1Z0-1127-25 LLM Fundamentals — Timed 40 Questions

Question 1

When using an LLM for code generation, a developer notices the model occasionally produces syntactically incorrect code. Which approach is most likely to reduce syntax errors while still allowing diverse output?

40 minute time limit — choose an answer to begin.

Accepted Answer

Set temperature to 0 and use greedy decoding. Lowering temperature reduces randomness, making outputs more deterministic and less prone to errors, while still allowing some variation.

Answer

Use top-k sampling with k=100

Answer

Increase the context window size

Answer

Increase the temperature to 1.5

Question 2

An OCI user notices that their Llama 3 model generates the same output sequence regardless of the input prompt when using default generation parameters. Which setting is most likely causing this lack of diversity?

Accepted Answer

Greedy decoding (temperature=0). Greedy decoding, which is equivalent to setting temperature=0, always selects the token with the highest probability at each step. This deterministic behavior causes the model to produce the exact same output sequence for any input prompt, as there is no randomness or variation in token selection. The lack of diversity is a direct consequence of eliminating all stochasticity from the generation process.

Answer

Top-k sampling with k=50

Answer

Temperature sampling with temperature=0.8

Answer

Top-p (nucleus) sampling with p=0.9

Question 3

A developer is building a code generation assistant and wants to minimize the number of API calls to the OCI Generative AI service. Which tokenization approach results in the lowest token count for a given code snippet?

Accepted Answer

BPE tokenizer trained on code corpora. BPE (Byte Pair Encoding) tokenizers trained specifically on code corpora learn subword units that align closely with programming language syntax (e.g., common keywords, operators, and variable patterns), resulting in fewer tokens for a given code snippet compared to general-purpose tokenizers. This reduces API calls by encoding more semantic meaning per token, directly minimizing token count.

Answer

WordPiece tokenizer

Answer

SentencePiece tokenizer with unigram LM

Answer

Character-level tokenization

Question 4

An organization wants to deploy a model that can summarize long financial reports (5000+ tokens) without losing context. Which model architecture is best suited for this requirement?

Accepted Answer

Encoder-decoder model (e.g., T5). Encoder-decoder models like T5 or BART are designed for sequence-to-sequence tasks such as summarization, and can handle long inputs with their encoder.

Answer

Mixture-of-experts model

Answer

Decoder-only model (e.g., GPT)

Answer

Encoder-only model (e.g., BERT)

Question 5

A data scientist needs to evaluate the quality of a text summarization model. Which TWO metrics are appropriate for this task?

Accepted Answer

BERTScore. BERTScore (D) is correct because it evaluates generated text against reference summaries by computing contextual embedding similarity (typically with token-level cosine similarity and greedy matching), which captures semantic equivalence rather than only surface overlap, making it well suited to judging summarization quality. ROUGE (E) is correct because it is the standard family of recall-oriented n-gram and longest-common-subsequence overlap metrics (e.g., ROUGE-1, ROUGE-2, ROUGE-L) designed specifically to measure how much reference-summary content a generated summary covers. Perplexity (A) is not appropriate here because it measures a language model's probability distribution over tokens and does not compare a summary to a reference for content quality. F1 score (C) is a generic precision/recall harmonic mean used for classification or token-level tasks, not a summarization-specific quality metric by itself. BLEU (B) is a precision-oriented n-gram metric designed primarily for machine translation, so it is not the standard choice for summarization evaluation.

Answer

Perplexity

Answer

BLEU

Answer

F1 score

Question 6

An organization wants to use OCI Generative AI for a multilingual translation task. They need high quality and must avoid biases present in the training data. Which THREE strategies should they consider? (Select THREE.)

Accepted Answer

Implement a human-in-the-loop review process to catch biased translations. Option B is correct because a human-in-the-loop review process adds expert validation that can detect and correct biased or culturally inappropriate translations that automated metrics may miss, directly addressing the requirement to avoid biases in the training data. Option C is correct because fine-tuning a pre-trained OCI Generative AI model on a curated, high-quality parallel corpus for the specific language pairs adapts the model to domain-appropriate terminology and reduces the influence of generic or biased patterns in the original training data. Option E is correct because encoder-decoder architectures such as T5 or BART are specifically designed for sequence-to-sequence tasks like translation, making them the appropriate model family for producing fluent, high-quality multilingual output. Option A is not among the marked correct answers; while RAG can improve factual grounding, it is not one of the three strategies identified here for improving translation quality and reducing bias. Option D is incorrect because raising the temperature to 1.5 increases randomness and hallucination rather than reducing bias, and would likely degrade translation quality.

