequence (LCS) between generated and reference summaries, focusing on lexical (word-level) overlap. It does not assess semantic meaning, fluency, or factual correctness. Human evaluators often prefer summaries that are coherent and capture key ideas, even if they use different wording, which explains why Model Y can score lower on ROUGE-L but be preferred 60% of the time.

Answer

Human evaluators are biased and cannot be trusted for objective assessment

Answer

Model Y overfits to the training data, causing poor generalisation

Answer

ROUGE-L is not a reliable metric for summarization because it only measures recall

Question 8

Which of the following is a primary limitation of large language models that can lead to generating factually incorrect information?

Accepted Answer

Hallucinations. Hallucinations are a primary limitation of large language models because they cause the model to generate text that is factually incorrect, nonsensical, or not grounded in the training data. This occurs due to the probabilistic nature of token prediction, where the model prioritizes fluency and coherence over factual accuracy, especially when the prompt lacks sufficient context or the model is asked to recall specific facts not well-represented in its training.

Answer

Bias in training data

Answer

Context window limitation

Answer

Knowledge cutoff

Question 9

A team is implementing a RAG pipeline in OCI. They have a large collection of PDF documents. After chunking and embedding the documents, retrieval quality is poor. Which step is MOST likely the root cause?

Accepted Answer

The chunk size is too large, causing each chunk to contain multiple topics. In a RAG pipeline, chunk size directly affects retrieval precision. If chunks are too large, each chunk contains multiple topics, so the embedding becomes a diluted average of several concepts. A query about one topic will match the chunk only weakly, and the retrieved context may include irrelevant information, degrading answer quality. Smaller, semantically coherent chunks (e.g., 200–500 tokens with overlap) typically improve retrieval.

Answer

The retrieval step uses greedy decoding

Answer

The embedding model is a generation model, not an embedding model

Answer

Cosine similarity is not appropriate for comparing embeddings

Question 10

Which of the following best describes the role of positional encoding in the Transformer architecture?

Accepted Answer

To provide information about the order of tokens in the input sequence. Positional encoding is essential in the Transformer architecture because the self-attention mechanism processes all tokens in parallel and has no inherent sense of order. By adding sinusoidal or learned positional vectors to the input embeddings, the model gains information about the relative or absolute position of each token in the sequence, enabling it to understand word order and sequence structure.

Answer

To compress the input sequence length

Answer

To increase the dimensionality of the hidden states

Answer

To reduce the effect of vanishing gradients

Question 11

A company has a large dataset of legal documents in multiple languages. They need to find documents semantically similar to a query. Which step is essential for this task?

Accepted Answer

Use a text embedding model to convert documents into dense vector representations. Semantic similarity search requires converting documents and the query into dense vector embeddings using a text embedding model, then comparing them via cosine similarity or similar metrics. This is the essential step that enables meaning-based matching across languages.

Answer

Apply BPE tokenization to all documents

Answer

Fine-tune a generation model on the legal documents

Answer

Use beam search to identify similar passages

Question 12

A research team is comparing two LLMs for a translation task. Model A uses greedy decoding, Model B uses beam search with width=5. Both models are otherwise identical. Which statement about their outputs is MOST likely true?

Accepted Answer

Model B will generally produce more fluent and accurate translations. Beam search considers multiple candidate sequences and selects the one with the highest overall probability, which often results in more fluent and accurate translations than greedy decoding, but at higher computational cost.

Answer

Model A will have higher BLEU scores than Model B

Answer

Model A will produce more diverse translations

Answer

Model B will have lower latency than Model A

Question 13

A data scientist is debugging a RAG system where the generated answers are not relevant to the retrieved documents. Which TWO factors are MOST likely causing this issue?

Accepted Answer

The retriever is returning irrelevant chunks due to poor embeddings or low similarity threshold. Option A is correct because if the retriever returns irrelevant chunks—due to weak or mismatched embeddings, an inappropriate similarity metric, or a similarity threshold set too low—then the generator receives off-topic context and its answers will not align with the intended documents. Option B is correct because in a RAG pipeline the generator must be explicitly conditioned on the retrieved chunks; if the prompt template does not instruct the model to ground its answer in the provided context (or the chunks are not inserted into the prompt at all), the model falls back on its parametric knowledge and produces answers unrelated to the retrieved documents. Option C is not the most likely cause: a context window smaller than the retrieved chunks would typically cause truncation or an error, not a systematic relevance mismatch, and modern pipelines truncate or re-rank to fit. Option D is incorrect because low temperature only makes sampling more deterministic; it does not decouple the answer from the retrieved context. Option E is not among the most likely causes here: overly small chunks can lose context and hurt retrieval quality, but that is a retrieval-side concern already captured by A, and it does not directly explain the generator ignoring retrieved documents.

Answer

The context window of the generation model is smaller than the retrieved chunks

Answer

The temperature is set too low, making outputs deterministic

Answer

The chunking strategy produces chunks that are too small, losing context

Question 14

Which component of the Transformer architecture allows the model to weigh the importance of different tokens in the input sequence when generating an output?

Accepted Answer

Self-attention mechanism. The self-attention mechanism is the core component of the Transformer architecture that computes attention scores between every pair of tokens in the input sequence. These scores determine how much each token should influence the representation of every other token, allowing the model to dynamically weigh the importance of different tokens when generating an output. This is achieved through scaled dot-product attention, where queries, keys, and values are derived from the input embeddings.

Answer

Positional encoding

Answer

Layer normalization

Answer

Feed-forward network

Question 15

An LLM generates a response that contains a plausible-sounding but factually incorrect statement about a historical event. This is an example of which known limitation?

Accepted Answer

Hallucination. Hallucination in LLMs refers to the generation of content that is plausible-sounding but factually incorrect or nonsensical. This occurs when the model's probabilistic next-token prediction produces statements that are not grounded in its training data or real-world facts, often due to overgeneralization or lack of factual recall mechanisms.

Answer

Knowledge cutoff

Answer

Bias in training data

Answer

Context length constraint