Knowledge + Practice

CCNA Fundamentals of Large Language Models Questions

75 of 105 questions · Page 1/2 · Fundamentals of Large Language Models · Answers revealed

Practice these questions Domain overview All questions

1

Multi-Selectmedium

Which three factors most significantly affect the quality of an LLM's output? (Select THREE)

Select 3 answers

A.Model's context window size

B.Clarity of the prompt

C.Number of GPUs used during inference

D.Temperature setting

E.Quality of training data

AnswersB, D, E

Correct: Clear prompts yield more accurate responses.

Why this answer

The clarity of the prompt directly determines how well the LLM interprets the user's intent. A well-structured, unambiguous prompt reduces ambiguity and guides the model toward generating relevant and coherent responses, while a vague or poorly worded prompt often leads to off-target or nonsensical output.

Exam trap

Oracle often tests the misconception that hardware resources like GPU count directly improve output quality, whereas in reality they only affect performance metrics like latency and throughput, not the semantic quality of the generated text.

Practice this question →

2

Multi-Selecteasy

Which TWO are advantages of using retrieval-augmented generation (RAG) over fine-tuning for incorporating new knowledge?

Select 2 answers

A.Better at capturing domain-specific writing style

B.Enables the model to access up-to-date information without retraining

C.Eliminates the need for a vector database

D.Reduces token usage and latency compared to fine-tuning

E.Cost-effective for large corpora that change frequently

AnswersB, E

RAG retrieves fresh data from external sources.

Why this answer

RAG retrieves relevant, up-to-date information from an external knowledge base at inference time, allowing the model to answer questions about recent events or proprietary data without requiring any retraining. This is a key advantage over fine-tuning, which would need a new training cycle to incorporate the same new knowledge.

Exam trap

Oracle often tests the misconception that RAG is always faster or cheaper than fine-tuning, when in reality RAG introduces retrieval latency and higher token usage, making it less suitable for low-latency or high-throughput scenarios.

Practice this question →

3

MCQeasy

A user wants to use OCI Generative AI to generate marketing copy. They want the output to be more creative and varied. Which parameter should they adjust?

A.Set temperature to 0.

B.Increase the temperature parameter.

C.Decrease the temperature parameter.

D.Increase the max_tokens parameter.

AnswerB

Higher temperature increases randomness, leading to more creative and varied text generation.

Why this answer

Increasing the temperature parameter makes the model's output more random and diverse, which is ideal for creative tasks like generating marketing copy. A higher temperature (e.g., 0.7–1.0) increases the probability of sampling less likely tokens, leading to more varied and imaginative text. Setting temperature to 0 would make the output deterministic and repetitive, which is the opposite of what the user wants.

Exam trap

Oracle often tests the misconception that increasing max_tokens or adjusting other parameters like top_p can substitute for temperature when the goal is to increase creativity, but only temperature directly controls randomness and diversity in token selection.

How to eliminate wrong answers

Option A is wrong because setting temperature to 0 forces the model to always choose the most likely token, resulting in deterministic, repetitive, and less creative output. Option C is wrong because decreasing the temperature reduces randomness, making the output more conservative and less varied, which contradicts the goal of creativity. Option D is wrong because increasing max_tokens only extends the maximum length of the generated text; it does not affect the creativity or variability of the output.

Practice this question →

4

MCQhard

A research team is experimenting with few-shot prompting to improve a model's performance on a complex reasoning task. They find that the model's performance degrades when the few-shot examples are too similar to each other. What is the likely cause and best remedy?

A.The model has not seen enough examples. Increase the number of few-shot examples.

B.The examples are presented in a confusing order. Reorder them by difficulty.

C.The examples lack diversity, causing the model to overfit to a narrow pattern. Use more diverse examples.

D.The temperature is too low, making the model too deterministic. Increase temperature slightly.

AnswerC

Diverse examples reduce bias and improve generalization.

Why this answer

When few-shot examples are too similar, the model overfits to a narrow pattern, reducing its ability to generalize to the diverse reasoning paths required by the task. This is a known limitation of in-context learning: the model treats the examples as a template rather than as diverse demonstrations. Using more diverse examples exposes the model to a wider range of reasoning patterns, improving robustness.

Exam trap

Oracle often tests the misconception that more examples always improve performance, when in fact diversity is critical to prevent overfitting in few-shot prompting.

How to eliminate wrong answers

Option A is wrong because the issue is not the quantity of examples but their lack of diversity; adding more similar examples would worsen overfitting. Option B is wrong because the order of examples (by difficulty) does not address the core problem of pattern overfitting; confusing order may affect performance but is not the likely cause here. Option D is wrong because temperature controls randomness in token sampling, not the model's sensitivity to example diversity; a low temperature would make outputs more deterministic but does not cause or fix overfitting to narrow patterns.

Practice this question →

5

Multi-Selectmedium

Which TWO factors most directly impact the consistency of text generated by an LLM when the same prompt is used multiple times?

Select 2 answers

A.Top_p

B.Batch size

C.Max_tokens

D.Temperature

E.Seed

AnswersA, D

Top_p controls the nucleus of tokens considered; lower values make output more focused.

Why this answer

Top_p (nucleus sampling) directly impacts consistency by controlling the cumulative probability threshold for token selection. A lower Top_p (e.g., 0.1) restricts the model to only the most probable tokens, reducing randomness and making outputs more deterministic across repeated prompts. This parameter, along with Temperature, is a primary lever for managing output variability in LLMs.

Exam trap

Oracle often tests the distinction between inference-time parameters (Temperature, Top_p) and training/hardware parameters (Batch size), or between parameters that control randomness (Temperature, Top_p) versus those that control output length (Max_tokens), leading candidates to mistakenly select Seed as a primary consistency factor.

