1Z0-1127-25 LLM Fundamentals Practice Test 9 — 15 Questions

Question 1

A company wants to translate legal documents from English to Spanish. They have a small parallel corpus of 500 sentence pairs. Which approach is MOST likely to yield the best translation quality?

Accepted Answer

Fine-tune an encoder-decoder model like T5 on the parallel corpus. Fine-tuning an encoder-decoder model like T5 on the 500 sentence pairs is the best approach because it leverages the model's pre-trained knowledge of language structure and translation patterns, then adapts it to the specific legal domain with a small but relevant parallel corpus. This transfer learning method requires far less data than training from scratch and typically outperforms zero-shot prompting for specialized, low-resource translation tasks.

Answer

Use a zero-shot prompt with a decoder-only model like GPT

Answer

Train a new model from scratch on the 500 sentence pairs

Answer

Use a rule-based machine translation system

Question 2

An AI engineer is deploying a RAG pipeline using OCI Generative AI. They notice the generated answers sometimes include information not present in the retrieved documents. What is the MOST likely cause?

Accepted Answer

The generation model is hallucinating because it does not rely solely on the retrieved context. The generation model in a RAG pipeline is designed to leverage retrieved context but can still produce outputs not grounded in that context, a phenomenon known as hallucination. This occurs when the model relies on its parametric knowledge or statistical patterns rather than strictly adhering to the retrieved documents, especially if the prompt does not enforce strict grounding or the model's training biases override the context.

Answer

The chunking strategy splits sentences across chunks

Answer

The embedding model is too small to represent document semantics

Answer

The context window of the generation model is exceeded

Question 3

A team is designing a text classification system using OCI Generative AI. They have a small labeled dataset of 200 examples per class. Which THREE techniques can help improve model performance without requiring additional labeled data?

Accepted Answer

Generate synthetic labeled examples using an LLM and include them in training. Option A is correct because using an LLM to generate synthetic labeled examples (data augmentation) expands the small 200-example-per-class training set without collecting new human-labeled data, which typically improves classifier robustness and performance. Option C is correct because extracting dense embeddings from a pretrained embedding model and training a lightweight classifier such as an SVM is a standard few-shot approach that leverages pretrained representations and works well with limited labeled data. Option E is correct because in-context learning places a few labeled examples directly in the prompt, letting a foundation model perform classification without any parameter updates or additional labeled data. Option B does not belong because enlarging the context window to hold all training examples is not a training technique and does not by itself improve a classifier's learned performance. Option D does not belong because fine-tuning a large decoder-only model on only 200 examples per class is prone to overfitting and is not a reliable way to boost performance without more labeled data.

Answer

Increase the context window of the model to include all training examples

Answer

Fine-tune a large decoder-only model on the 200 examples

Question 4

An organization is concerned about bias in their LLM-powered hiring assistant. Which TWO actions are MOST effective in mitigating bias?

Accepted Answer

Implement human-in-the-loop evaluation with fairness-focused rubrics. Option D is correct because human-in-the-loop evaluation with fairness-focused rubrics directly targets bias by having reviewers assess model outputs against explicit fairness criteria, catching discriminatory patterns that automated metrics may miss and enabling corrective feedback before decisions are finalized. Option E is correct because fine-tuning on a carefully curated dataset that balances demographic representation addresses bias at its source—the training data—reducing the model's tendency to learn and reproduce skewed associations from underrepresented or overrepresented groups. Options A, B, and C do not belong: a larger context window (A) only provides more input tokens and does nothing to correct biased training or outputs; increasing temperature (B) merely adds randomness and can actually amplify inconsistent or biased responses rather than mitigate bias; and using encoder-only models like BERT (C) is a task/architecture choice for classification, not a bias-mitigation measure, since such models can still inherit and reproduce bias from their training data.

Answer

Use a larger context window to include more examples

Answer

Increase the temperature parameter to introduce more randomness

Answer

Use only encoder-only models like BERT for classification

Question 5

A data scientist is evaluating a summarization model on a news article dataset. They compute ROUGE-L and BLEU scores. The ROUGE-L score is high, but the BLEU score is low. Which of the following best explains this discrepancy?

Accepted Answer

The summaries are very short and use different vocabulary from the reference. ROUGE-L measures longest common subsequence, which captures fluency and structural similarity, while BLEU measures n-gram precision, which is sensitive to vocabulary overlap and length. A high ROUGE-L with low BLEU suggests the summaries are structurally similar but use different words and are shorter, leading to low n-gram precision.

Answer

The BLEU score is unreliable for summarization tasks

Answer

The model is overfitting to the training data

Answer

The summaries are too long and contain many irrelevant n-grams

Question 6

An organization needs to select a tokenisation algorithm for a multilingual LLM that will process English, Chinese, and Korean text efficiently. Which tokenisation method is BEST suited for this requirement?

Accepted Answer

SentencePiece. SentencePiece is language-agnostic and treats input as a raw byte stream, making it highly effective for multilingual models covering English, Chinese, and Korean without language-specific pre-tokenization. It supports subword regularization and handles scripts without whitespace delimiters (Chinese, Korean) gracefully.

Answer

WordPiece

Answer

Byte-Pair Encoding (BPE)

Answer

Character-level tokenisation

Question 7

Which of the following is a key limitation of large language models that RAG (Retrieval-Augmented Generation) aims to address?