Answer

Use a RAG pipeline to retrieve canonical translations from a trusted database

Answer

Increase the temperature parameter to 1.5 to reduce repetitive biases

Question 7

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

Accepted Answer

Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store. RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

Answer

Train a custom model from scratch on the policy documents each month

Answer

Use a larger foundation model with a longer context window and paste all documents into each prompt

Answer

Fine-tune a base LLM on the policy documents monthly

Question 8

What is the primary purpose of the self-attention mechanism in a transformer model?

Accepted Answer

To process tokens in parallel while modeling long-range dependencies. The self-attention mechanism allows each token in the input sequence to attend to every other token, computing a weighted sum of their representations. This enables the model to capture long-range dependencies directly without the sequential processing constraints of RNNs, and because the attention scores for all tokens can be computed simultaneously, the mechanism supports parallel processing of the entire sequence.

Answer

To reduce the number of parameters in the model

Answer

To convert tokens into fixed-length vectors

Answer

To ensure the model is autoregressive

Question 9

Which tokenization algorithm is used by models like BERT and GPT-2?

Accepted Answer

Byte-Pair Encoding (BPE). The question asks which tokenization algorithm is used by models like BERT and GPT-2. While BERT uses WordPiece, GPT-2 uses Byte-Pair Encoding (BPE). Since GPT-2 is one of the models cited, BPE is the correct answer for that model. WordPiece is a distractor that applies only to BERT.

Answer

SentencePiece

Answer

WordPiece

Answer

Unigram Language Model

Question 10

A researcher is comparing BLEU and ROUGE scores for a machine translation model. They notice that the BLEU score is high but the ROUGE score is low. Which scenario is MOST consistent with this observation?

Accepted Answer

The model outputs concise translations that capture key words but miss some reference phrases. High BLEU and low ROUGE indicate that the generated text has high precision (many n-grams match reference) but low recall (missing many reference n-grams). This often occurs when the output is short or overly cautious.

Answer

The model outputs very long and verbose translations

Answer

The model is overfitting to the training data

Answer

The reference translations are of poor quality

Question 11

An enterprise is building a document Q&A application with OCI Generative AI. They want to minimize hallucinations. Which TWO techniques should they implement? (Choose two.)

Accepted Answer

Fine-tune the model on domain-specific data. RAG grounds answers in retrieved documents, and fine-tuning on domain-specific data reduces errors. Increasing context window alone does not reduce hallucinations; greedy decoding reduces creativity but not factual errors.

Answer

Increase the model's context window to its maximum

Answer

Use a higher temperature in sampling

Answer

Apply greedy decoding

Question 12

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

Accepted Answer

Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store. RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

Answer

Fine-tune a base LLM on the policy documents monthly

Answer

Use a larger foundation model with a longer context window and paste all documents into each prompt

Answer

Train a custom model from scratch on the policy documents each month

Question 13

An ML engineer notices that when using temperature sampling with temperature=0.8 for code generation, the model sometimes produces syntactically incorrect code. The engineer needs to ensure syntactically valid outputs while maintaining some creativity. Which combination of sampling parameters is MOST appropriate?

Accepted Answer

Reduce temperature to 0.2 and use top-p=0.9. Reducing temperature to 0.2 makes the output distribution more peaked, favoring high-probability tokens, which reduces syntax errors. Combining this with top-p=0.9 (nucleus sampling) limits the sampling pool to the smallest set of tokens whose cumulative probability reaches 0.9, further filtering out low-probability tokens that often cause invalid syntax. This balance preserves some creativity while ensuring syntactically valid code.

Answer

Increase temperature to 1.2 and use top-k=50

Answer

Use beam search with width=5

Answer

Set temperature=1.0 and use greedy decoding

Question 14

A practitioner needs to choose a pre-trained model for a sentiment analysis task on customer reviews. The model must be efficient for inference and capable of handling multiple languages. Which architecture is MOST suitable?

Accepted Answer

Mixture of Experts model. Mixture of Experts (MoE) models are designed for efficient inference by activating only a subset of parameters per input, reducing computational cost. They can also support multiple languages by allocating different experts to different language patterns, making them highly suitable for multilingual sentiment analysis. In contrast, encoder-only BERT models are efficient but may not scale as well for multilingual tasks without large capacity, encoder-decoder models are optimized for sequence-to-sequence tasks, and decoder-only models are primarily generative and less efficient for classification.