Practice this question →

6

MCQeasy

A startup needs to deploy a large language model for a customer support chatbot that requires low latency and cost efficiency. They are evaluating OCI Generative AI models. Which model type is most appropriate?

A.Embedding model (e.g., cohere.embed)

B.Instruct model (e.g., cohere.command)

C.Image generation model

D.Base model (e.g., cohere.base)

AnswerB

Instruct models are fine-tuned to follow instructions, making them ideal for chatbots.

Why this answer

The startup requires low latency and cost efficiency for a customer support chatbot. Instruct models like cohere.command are specifically fine-tuned to follow conversational instructions and generate concise, task-oriented responses, making them ideal for interactive chatbot applications. They balance performance and cost better than base models, which lack instruction-following capability, and embedding models, which are designed for semantic search rather than text generation.

Exam trap

Oracle often tests the distinction between base models and instruct models, trapping candidates who assume a base model can be used directly for task-specific applications without fine-tuning or instruction alignment.

How to eliminate wrong answers

Option A is wrong because embedding models (e.g., cohere.embed) are designed to convert text into vector representations for tasks like semantic search or clustering, not for generating conversational responses. Option C is wrong because image generation models are used for creating or editing images, not for text-based customer support interactions. Option D is wrong because base models (e.g., cohere.base) are general-purpose language models that have not been fine-tuned for instruction following, leading to less relevant and less controllable outputs for a chatbot use case.

Practice this question →

7

MCQeasy

A team wants to deploy an LLM for real-time inference with low latency. Which OCI deployment option is best?

A.OCI Data Science Model Deployment with GPU shapes

B.OCI Functions with CPU

C.OCI Events

D.OCI Streaming

AnswerA

GPU shapes provide the compute power needed for low-latency LLM inference.

Why this answer

OCI Data Science Model Deployment with GPU shapes is the best option because it provides managed, scalable, low-latency inference endpoints for LLMs. GPU shapes (e.g., VM.GPU.A10) are essential for the parallel matrix computations required by transformer-based models, and the deployment service supports auto-scaling and load balancing to maintain real-time response times.

Exam trap

The trap here is that candidates may confuse OCI Functions (a serverless compute service) with a viable inference platform, overlooking the GPU requirement for LLM workloads, or mistakenly think OCI Streaming can process inference requests because of its 'real-time' label.

How to eliminate wrong answers

Option B (OCI Functions with CPU) is wrong because OCI Functions is a serverless compute service designed for short-lived, stateless functions, and CPU-only execution cannot meet the low-latency requirements of LLM inference due to the lack of GPU acceleration for large matrix operations. Option C (OCI Events) is wrong because OCI Events is a notification and orchestration service for reacting to infrastructure changes, not a compute platform for running inference workloads. Option D (OCI Streaming) is wrong because OCI Streaming is a real-time data ingestion and messaging service (based on Apache Kafka) for handling event streams, not for executing LLM inference.

Practice this question →

8

MCQmedium

A developer is building a code generation assistant. The model occasionally produces syntactically correct but semantically wrong code. Which technique directly addresses semantic correctness?

A.Expand the token vocabulary

B.Lower the temperature to 0

C.Apply RLHF using human-validated code examples

D.Increase beam search width

AnswerC

RLHF directly optimizes for desired outcomes like semantic correctness.

Why this answer

Reinforcement Learning from Human Feedback (RLHF) directly addresses semantic correctness by fine-tuning the model using human-validated code examples. This process teaches the model to prefer outputs that are not only syntactically valid but also logically correct and aligned with developer intent, reducing semantically wrong code generation.

Exam trap

Oracle often tests the misconception that adjusting decoding parameters (temperature, beam search) or tokenization can fix semantic errors, when in fact only training techniques like RLHF that incorporate human feedback can directly improve semantic correctness.

How to eliminate wrong answers

Option A is wrong because expanding the token vocabulary increases the range of tokens the model can generate but does not improve the model's ability to reason about code semantics or correct logical errors. Option B is wrong because lowering the temperature to 0 makes the model deterministic, reducing randomness but not fixing underlying semantic misunderstandings; it may still produce the same incorrect logic repeatedly. Option D is wrong because increasing beam search width explores more candidate sequences during decoding, which can improve syntactic fluency but does not directly address semantic correctness or logical accuracy.

Practice this question →

9

MCQeasy

Refer to the exhibit. A user in group GenAIUsers reports that they cannot call the OCI Generative AI API. What is the most likely issue?

A.The policy statement is missing the 'inspect' verb.

B.The policy is in INACTIVE state.

C.The compartment ID in the policy does not match the user's compartment.

D.The user is not in the group GenAIUsers.

AnswerC

The policy applies to 'ExampleCompartment' by name, but the user may be in a different compartment. The compartment OCID in the policy header does not match the compartment name in the statement, indicating a mismatch.

Why this answer

The policy is scoped to a specific compartment ID, but the user's compartment does not match that ID. For OCI IAM policies to grant access to resources like the Generative AI API, the policy must be written for the compartment where the resource resides or where the user operates. Since the user is in a different compartment, the policy does not apply, causing the API call to fail.

Exam trap

Oracle often tests the misconception that a user's group membership is the sole factor for policy applicability, ignoring that the compartment scope in the policy statement must match the user's compartment or resource compartment for the policy to take effect.