Accepted Answer

Hallucinations (factual errors). RAG addresses hallucinations by grounding the model's output in retrieved documents that contain factual information. Knowledge cutoff, bias, and context length constraints are separate issues that RAG may partially help with, but its primary purpose is to reduce factual errors.

Answer

Context length constraints

Answer

Bias in training data

Answer

Knowledge cutoff date

Question 8

A researcher wants to compare two summarization models. Model A achieves a higher ROUGE-L score than Model B, but human evaluators prefer Model B's summaries. Which of the following is the MOST likely reason?

Accepted Answer

ROUGE-L measures n-gram overlap, which may not align with human judgment of quality. ROUGE-L measures n-gram overlap, which may not capture semantic quality. Human evaluators often prefer summaries that are fluent, coherent, and concise, even if they use different wording. The discrepancy indicates that ROUGE-L alone is insufficient for evaluation.

Answer

Model A overfits to the training data

Answer

Model B has a larger context window

Answer

Model A is an encoder-decoder model while Model B is decoder-only

Question 9

Which model architecture is used by BERT for natural language understanding tasks?

Accepted Answer

Encoder-only. BERT (Bidirectional Encoder Representations from Transformers) uses a Transformer encoder-only architecture. It processes the entire input sequence bidirectionally using self-attention, which is why it excels at natural language understanding tasks such as classification, named entity recognition, and question answering. There is no decoder because BERT does not generate text autoregressively.

Answer

Recurrent neural network

Answer

Encoder-decoder

Answer

Decoder-only

Question 10

An OCI user wants to generate embeddings for a large corpus of technical documents to enable semantic search. Which type of model should they use?

Accepted Answer

An embedding model like Cohere Embed. Embedding models are specifically designed to produce dense vector representations that capture semantic meaning. They are distinct from generation models. For semantic search, embeddings from an embedding model are compared using cosine similarity.

Answer

A summarization model

Answer

A classification model

Answer

A generation model like Cohere Command

Question 11

A developer notices that an LLM-based question-answering system sometimes provides answers that are correct but from an outdated version of the knowledge base. The system uses RAG with a vector database updated daily. What is the MOST likely root cause?

Accepted Answer

The embedding model was not re-run on the updated documents, so the index contains old embeddings. The core issue is that the vector database index still contains old embeddings. Even though the knowledge base documents are updated daily, if the embedding model is not re-run on those updated documents, the vector representations in the index remain stale. When the RAG system retrieves, it fetches these outdated embeddings, leading to correct but outdated answers. This is a classic index synchronization problem in RAG pipelines.

Answer

The retrieval top-k parameter is set too high

Answer

The chunking strategy splits documents into too-small pieces

Answer

The LLM's training data has a knowledge cutoff date

Question 12

Which of the following is a distinguishing feature of in-context learning compared to fine-tuning?

Accepted Answer

In-context learning does not update the model's weights; instead, examples are provided in the prompt. In-context learning does not update model weights; it provides examples in the prompt at inference time. Fine-tuning updates the model weights through additional training on a dataset.

Answer

In-context learning modifies the model's weights based on examples

Answer

In-context learning is only possible with encoder-only models

Answer

In-context learning requires additional training on a labeled dataset

Question 13

An enterprise is deploying an LLM application on OCI and must minimize hallucinations. Which TWO strategies should they implement? (Choose two.)

Accepted Answer

Apply prompt engineering techniques such as asking the model to cite sources. Option A is correct because prompt engineering techniques such as instructing the model to cite sources or to answer only from provided evidence constrain the model's generation and reduce unsupported claims, which directly lowers hallucination rates. Option E is correct because Retrieval-Augmented Generation (RAG) grounds the LLM's responses in a curated, authoritative knowledge base retrieved at inference time, so answers are anchored to verifiable content rather than the model's parametric memory. Option B is incorrect because increasing the temperature parameter raises randomness and diversity, which typically increases hallucination rather than reducing it. Option C is incorrect because using a smaller model does not inherently reduce hallucinations and may even degrade factual accuracy. Option D is incorrect because fine-tuning on factual question-answer pairs can improve style and domain fit but does not reliably prevent hallucinations and can even reinforce confident but incorrect outputs.

Answer

Increase the temperature parameter to encourage more diverse outputs

Answer

Use a smaller model to reduce complexity

Answer

Fine-tune the model on a dataset of factual question-answer pairs

Question 14

A machine learning engineer is evaluating the performance of a translation model using BLEU score. Which THREE statements about BLEU are correct? (Choose three.)

Accepted Answer

BLEU includes a brevity penalty to penalize outputs that are too short. BLEU is a precision-based metric (not recall). It uses modified n-gram precision with a brevity penalty. It correlates reasonably well with human judgment at the corpus level but has known limitations such as not capturing semantic equivalence.

Answer

BLEU measures recall of n-grams by comparing the output to the reference

Answer

BLEU is a recall-oriented metric

Question 15

Which component of the Transformer architecture allows each token to consider the relevance of every other token in the input sequence?

Accepted Answer

Self-attention. Self-attention computes attention scores between all pairs of tokens, enabling the model to capture dependencies across the entire sequence.

Answer

Multi-head attention

Answer

Feed-forward network

Answer

Positional encoding