Answer

Encoder-only BERT model

Answer

Encoder-decoder T5 model

Answer

Decoder-only GPT model

Question 15

An organization wants to deploy an LLM for legal document analysis where accuracy is critical, and the model must not reference any external data outside the provided legal corpus. Which approach BEST satisfies these requirements?

Accepted Answer

Use RAG with a vector store containing only the legal documents, and set the retriever to return a fixed number of chunks with high similarity threshold. RAG can ground generation in a curated corpus, and with strict retrieval settings (e.g., only retrieving from the legal corpus), the model will not use any outside knowledge, reducing hallucinations.

Answer

Use a decoder-only model with zero-shot prompting

Answer

Use a fine-tuned encoder-only model for classification only

Answer

Use a large foundation model with a high temperature setting

Question 16

What is the primary difference between pre-training and fine-tuning in the context of large language models?

Accepted Answer

Pre-training trains from scratch, fine-tuning updates all weights on a new dataset. Pre-training trains a model from scratch on a large corpus, updating all weights via self-supervised objectives like masked language modeling or next-token prediction. Fine-tuning then adapts the pre-trained model to a specific task by updating weights (often all, sometimes a subset) on a smaller labeled dataset. Option A correctly captures that pre-training starts from random initialization and fine-tuning updates weights on a new dataset.

Answer

Pre-training uses a smaller dataset, fine-tuning uses a larger dataset

Answer

Pre-training produces embeddings, fine-tuning produces text generation

Answer

Pre-training is unsupervised, fine-tuning is always supervised

Question 17

An ML engineer is choosing an LLM for a code generation assistant. The model must generate syntactically correct code, handle multiple programming languages, and be cost-efficient. Which THREE characteristics should the engineer prioritize?

Accepted Answer

Model with a large context window (e.g., 128K tokens). Option A is correct because a large context window (e.g., 128K tokens) lets the assistant ingest entire files, multiple related modules, or long documentation in one prompt, which improves the model's ability to generate syntactically consistent and context-aware code across a large codebase. Option C is correct because high pass@k on coding benchmarks such as HumanEval directly measures functional correctness of generated programs, which is the core requirement for a code generation assistant. Option D is correct because training on a diverse multilingual code corpus enables the model to handle multiple programming languages and varied syntax, idioms, and libraries, satisfying the multi-language requirement. Option B is not appropriate because text summarization optimization targets a different task and does not specifically improve code synthesis or syntactic correctness. Option E is not appropriate because maximizing parameter count regardless of performance ignores cost-efficiency and does not guarantee better coding ability; larger models typically increase inference cost and latency.

Answer

Model optimized for text summarization tasks

Answer

Model with the highest number of parameters regardless of performance

Question 18

A team is evaluating two LLMs for a summarization task. Model X has a BERTScore of 0.85, Model Y has a BERTScore of 0.82. However, human evaluators prefer Model Y. Which TWO reasons could explain this discrepancy?

Accepted Answer

BERTScore is based on BERT embeddings, which may not fully capture summary-specific qualities like conciseness or readability. BERTScore correlates with human judgment but is not perfect; it may favor certain styles. Additionally, BERTScore may be inflated if the reference summaries are similar to the model's training data.

Answer

BERTScore uses precision only, so it misses recall aspects

Answer

Human evaluators were not given clear criteria for evaluation

Answer

Model Y was fine-tuned on a different dataset, causing distribution shift

Question 19

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

Accepted Answer

Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store. Retrieval-Augmented Generation (RAG) with a vector store allows the chatbot to retrieve relevant policy documents at query time and generate answers without retraining the model. This approach is ideal for monthly document updates because the vector store can be updated independently, and the LLM does not need fine-tuning.

Answer

Train a custom model from scratch on the policy documents each month

Answer

Fine-tune a base LLM on the policy documents monthly

Answer

Use a larger foundation model with a longer context window and paste all documents into each prompt

Question 20

A data scientist is fine-tuning a Llama 2 model on a custom dataset for a summarization task. After fine-tuning, the model produces summaries that are too similar to the input text, often copying sentences verbatim. Which adjustment is MOST likely to reduce copying and improve abstractive summarization?

Accepted Answer

Increase the temperature to 0.8 and use top-p sampling. The model is copying input text verbatim because decoding is too deterministic — likely using greedy or low-temperature sampling. Increasing temperature to 0.8 and using top-p (nucleus) sampling introduces controlled randomness, encouraging the model to explore alternative phrasings and produce more abstractive summaries rather than extractive copies.

Answer

Switch to greedy decoding

Answer

Increase the beam search width

Answer

Reduce the context window size