How to eliminate wrong answers

Option A is wrong because the 'inspect' verb is not required for calling the Generative AI API; the policy uses 'allow group GenAIUsers to manage generative-ai-family in compartment ...', which includes all verbs (inspect, read, use, manage) and is sufficient. Option B is wrong because the exhibit shows the policy is in ACTIVE state, not INACTIVE; an INACTIVE policy would be explicitly marked and would not enforce any rules. Option D is wrong because the user reports being in group GenAIUsers, and the policy targets that group; if the user were not in the group, the error would be an authorization failure, but the issue here is compartment mismatch.

Practice this question →

10

MCQhard

During inference with OCI Generative AI, you notice that the model is generating repetitive phrases. Which combination of parameters can help reduce repetition?

A.Top_p = 0.1, frequency_penalty = 0.5

B.Top_p = 0.9, frequency_penalty = 0.5

C.Top_p = 0.9, frequency_penalty = 0.0

D.Top_p = 1.0, frequency_penalty = 0.0

AnswerB

This combination applies a gentle penalty on repeated tokens while keeping token selection diverse, effectively reducing repetition.

Why this answer

A high Top_p value (0.9) allows the model to consider a diverse set of tokens, reducing the chance of getting stuck in repetitive loops, while a positive frequency_penalty (0.5) actively penalizes tokens that have already been generated, discouraging the model from repeating the same phrases. Together, these parameters balance creativity and repetition suppression.

Exam trap

Oracle often tests the misconception that lowering Top_p (making it more restrictive) reduces repetition, when in fact it can worsen repetition by limiting the model to only the most probable tokens, which are often the same ones already used.

How to eliminate wrong answers

Option A is wrong because Top_p = 0.1 is too restrictive, forcing the model to sample from only the top 10% of probable tokens, which actually increases the likelihood of repetitive patterns by narrowing the token pool. Option C is wrong because frequency_penalty = 0.0 means no penalty is applied for repeated tokens, so even with a high Top_p, the model has no disincentive to repeat phrases. Option D is wrong because Top_p = 1.0 (no nucleus sampling) combined with frequency_penalty = 0.0 provides no mechanism to reduce repetition, effectively using raw probability sampling without any diversity-enhancing constraints.

Practice this question →

11

MCQhard

An architect is optimizing an LLM application that processes long documents. The model has a 4096 token limit, but the documents are often 8000 tokens. They are using a chunking strategy. However, model responses sometimes miss key information that spans across chunks. Which technique most directly addresses this issue?

A.Randomly select parts of the document to include.

B.Increase the max_tokens parameter for longer outputs.

C.Use overlapping chunks to maintain context continuity.

D.Use a model with a larger context window.

AnswerC

Overlapping ensures that information at chunk boundaries is not lost.

Why this answer

Overlapping chunks ensure that tokens at the boundaries of one chunk are also present at the start of the next, preserving context continuity. This prevents the model from losing information that spans across chunk boundaries, which is a common issue when processing documents longer than the model's 4096-token context window.

Exam trap

Oracle often tests the misconception that increasing output length (max_tokens) can compensate for input context limitations, but the trap here is that max_tokens only affects the response length, not the model's ability to see the full document.

How to eliminate wrong answers

Option A is wrong because randomly selecting parts of the document discards structured information and introduces unpredictability, making it impossible to reliably capture cross-chunk dependencies. Option B is wrong because increasing the max_tokens parameter only controls the length of the generated output, not the input context size; the model still cannot process the full 8000-token document at once. Option D is wrong because while using a model with a larger context window would solve the problem, it is not a chunking strategy and may not be feasible due to cost, latency, or model availability; the question specifically asks for a technique that addresses the issue within the current chunking approach.

Practice this question →

12

Multi-Selecthard

Which TWO factors most significantly influence the computational cost of fine-tuning a large language model?

Select 2 answers

A.Batch size

B.Number of model parameters

C.Maximum sequence length

D.Quantization bits

E.Dataset size

AnswersB, C

More parameters increase compute and memory requirements.

Why this answer

The number of model parameters directly determines the size of the weight matrices that must be updated during backpropagation. Fine-tuning requires storing gradients and optimizer states for each parameter, so the memory and compute scale linearly with parameter count. This is the primary driver of FLOPs (floating-point operations) per training step.

Exam trap

Candidates often confuse factors affecting per-step computational cost (parameters, sequence length) with those affecting total training time or memory efficiency (batch size, quantization, dataset size), leading to incorrect selections.

Practice this question →

13

MCQmedium

A multi-turn chatbot needs to maintain context across user queries. The context window is limited. What design should be used?

A.Use a summary of previous turns and add new input.

B.Store context in a separate database and retrieve each time.

C.Reset context after each turn.

D.Keep the entire conversation history in each request.

AnswerA

Correct: Summarization preserves context within limits.

Why this answer

Summarizing previous turns and appending the new input efficiently manages the limited context window of large language models (LLMs). This approach preserves essential conversational context without exceeding token limits, ensuring coherent multi-turn interactions.

Exam trap

Oracle often tests the misconception that storing context externally (Option B) bypasses the context window limit, but the retrieved data must still be injected into the model's input, which is constrained by the same token budget.

How to eliminate wrong answers

Option B is wrong because storing context in a separate database and retrieving it each time introduces latency and does not inherently solve the context window limitation; the retrieved context still needs to fit into the model's input. Option C is wrong because resetting context after each turn breaks conversational continuity, making the chatbot unable to reference prior exchanges. Option D is wrong because keeping the entire conversation history in each request quickly exceeds the context window's token limit, causing truncation or errors.

Practice this question →

14

MCQeasy

A company wants to use OCI Generative AI to summarize customer reviews. Which model parameter should be adjusted to control the creativity of the summary?

A.Temperature

B.Frequency penalty

C.Top-k

D.Presence penalty

AnswerA

Temperature directly controls randomness and creativity.

