1Z0-1127-25 Practice Test 38 — 15 Questions

Question 1

Which THREE are known challenges when deploying large language models in production?

Accepted Answer

Bias in training data perpetuating stereotypes. Option A is correct because LLMs trained on large, unfiltered corpora can learn and reproduce societal stereotypes, and mitigating this bias in production requires ongoing data curation, evaluation, and guardrails. Option B is correct because serving LLMs demands substantial GPU/accelerator resources for inference, and costs scale with model size, context length, and request volume, making latency and throughput expensive to sustain. Option C is correct because LLMs are probabilistic generators that can produce fluent, confident, but factually wrong outputs (hallucinations), which is a well-known production risk requiring grounding, retrieval, or verification. Option D is not a challenge but a claimed benefit, and it is also misleading because autoregressive token-by-token decoding limits parallelism during inference. Option E is incorrect because large models typically have very high memory footprints, especially when weights, KV cache, and activations are considered.

Answer

Fast inference speed due to parallelization

Answer

Low memory footprint

Question 2

A company notices that their OCI GenAI managed serving endpoint returns incomplete responses for long prompts. What is the most likely cause?

Accepted Answer

The model's context window is exceeded.. When a prompt exceeds the model's context window, the model cannot process the entire input, leading to truncated or incomplete responses. OCI GenAI managed serving endpoints enforce a maximum context length (e.g., 4096 tokens for many models), and any prompt longer than this limit will be cut off, causing incomplete output.

Answer

The max tokens parameter is set too high.

Answer

The top_p value is too low.

Answer

The temperature is set too high.

Question 3

A developer wants to ensure that a LangChain RetrievalQA chain returns diverse results, avoiding near-duplicate passages in the retrieved context. Which retrieval parameter should be configured on the retriever?

Accepted Answer

Maximum Marginal Relevance (MMR). Maximum Marginal Relevance (MMR) is specifically designed to reduce redundancy by balancing relevance with diversity. When configured on a retriever, MMR selects documents that are both relevant to the query and dissimilar to each other, thus avoiding near-duplicate passages in the retrieved context.

Answer

top_k

Answer

similarity_search

Answer

chunk_overlap

Question 4

During a RAG implementation, the response quality degrades because the LLM receives too many irrelevant document chunks. Which technique can best filter out irrelevant chunks before sending them to the LLM?

Accepted Answer

Implement a reranking step using a cross-encoder model.. A reranking step using a cross-encoder model (Option C) directly addresses the problem of irrelevant chunks by scoring each retrieved chunk against the user query with a deep, pairwise similarity computation. Unlike a bi-encoder used in initial retrieval, a cross-encoder evaluates query-chunk pairs jointly, allowing it to reorder or filter out chunks that are semantically irrelevant, thus improving the quality of context fed to the LLM.

Answer

Use a larger LLM for generation, hoping it ignores irrelevant chunks.

Answer

Reduce the top-k retrieval count.

Answer

Increase the chunk size to provide more context.

Question 5

An enterprise is fine-tuning a Cohere model using OCI Generative AI for a domain-specific task. After training, the model shows high accuracy on validation data but poor performance on unseen test data. What is the most likely cause?

Accepted Answer

The model overfitted to training data. The scenario describes high accuracy on validation data but poor performance on unseen test data, which is the classic symptom of overfitting. In OCI Generative AI, when fine-tuning a Cohere model, overfitting occurs when the model memorizes the training data rather than learning generalizable patterns, often due to excessive training epochs, a small dataset, or high model capacity relative to the data.

Answer

The training dataset was too small

Answer

The number of training epochs was too low

Answer

The learning rate was set too high

Question 6

Which TWO of the following are valid similarity metrics used in vector search?

Accepted Answer

Cosine similarity. Cosine similarity (B) is a valid vector-search similarity metric because it measures the cosine of the angle between two embedding vectors, giving a value from -1 to 1 that reflects semantic closeness independent of vector magnitude. Euclidean distance (C) is also valid, as it computes the straight-line L2 distance between two vectors in the embedding space, and vector databases commonly support it as a distance metric (often converted to a similarity score). Levenshtein distance (A) is not a vector metric; it measures the minimum number of single-character edits between two strings, so it applies to text/string comparison rather than dense vector embeddings. Hamming distance (D) counts differing bit positions between equal-length strings or binary codes, which is not the standard similarity measure for continuous embedding vectors. Jaccard index (E) measures set overlap (intersection over union) for discrete sets or token sets, not the geometric similarity between dense vectors used in vector search.

Answer

Levenshtein distance

Answer

Hamming distance

Answer

Jaccard index

Question 7

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

Question 8

A developer wants to interactively test different prompts and parameters (temperature, top_p, frequency_penalty) with a Cohere Command R model before integrating it into an application. Which tool should they use?

