1Z0-1127-25 Fundamentals of Large Language Models Practice Test 10 — 15 Questions

Question 1

A financial analyst is using an OCI Generative AI model to generate quarterly earnings summaries. The model occasionally produces summaries that include fabricated financial figures not present in the source data. Which phenomenon is primarily responsible for this behavior?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Hallucination. Hallucination is the correct term for when an LLM generates factually incorrect or fabricated information that is not supported by the input data. In this scenario, the model invents financial figures, which is a hallmark of hallucination. Overfitting, catastrophic forgetting, and vanishing gradients are training-related issues that do not directly cause inference-time fabrication.

Answer

Catastrophic forgetting

Answer

Vanishing gradients

Answer

Overfitting

Question 2

A machine learning engineer is fine-tuning a large language model on OCI Generative AI. The training job completes, but the resulting model produces repetitive and nonsensical outputs when given new prompts. The engineer suspects overfitting. Which training hyperparameter adjustment is most likely to mitigate this issue?

Accepted Answer

Decrease the number of training epochs. Overfitting in fine-tuning often manifests as repetitive or nonsensical outputs on new data. Reducing the number of training epochs is a direct way to prevent the model from memorizing the training set. This allows the model to capture general patterns rather than noise, improving generalization to unseen prompts.

Answer

Increase the model's context window

Answer

Increase the learning rate

Answer

Increase the batch size

Question 3

An AI engineer is using OCI Generative AI to build a retrieval-augmented generation (RAG) system. They notice that the model sometimes ignores the retrieved context and generates answers from its parametric knowledge. Which technique would most directly improve the model's adherence to the retrieved context?

Accepted Answer

Fine-tune the model on examples where the correct answer is derived solely from the provided context.. When a model ignores retrieved context, it is often because it has not been trained to prioritize that context. Fine-tuning with examples that explicitly require using the provided context can realign the model's behavior. This approach is more effective than adjusting inference parameters or context size, which do not change the model's underlying tendency to favor its pretrained knowledge.

Answer

Use a larger context window model and include more retrieved documents.

Answer

Reduce the number of retrieved documents to only the most relevant one.

Answer

Increase the temperature parameter to encourage more creative responses.

Question 4

A developer is new to OCI Generative AI and wants to generate a short poem. They call the model with a prompt but receive an error indicating the model is not available. Which factor is most likely the cause?

Accepted Answer

The developer has not selected a region where the model is available.. OCI Generative AI models are hosted in specific regions. If a developer attempts to use a model in a region where it is not offered, the service returns an error indicating the model is unavailable. The solution is to switch to a supported region or use a different model that is available in the current region. This is a common initial setup issue.

Answer

The developer's account lacks the necessary IAM policy to use the model.

Answer

The model requires a dedicated AI cluster to be provisioned before use.

Answer

The prompt exceeds the maximum token limit, causing the model to be rejected.

Question 5

A developer is building an OCI Generative AI application that requires the model to answer questions based on a large, dynamic knowledge base. The knowledge base is updated frequently, and the developer wants to avoid retraining the model. Which approach is most suitable?

Accepted Answer

Use retrieval-augmented generation (RAG) with a vector database. Retrieval-augmented generation (RAG) combines an LLM with a retrieval system, allowing the model to access up-to-date information from a vector database without retraining. This is efficient for dynamic knowledge bases and ensures answers are grounded in current data. Fine-tuning or larger models do not provide real-time updates.

Answer

Increase the model's temperature to encourage creativity

Answer

Fine-tune the model on the entire knowledge base daily

Answer

Use a larger model with more parameters

Question 6

A solutions architect is designing a RAG pipeline on OCI Generative AI for a legal research assistant. The knowledge base contains 2,000-page contracts, and the assistant must answer clause-level questions with precise citations. Which chunking strategy should the architect implement to maximize retrieval precision and keep the generation grounded?

Accepted Answer

Split each contract into overlapping chunks aligned to semantic section boundaries, embed each chunk, and retrieve only the top-scoring chunks that fit the context window.. Semantic chunking with overlap aligned to document structure preserves clause meaning across boundaries, and top-k retrieval restricts the prompt to the most relevant passages. This combination keeps inputs within the context window while grounding generation in exact contract text, which is what enables precise clause-level citations and minimizes hallucinated legal content.

Answer

Embed the entire 2,000-page contract as one vector and pass the full document to the model on every query.

Answer

Fine-tune the base model on the full contract corpus and skip retrieval entirely for each query.

Answer

Chunk the contract into fixed 128-token segments with no overlap and retrieve a single best-matching segment.

Question 7

A machine learning engineer is evaluating an LLM for a text classification task. They notice that the model's output varies significantly each time the same input is provided, even though the task requires deterministic results. Which parameter should they adjust to reduce variability?

Accepted Answer

temperature. Temperature controls the randomness of the model's output. A temperature of 0 makes the model greedy, always choosing the most likely next token. This results in deterministic behavior, which is essential for tasks like classification where consistent output is required. Adjusting other parameters like top_p or max_tokens does not guarantee determinism.

Answer

max_tokens

Answer

top_p

Answer

beam search width

Question 8

A machine learning engineer is serving a fine-tuned OCI Generative AI model for an internal knowledge assistant. Users report that identical prompts return noticeably different answers between requests, and a compliance reviewer requires deterministic output for audit replay. Which inference configuration should the engineer apply to make responses reproducible?

Accepted Answer

Set temperature to 0 and disable top_p and top_k sampling so the model always selects the highest-probability token.. Reproducibility requires removing every stochastic element from decoding. Setting temperature to zero turns sampling into greedy selection of the most probable token, and disabling top_p and top_k prevents nucleus or top-k filtering from reintroducing randomness. With those settings, identical prompts and parameters yield identical completions, satisfying the audit replay requirement.