Why this answer

Temperature controls the randomness of token selection in the model's output distribution. A higher temperature (e.g., 0.9) makes the summary more creative and diverse, while a lower temperature (e.g., 0.1) makes it more deterministic and focused. For summarizing customer reviews, adjusting temperature directly influences how novel or conservative the generated text will be.

Exam trap

Oracle often tests the distinction between parameters that control randomness (temperature) versus those that control repetition (frequency/presence penalties) or sampling pool size (top-k), leading candidates to confuse diversity with creativity.

How to eliminate wrong answers

Option B (Frequency penalty) is wrong because it reduces the likelihood of repeating the same tokens or phrases, which controls redundancy rather than creativity. Option C (Top-k) is wrong because it limits the sampling pool to the k most likely next tokens, which affects diversity but not the overall creativity or randomness of the output. Option D (Presence penalty) is wrong because it penalizes tokens that have already appeared in the text, encouraging the model to introduce new topics, but this does not directly control the creativity of the summary.

Practice this question →

15

Multi-Selecthard

Which three statements about transformer architecture are correct? (Choose three.)

Select 3 answers

A.The softmax function is used in the attention mechanism to normalize attention scores.

B.The feed-forward network applies a different set of weights for each token position.

C.Positional encodings are necessary because the model is not recurrent.

D.The self-attention layer allows the model to weigh the importance of different tokens.

E.The encoder-decoder structure is used in GPT models.

AnswersA, C, D

Softmax converts attention scores into probabilities.

Why this answer

The softmax function is applied to the raw attention scores (the dot products between queries and keys) to convert them into a probability distribution that sums to 1. This normalization allows the model to assign a relative weight to each token in the sequence, ensuring that the weighted sum of values is stable and interpretable.

Exam trap

Oracle often tests the distinction between encoder-decoder and decoder-only architectures, trapping candidates who assume all transformer-based models follow the original encoder-decoder design, when in fact GPT and other autoregressive models use only the decoder stack.

Practice this question →

16

MCQeasy

Which OCI service provides pre-trained models for custom text classification without requiring fine-tuning?

A.OCI Generative AI

B.OCI AI Language

C.OCI Data Science

D.OCI Vision

AnswerB

OCI AI Language provides pre-trained models for text classification without fine-tuning.

Why this answer

B is correct because OCI AI Language provides pre-trained models that can perform custom text classification out-of-the-box without requiring fine-tuning. It offers built-in models for common NLP tasks like sentiment analysis, entity extraction, and text classification, allowing users to classify text into custom categories defined by their own labels without additional training.

Exam trap

Oracle often tests the distinction between pre-trained models that require no fine-tuning versus platforms that require custom model training, leading candidates to mistakenly choose OCI Data Science or OCI Generative AI when the question specifically asks for a service that provides pre-trained models for custom text classification without fine-tuning.

How to eliminate wrong answers

Option A is wrong because OCI Generative AI focuses on generating text, images, and code using large language models, not on pre-trained models for custom text classification without fine-tuning. Option C is wrong because OCI Data Science is a platform for building, training, and deploying custom machine learning models, requiring users to fine-tune or train models from scratch rather than providing pre-trained classification models. Option D is wrong because OCI Vision is designed for image analysis tasks such as object detection and image classification, not for text classification.

Practice this question →

17

MCQeasy

A team wants to evaluate an LLM's performance on a text classification task. Which metric is most appropriate for a balanced dataset?

A.BLEU score

B.Perplexity

C.Accuracy

D.ROUGE score

AnswerC

Accuracy directly measures correct predictions, appropriate for balanced data.

Why this answer

Accuracy is the most appropriate metric for evaluating an LLM on a text classification task with a balanced dataset because it directly measures the proportion of correctly predicted labels out of total predictions. For balanced classes, accuracy provides a reliable and intuitive performance indicator without the distortion caused by class imbalance.

Exam trap

Oracle often tests the distinction between metrics for generation tasks (BLEU, ROUGE, perplexity) versus classification tasks (accuracy, F1-score), and the trap here is assuming a language model metric like perplexity applies to any NLP task, when it is specific to probabilistic language modeling.

How to eliminate wrong answers

Option A is wrong because BLEU score is designed for evaluating machine translation quality by comparing n-gram overlap between generated and reference text, not for classification tasks. Option B is wrong because perplexity measures how well a language model predicts a sequence of tokens, typically used for language modeling or generation, not for discrete label classification. Option D is wrong because ROUGE score is used for summarization evaluation by measuring recall-oriented overlap of n-grams, not for classification accuracy.

Practice this question →

18

MCQeasy

A developer notices that an LLM's responses are too verbose. Which parameter adjustment would most effectively reduce verbosity?

A.Increase frequency_penalty

B.Increase top_p

C.Decrease max_tokens

D.Decrease temperature

AnswerC

Max_tokens directly controls the maximum output length, reducing verbosity.

Why this answer

Decreasing max_tokens directly limits the maximum length of the LLM's response, which is the most straightforward way to reduce verbosity. This parameter caps the number of tokens the model can generate, forcing it to produce shorter completions. Other parameters like frequency_penalty, top_p, and temperature influence the style, diversity, or randomness of the output but do not directly control response length.

Exam trap

The trap here is that candidates confuse parameters that affect output style (temperature, top_p, frequency_penalty) with the one that directly controls output length (max_tokens), leading them to choose a parameter that changes how the model says something rather than how much it says.

How to eliminate wrong answers

Option A is wrong because increasing frequency_penalty reduces repetition by penalizing tokens that have already appeared, which can actually make responses more varied and potentially longer as the model avoids reusing words. Option B is wrong because increasing top_p (nucleus sampling) considers a larger set of probable tokens, which can increase diversity and often leads to longer, more exploratory responses. Option D is wrong because decreasing temperature makes the model more deterministic and focused on high-probability tokens, but it does not cap the length of the response; the model can still generate verbose text if it deems it likely.