Accepted Answer

OCI Generative AI Playground. The OCI Generative AI Playground is a web-based interactive console designed specifically for experimenting with prompts and parameters like temperature, top_p, and frequency_penalty against models such as Cohere Command R. It provides immediate visual feedback, making it ideal for prompt tuning before coding an application integration.

Answer

OCI Console Resource Manager

Answer

OCI SDK for Python

Answer

OCI Data Science Notebooks

Question 9

A team is fine-tuning a foundation model on a large dataset stored in OCI Object Storage. They want to minimize data transfer costs. What is the best practice for locating the storage?

Accepted Answer

Place the bucket in the same region as the fine-tuning job. Placing the Object Storage bucket in the same OCI region as the fine-tuning job eliminates cross-region data transfer charges. OCI charges egress fees when data moves between regions, but intra-region data transfer between services in the same region is free. This minimizes costs while keeping the data accessible for the fine-tuning workload.

Answer

Place the bucket in the same region and availability domain as the fine-tuning job

Answer

Use OCI File Storage instead of Object Storage

Answer

Use a cross-region bucket to leverage geographically distributed data

Question 10

An administrator needs to ensure that only specific users in the finance department can invoke a generative AI model deployed on OCI. Which IAM policy should be used?

Accepted Answer

allow group finance_group to use generative-ai-model in compartment finance. The 'use' verb in an OCI IAM policy grants the minimum required permissions to invoke a generative AI model without allowing management actions like creating or deleting models. The policy scopes access to the 'finance_group' group and the 'finance' compartment, ensuring only specific users in the finance department can invoke the model.

Answer

allow group admins to use generative-ai-model in compartment finance

Answer

allow group finance_group to manage generative-ai-model in compartment finance

Answer

allow any-user to use generative-ai-model in compartment finance

Question 11

Which TWO statements about tokens in large language models are correct?

Accepted Answer

Common tokenization methods include word-based and subword-based.. Option A is correct because tokenization schemes in LLMs fall into categories such as word-based tokenization (splitting on whitespace/punctuation) and subword-based tokenization (e.g., BPE, WordPiece, SentencePiece), which balance vocabulary size against out-of-vocabulary issues. Option E is correct because the context window is precisely the maximum number of tokens (input plus generated output) that a model can attend to in a single forward pass, such as 4,096 or 128,000 tokens depending on the model. Option B is incorrect because embedding size is a fixed model dimension shared by all tokens in the vocabulary, not a per-token property that varies. Option C is incorrect because tokens are used at both training and inference time, since prompts and generated text are tokenized during inference as well. Option D is incorrect because subword tokenization deliberately splits words into fragments (e.g., 'tokenization' into 'token' + 'ization'), so tokens are not always whole words.

Answer

All tokens have the same embedding size.

Answer

Tokens are only used during training.

Answer

Tokens are always whole words.

Question 12

In prompt engineering, what is the role of a system prompt?

Accepted Answer

To define the model's persona, tone, and behavioral constraints. The system prompt in prompt engineering defines the model's persona, tone, and behavioral constraints, setting the overarching context for all interactions. This is distinct from user prompts, which carry the current query. By establishing these parameters, the system prompt guides the model's responses consistently across a session, ensuring alignment with desired output characteristics.

Answer

To handle error messages from the API

Answer

To specify the user's current query

Answer

To provide examples of desired outputs

Question 13

A security audit reveals that the RAG application exposes internal documents through the chatbot. The vector search index contains sensitive data. Which action should be taken FIRST to mitigate?

Accepted Answer

Implement access control at the OpenSearch index level. The correct first step is to implement access control at the OpenSearch index level because the core issue is that sensitive internal documents are accessible to unauthorized users through the chatbot. By enforcing index-level permissions (e.g., using OpenSearch's fine-grained access control with roles and users), you can restrict which documents the RAG pipeline can retrieve, directly addressing the security vulnerability at the data source. This is a foundational security measure that should precede any other adjustments.

Answer

Reduce the number of retrieved chunks

Answer

Redact sensitive terms from documents before embedding

Answer

Use a different embedding model

Question 14

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 15

A developer is using LangChain's RecursiveCharacterTextSplitter with chunk_size=1000 and chunk_overlap=200. Which statement best describes the resulting chunks?

Accepted Answer

Chunks are at most 1000 characters, and consecutive chunks overlap by 200 characters. RecursiveCharacterTextSplitter splits text into chunks of at most chunk_size characters, with consecutive chunks overlapping by chunk_overlap characters. So with chunk_size=1000 and chunk_overlap=200, each chunk is up to 1000 characters, and the last 200 characters of one chunk are repeated at the beginning of the next. This overlap helps maintain context across chunk boundaries.

Answer

Chunks are at most 1200 characters, overlapping by 200

Answer

Chunks are exactly 800 characters, with no overlap

Answer

Each chunk is exactly 1000 characters, with no overlap