Answer

Enable beam search with a large beam width and leave temperature at its default.

Answer

Set temperature to a very high value and increase top_p to 1.0.

Answer

Increase max_tokens and set frequency_penalty to a negative value.

Question 9

A product manager is evaluating OCI Generative AI for a customer support application. They need a model that can understand and generate human-like text across a wide range of topics. Which type of model is best suited for this purpose?

Accepted Answer

Large language model (LLM). Large language models are specifically designed to understand and generate human-like text across many domains. They excel at conversational tasks due to their training on diverse text data. CNNs, RNNs, and decision trees are not suited for broad natural language generation. Therefore, an LLM is the correct choice.

Answer

Decision tree

Answer

Recurrent neural network (RNN)

Answer

Convolutional neural network (CNN)

Question 10

A team is planning to use OCI Generative AI to build a conversational agent that must maintain context over multiple turns. They want to ensure the agent remembers previous user inputs. Which two strategies are most effective for maintaining conversation context? (Choose two.)

Accepted Answer

Store conversation history in an external database and retrieve relevant past messages to include in the prompt.. To maintain conversation context, the model must receive the relevant history in each prompt. The two primary methods are including the full history (within token limits) or using an external store to retrieve and inject relevant past messages. Both ensure the model has the necessary information to generate context-aware responses. Other parameters like temperature or max_tokens do not provide memory.

Answer

Set a higher max_tokens limit to allow the model to generate longer responses that reference past turns.

Answer

Use a fine-tuned model that has been trained on similar conversations.

Answer

Increase the temperature parameter to encourage the model to recall previous messages.

Question 11

A data scientist is fine-tuning a large language model on OCI Generative AI for a specialized medical diagnosis task. They have a small dataset of 500 labeled examples. Which fine-tuning approach is most appropriate to avoid overfitting?

Accepted Answer

Parameter-efficient fine-tuning (PEFT) with LoRA. PEFT with LoRA is designed for scenarios with limited data, as it only trains a small number of parameters, reducing overfitting risk. Full fine-tuning and training from scratch require large datasets. Zero-shot learning lacks domain adaptation. Therefore, PEFT with LoRA is the best choice.

Answer

Full fine-tuning of all model parameters

Answer

Training from scratch

Answer

Zero-shot learning without fine-tuning

Question 12

A developer is building an OCI Generative AI application that must translate short product descriptions from English into Spanish. The descriptions are one or two sentences, and the developer wants the highest translation quality without writing training data. Which approach should the developer take?

Accepted Answer

Use prompt engineering with a pretrained instruction-following model, providing the source text and an explicit instruction to translate into Spanish.. A pretrained instruction-following model already handles multilingual generation, so a prompt that supplies the source description and explicitly requests Spanish translation delivers strong quality with zero training data and no infrastructure to manage. This matches the developer's goal of high quality with minimal effort for short texts.

Answer

Fine-tune a base model on the company's existing English product catalog before translating.

Answer

Deploy an embedding model and compute cosine similarity between English and Spanish product vectors.

Answer

Train a custom model from scratch on a parallel English-Spanish corpus using OCI Data Science.

Question 13

A team is evaluating OCI Generative AI for a customer service chatbot. They need to ensure the model's responses are factually accurate and do not include harmful content. Which TWO techniques should they implement to mitigate hallucinations and toxic outputs? (Choose two.)

Accepted Answer

Implement content moderation filters on both input and output. To mitigate hallucinations and toxic outputs, the team should ground responses in verified documents using RAG and apply content moderation filters to both inputs and outputs. These techniques directly address factual accuracy and harmful content. Other options like increasing temperature or using unverified data would worsen the issues.

Answer

Fine-tune the model on a large dataset of unverified social media posts

Answer

Set max tokens to a very low value to limit response length

Answer

Increase the temperature to make responses more creative

Question 14

A financial analyst is using an OCI Generative AI model to generate quarterly earnings summaries. The model occasionally produces summaries that include fabricated financial figures not present in the source data. Which phenomenon is primarily responsible for this behavior?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Provide few-shot examples that pair transcript excerpts with the correct metric labels and require the model to quote the supporting sentence for each extracted value.. Few-shot examples teach the model the precise mapping between transcript phrasing and metric labels, while requiring a supporting quote enforces grounding in the source text. The model must identify evidence before producing a value, which measurably reduces mismatches when several similar numbers appear in one transcript.

Answer

Increase the temperature so the model explores alternative interpretations of each figure.

Answer

Shorten the prompt by removing metric definitions to reduce token usage.

Answer

Request the output as a single free-form paragraph describing all financial results.

Question 15

A platform team is evaluating an OCI Generative AI deployment for a customer-facing assistant. They must reduce the risk of the model producing fabricated facts and of leaking sensitive prompt content. Which two techniques should the team implement to address these risks? (Choose two.)

Accepted Answer

Ground responses in retrieved enterprise documents and instruct the model to answer only from the supplied context, declining when the context is insufficient.. Grounding generation in retrieved documents with an explicit abstain instruction constrains the model to verifiable source material, which curbs fabricated facts. Guardrails that inspect and redact sensitive entities on both the request and the response prevent confidential values from reaching the model or the user, covering the leakage risk. Together they address both concerns.

Answer

Increase the context window by selecting a model with the largest available token limit.

Answer

Raise the temperature and top_p values so the assistant produces more varied and natural-sounding answers.

Answer

Disable logging on the inference endpoint so prompts and completions are never stored.