Practice this question →

19

MCQmedium

An enterprise is deploying a chat application using a large language model. Users report that the model sometimes generates toxic or biased responses. Which best practice should be applied to mitigate this issue?

A.Use few-shot prompting with examples of toxic responses so the model learns to avoid them.

B.Increase the max_tokens parameter to allow the model more context to correct itself.

C.Disable the temperature parameter to make outputs deterministic.

D.Implement a content filtering layer using a safety classifier to detect and block toxic outputs.

AnswerD

Safety classifiers directly filter toxic content.

Why this answer

Implementing a content filtering layer using a safety classifier is a proven best practice to detect and block toxic or biased outputs in real-time. This approach acts as a guardrail, intercepting harmful responses before they reach users, and is independent of the model's internal parameters or training data.

Exam trap

Oracle often tests the misconception that adjusting model parameters (like temperature or max_tokens) can fix safety issues, when in reality, safety requires external guardrails like content filters.

How to eliminate wrong answers

Option A is wrong because few-shot prompting with examples of toxic responses would not teach the model to avoid them; instead, it could inadvertently reinforce undesirable patterns, as the model may learn to mimic the toxic examples rather than suppress them. Option B is wrong because increasing the max_tokens parameter does not help the model correct its own toxicity; it simply allows longer outputs, which could include more harmful content. Option C is wrong because disabling the temperature parameter (setting it to 0) makes outputs deterministic but does not address the underlying issue of toxic or biased generation; the model can still produce harmful responses consistently.

Practice this question →

20

Multi-Selectmedium

An organization is implementing a RAG system using OCI GenAI. Which two are best practices for optimizing retrieval and generation? (Choose two.)

Select 2 answers

A.Use the same embedding model for both retrieval and generation

B.Store all documents in a single large index

C.Use semantic search (embeddings) for document retrieval

D.Implement caching for frequently asked questions

E.Disable summarization to save inference costs

AnswersC, D

Semantic search captures meaning beyond keywords, improving relevance.

Why this answer

Semantic search using embeddings retrieves documents based on meaning rather than keyword matching, which significantly improves the relevance of context provided to the LLM in a RAG system. This aligns with best practices for OCI GenAI, where embedding models convert text into vector representations for similarity search in a vector database.

Exam trap

Oracle often tests the misconception that retrieval and generation should share the same model, but in practice they are optimized separately, and candidates may confuse 'embedding model' with 'generation model' in a RAG context.

Practice this question →

21

Multi-Selectmedium

A data scientist is evaluating different models for a summarization task. Which two metrics are commonly used to evaluate the quality of generated summaries?

Select 2 answers

A.F1 score

B.Mean Average Precision

C.ROUGE

D.Perplexity

E.BLEU

AnswersC, E

ROUGE measures overlap of n-grams between generated and reference summaries, commonly used for summarization.

Why this answer

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is a standard metric for summarization that measures the overlap of n-grams, word sequences, or word pairs between the generated summary and reference summaries. It focuses on recall, making it well-suited for evaluating how well the generated summary captures the key content from the reference.

Exam trap

Oracle often tests the distinction between metrics used for summarization (ROUGE) versus translation (BLEU) versus language modeling (Perplexity), and candidates may confuse BLEU as a summarization metric because it also evaluates text generation, but it is primarily designed for translation tasks.

Practice this question →

22

MCQhard

Refer to the exhibit. The output is very short and cuts off mid-sentence. Which parameter is most likely the cause?

A.Max-tokens is too low

B.Temperature is too high

C.Model ID incorrect

D.Top-p is too high

AnswerA

If the output exceeds max-tokens, it gets truncated, causing the cut-off.

Why this answer

The 'max-tokens' parameter controls the maximum number of tokens (words or subwords) the model can generate in a single response. When set too low, the model is forced to stop generating output before completing its thought, resulting in a truncated or cut-off sentence. This is the most direct cause of an output that ends abruptly mid-sentence.

Exam trap

Oracle often tests the distinction between parameters that control output length (max-tokens) versus those that control output diversity (temperature, top-p), leading candidates to confuse the effect of randomness with truncation.

How to eliminate wrong answers

Option B is wrong because a high temperature increases randomness in token selection, leading to more creative or divergent responses, but it does not cause premature truncation. Option C is wrong because an incorrect model ID would typically result in a model not found error or unexpected behavior, not a clean but short output. Option D is wrong because a high top-p value allows the model to consider a larger pool of probable tokens, which can make output more varied but does not limit the length of the response.

Practice this question →

23

MCQmedium

A company uses OCI Generative AI service to power a chatbot. After deployment, the chatbot starts generating inappropriate responses. Which action should be taken first?

A.Increase the temperature parameter.

B.Fine-tune the model on customer-specific data.

C.Switch to a larger model.

D.Adjust the prompt template to include safety instructions.

AnswerD

Adding safety instructions in the prompt is a quick and effective safeguard.

Why this answer

Adjusting the prompt template to include safety instructions is the fastest and most direct way to mitigate inappropriate responses without retraining or changing model parameters. In OCI Generative AI, prompt engineering—including explicit safety guidelines—can immediately constrain the model's output behavior by providing clear guardrails in the context window.

Exam trap

Oracle often tests the misconception that safety issues require model retraining or parameter tuning, when in fact prompt engineering is the first-line, low-cost intervention recommended in OCI documentation.

How to eliminate wrong answers

Option A is wrong because increasing the temperature parameter would make the model's output more random and creative, likely worsening inappropriate responses rather than fixing them. Option B is wrong because fine-tuning on customer-specific data requires significant time, cost, and labeled data, and does not directly address safety issues—it is an over-engineered solution for a problem that can be solved with prompt adjustments. Option C is wrong because switching to a larger model does not inherently improve safety; larger models may even generate more complex or unexpected inappropriate content without proper guardrails.

Practice this question →

24

MCQhard

A team uses OCI Generative AI's summarization feature to condense legal documents. The summaries sometimes omit critical clauses. Which parameter adjustment is most likely to improve completeness?

A.Adjust frequencyPenalty.

B.Increase temperature.

C.Decrease topP.

D.Increase maxTokens.

AnswerD

A larger token limit enables longer summaries, helping to include critical clauses.

Why this answer

Increasing maxTokens (option D) is the most direct way to improve completeness because it extends the maximum length of the generated summary, allowing the model to include more content from the source legal document. Critical clauses are often omitted when the token limit truncates the output before the model can cover all essential sections. This parameter controls the output length, not the style or randomness of the generation.

Exam trap

Oracle often tests the misconception that randomness parameters (temperature, topP) or repetition penalties control output length, when in fact only maxTokens directly determines how much text the model can produce.

How to eliminate wrong answers

Option A is wrong because frequencyPenalty reduces repetition by penalizing tokens that have already appeared, which does not address the omission of critical clauses—it only discourages the model from repeating itself. Option B is wrong because increasing temperature adds randomness to token selection, which can make the summary less coherent and more likely to skip important details, not improve completeness. Option C is wrong because decreasing topP narrows the set of candidate tokens to only the most probable ones, which can make the output more conservative and even more likely to omit less common but critical clauses.

Practice this question →

25

Multi-Selecteasy

Which two are essential components of the Transformer architecture? (Select TWO)

Select 2 answers

A.Pooling layers

B.Recurrent connections

C.Self-attention mechanism

D.Feed-forward neural network

E.Convolutional layers

AnswersC, D

Correct: Core component of Transformers.

Why this answer

The self-attention mechanism is essential because it allows each token in the input sequence to attend to every other token, capturing long-range dependencies without the sequential bottleneck of RNNs. This mechanism computes attention scores using queries, keys, and values, enabling parallel processing and forming the core of the Transformer's ability to model context.

Exam trap

Oracle often tests the misconception that Transformers still use recurrence or convolution for sequence processing, when in fact they rely solely on self-attention and feed-forward networks.

Practice this question →

26

MCQmedium

Refer to the exhibit. What is the solution?

A.Use a different base model that supports fine-tuning.

B.Change the learning rate.

C.Increase the training epochs.

D.Use a different compartment.

AnswerA

The error indicates the base model does not support fine-tuning; switch to a supported model.

Why this answer

The exhibit indicates that the base model does not support fine-tuning, which is a prerequisite for adapting a large language model to a specific task or domain. Using a different base model that supports fine-tuning allows the model to be customized through supervised learning on task-specific data, enabling it to learn new patterns and improve performance. This is the correct solution because without fine-tuning capability, the model cannot be effectively adapted regardless of other hyperparameter adjustments.

Exam trap

Oracle often tests the distinction between hyperparameter tuning (learning rate, epochs) and fundamental model capability (fine-tuning support), leading candidates to mistakenly choose a hyperparameter adjustment when the core issue is that the model cannot be fine-tuned at all.

How to eliminate wrong answers

Option B is wrong because changing the learning rate only affects the optimization process during training, but if the base model does not support fine-tuning, no amount of learning rate adjustment will enable the model to be trained on new data. Option C is wrong because increasing the training epochs will not help if the model cannot be fine-tuned at all; epochs only matter when the model is actually being trained or fine-tuned. Option D is wrong because using a different compartment (a tenancy or organizational boundary in Oracle Cloud Infrastructure) does not change the underlying model's architecture or its ability to be fine-tuned; it only affects resource isolation and access control.

Practice this question →

27

MCQhard

During multi-turn conversation with an OCI GenAI model, the model repeats user messages from earlier turns. What is the most likely cause?

A.Low top-p

B.High temperature

C.Low presence penalty

D.High frequency penalty

AnswerC

Low presence penalty means the model is less penalized for repeating topics, leading to repetition.

Why this answer

A low presence penalty reduces the model's incentive to avoid repeating previously mentioned content. In multi-turn conversations, this can cause the model to echo user messages from earlier turns because the penalty is too weak to discourage repetition of tokens that have already appeared in the context window.

Exam trap

Oracle often tests the distinction between presence penalty (which penalizes any occurrence) and frequency penalty (which penalizes based on count), leading candidates to mistakenly think a high frequency penalty causes repetition when it actually prevents it.

How to eliminate wrong answers

Option A is wrong because low top-p limits the cumulative probability mass for token sampling, which reduces diversity but does not directly cause repetition of earlier user messages; it may instead make outputs more deterministic. Option B is wrong because high temperature increases randomness in token selection, which can lead to more creative or even nonsensical outputs, not specifically the repetition of prior user messages. Option D is wrong because a high frequency penalty actively discourages the model from using tokens that have already appeared, which would reduce repetition, not cause it.

Practice this question →

28

MCQhard

A company uses an LLM to generate product descriptions. The outputs are consistently too verbose and include irrelevant details. The prompt includes a simple instruction: 'Describe the product.' Which adjustment to the prompt is most likely to yield concise, relevant descriptions?

A.Set temperature to 0.

B.Increase max_tokens to 500.

C.Add constraints like 'Max 30 words. Focus on key features.'

D.Include a few examples of desired short descriptions.

AnswerC

Explicit constraints directly limit length and scope.

Why this answer

Adding explicit constraints like 'Max 30 words. Focus on key features.' directly instructs the LLM to limit verbosity and prioritize relevant details. This technique, known as prompt engineering with constraints, is the most effective way to control output length and content without altering model parameters or relying on examples that may not generalize.

Exam trap

Oracle often tests the misconception that adjusting model parameters (temperature or max_tokens) is the primary way to control output quality, when in fact prompt engineering with explicit constraints is a more direct and reliable method for achieving specific formatting or length requirements.

How to eliminate wrong answers

Option A is wrong because setting temperature to 0 makes the model deterministic (greedy decoding), which reduces randomness but does not inherently shorten or focus the output—it may still produce verbose descriptions. Option B is wrong because increasing max_tokens to 500 actually allows the model to generate longer responses, which is counterproductive to achieving concise descriptions. Option D is wrong because including a few examples (few-shot prompting) can guide the style but does not guarantee brevity; the model may still extrapolate irrelevant details or exceed the desired length without explicit constraints.

Practice this question →

29

MCQmedium

An e-commerce company fine-tuned a Cohere Command model on their product catalog to generate product descriptions. During inference, they notice the model outputs are too repetitive: it often repeats similar phrases across different products, and the descriptions lack diversity. The team wants to increase the variety of the generated text without sacrificing relevance. They are currently using temperature=0.8, top_p=0.9, frequency_penalty=0, and presence_penalty=0. Which parameter adjustment should they make to most effectively increase diversity?

A.Decrease temperature from 0.8 to 0.5.

B.Set frequency_penalty to a negative value (e.g., -0.5).

C.Increase max_tokens from 200 to 500.

D.Increase top_p from 0.9 to 0.95.

AnswerD

Higher top_p includes more tokens in the sampling pool, increasing diversity.

Why this answer

Increasing top_p from 0.9 to 0.95 expands the nucleus of tokens considered during sampling, allowing the model to select from a wider set of plausible next tokens. This directly increases output diversity while still maintaining relevance, as tokens outside the top 90% probability mass are now included. The current settings already have moderate temperature and no penalties, so broadening top_p is the most effective single adjustment to reduce repetitiveness.

Exam trap

Oracle often tests the misconception that increasing temperature always increases diversity, when in fact decreasing temperature reduces randomness, and the most effective lever for diversity in a fine-tuned model is often adjusting top-p or adding a positive frequency penalty.

How to eliminate wrong answers

Option A is wrong because decreasing temperature from 0.8 to 0.5 makes the model more deterministic, reducing randomness and likely increasing repetitiveness, which is the opposite of the desired outcome. Option B is wrong because setting frequency_penalty to a negative value (e.g., -0.5) encourages the model to repeat tokens, exacerbating the repetitiveness problem rather than solving it. Option C is wrong because increasing max_tokens from 200 to 500 only extends the length of generated text; it does not alter the sampling strategy, so the model will continue to repeat phrases within the longer output.

Practice this question →

30

MCQmedium

A financial institution uses OCI GenAI to power a customer support chatbot. The compliance team requires that responses are strictly consistent with regulatory guidelines and approved responses. The company has a curated set of question-answer pairs that cover common scenarios. They want to ensure that the chatbot never deviates from these approved answers. The data science team is considering various approaches to enforce this consistency. Which approach is most effective?

A.Few-shot prompting with three example responses in every query.

B.Fine-tuning the model on the curated dataset of question-answer pairs.

C.Using a large context window to include all regulatory guidelines in the prompt.

D.Setting a low temperature (0.1) to make outputs deterministic.

AnswerB

Fine-tuning adapts the model to mimic the approved responses, providing strong consistency.

Why this answer

Fine-tuning the model on the curated dataset of approved responses teaches the model to output similar responses for related questions, ensuring consistency. Option A is wrong because few-shot prompting may fail for unseen variations and does not guarantee strict adherence. Option C is wrong because using a large context window does not enforce specific content.

Option D is wrong because setting a low temperature reduces randomness but does not guarantee the model will choose approved responses.

Practice this question →

31

MCQeasy

A company wants to build a customer support chatbot using OCI Generative AI. They have a large number of historical support tickets. Which approach is most effective for leveraging this data to improve the chatbot's responses?

A.Use a pre-loaded prompt template from the OCI console.

B.Fine-tune the Cohere Command model on the historical tickets using OCI Data Science.

C.Increase the temperature parameter to 1.0 to encourage diverse responses.

D.Use zero-shot prompting with the base model and include few-shot examples in the prompt.

AnswerB

Fine-tuning on the company's own support tickets adapts the model to the specific language, context, and resolutions, significantly improving response quality.

Why this answer

Fine-tuning the Cohere Command model on the historical support tickets using OCI Data Science is the most effective approach because it adapts the model's weights to the specific domain language, terminology, and resolution patterns found in the company's data. This supervised learning process creates a specialized model that can generate accurate, context-aware responses for customer support queries, unlike generic prompting methods that lack deep domain adaptation.

Exam trap

Oracle often tests the misconception that increasing temperature or using few-shot examples can substitute for fine-tuning when adapting a model to proprietary domain data, but in reality only fine-tuning modifies model weights to deeply learn domain-specific patterns from large datasets.

How to eliminate wrong answers

Option A is wrong because pre-loaded prompt templates in the OCI console are generic and not trained on the company's specific historical ticket data, so they cannot capture domain-specific nuances or improve response accuracy beyond basic instruction following. Option C is wrong because increasing the temperature parameter to 1.0 maximizes randomness in token selection, which reduces coherence and factual reliability—exactly the opposite of what is needed for a customer support chatbot that requires consistent, accurate answers. Option D is wrong because zero-shot prompting with few-shot examples only provides a few static examples in the context window, which does not modify the model's underlying weights and cannot match the depth of learning achieved by fine-tuning on thousands of historical tickets.

Practice this question →

32

MCQhard

A data scientist observes that their fine-tuned LLM performs well on training data but generates repetitive and dull responses in production. What is the most likely cause and best solution?

A.The model is overfitted; apply stronger regularization

B.The temperature is set too low; increase temperature during inference

C.The training data lacks diversity; add more varied examples

D.The model has too many layers; reduce model size

AnswerB

Low temperature makes outputs deterministic and repetitive; increasing it adds variability.

Why this answer

The model's repetitive and dull responses indicate that the temperature parameter is too low, causing the model to always select the most probable tokens, leading to deterministic and monotonous outputs. Increasing temperature during inference introduces randomness into token sampling, allowing for more diverse and creative responses. This is a common issue in production LLMs where low temperature settings optimized for training metrics fail to produce engaging real-world outputs.

Exam trap

Oracle often tests the misconception that poor production performance is always due to overfitting or data issues, when in fact inference-time hyperparameters like temperature are the direct cause of repetitive/dull outputs.

How to eliminate wrong answers

Option A is wrong because overfitting would cause poor generalization to new inputs, not specifically repetitive/dull outputs; regularization reduces overfitting but does not address the deterministic token selection caused by low temperature. Option C is wrong because while training data diversity affects model knowledge, the described symptom of repetitive outputs in production despite good training performance points to inference-time sampling issues, not data diversity. Option D is wrong because having too many layers might cause overfitting or computational inefficiency, but it does not directly cause repetitive or dull responses; reducing model size would not fix the temperature-related sampling behavior.

Practice this question →

33

MCQhard

A team is fine-tuning a large language model for a domain-specific Q&A application. After fine-tuning, they observe that the model performs well on the training distribution but struggles with out-of-distribution (OOD) questions. Which approach would best improve OOD robustness?

A.Include a diverse set of examples from related domains in the fine-tuning dataset.

B.Use early stopping based on training loss to avoid overfitting.

C.Reduce the model size to prevent overfitting to the training data.

D.Increase the learning rate during fine-tuning to adapt faster to new patterns.

AnswerA

Diverse data improves generalization and OOD performance.

Why this answer

Including diverse examples from related domains (Option A) exposes the model to varied patterns, reducing overfitting to the training distribution and improving generalization to out-of-distribution inputs. Option B (early stopping) helps prevent overfitting on the validation set but does not specifically address OOD robustness. Option C (reducing model size) limits capacity, which can harm performance on both in-distribution and OOD data.

Option D (increasing learning rate) risks catastrophic forgetting and training instability, not OOD robustness.

Practice this question →

34

MCQeasy

A developer is testing the OCI Generative AI API by sending a request to generate text using the Cohere Command R model. The request returns the following error: 'The model 'cohere.command-r-08-2024' is not available in this region. Please check the model availability in your region.' The developer is using the us-ashburn-1 region. What is the most likely cause of this error?

A.The request body format is incorrect.

B.The model is not deployed in the us-ashburn-1 region.

C.The model name is misspelled (e.g., 'cohere.command-r-08-2024' vs 'cohere.command-r-08-2024').

D.The API key used in the request is invalid.

AnswerB

Cohere Command R may not be available in all regions; check supported regions in OCI documentation.

Why this answer

The error message explicitly states that the model 'cohere.command-r-08-2024' is not available in the region. OCI Generative AI models are deployed regionally, and the Cohere Command R model is not available in the us-ashburn-1 (Ashburn) region. The developer must select a supported region, such as us-chicago-1, where this model is deployed.

Exam trap

Oracle often tests the misconception that model names must be perfectly spelled or that API keys are the cause of all errors, but here the trap is that candidates overlook regional availability and assume the error is due to a typo or authentication failure.

How to eliminate wrong answers

Option A is wrong because an incorrect request body format would typically result in a 400 Bad Request or validation error, not a model availability error. Option C is wrong because the model name in the error matches the one sent, so a misspelling would cause a different error (e.g., 'model not found'), not a region availability error. Option D is wrong because an invalid API key would result in a 401 Unauthorized or 403 Forbidden error, not a model availability error.

Practice this question →

35

MCQmedium

A company is deploying a large language model for a customer service chatbot. The model needs to understand industry-specific jargon and maintain low latency. Which approach best balances these requirements?

A.Employ retrieval-augmented generation (RAG) with a general model

B.Rely solely on prompt engineering with a general model

C.Use a large general-purpose LLM with zero-shot prompting

D.Fine-tune a small open-source LLM on domain-specific data

AnswerD

Fine-tuning adapts the model to jargon and a smaller model keeps latency low.

Why this answer

Fine-tuning a small open-source LLM on domain-specific data is the best approach because it adapts the model to understand industry-specific jargon while keeping the model small enough to maintain low latency. Unlike larger models, a fine-tuned small model can run efficiently on local hardware, reducing inference time and avoiding the overhead of external API calls or large model sizes.

Exam trap

Oracle often tests the misconception that larger models always perform better or that RAG alone solves domain adaptation, ignoring the latency and efficiency trade-offs that make fine-tuning a smaller model the optimal choice for production systems with strict response time requirements.

How to eliminate wrong answers

Option A is wrong because retrieval-augmented generation (RAG) with a general model still relies on a general model that may not inherently understand industry-specific jargon, and the retrieval step adds latency, which conflicts with the low-latency requirement. Option B is wrong because relying solely on prompt engineering with a general model does not embed domain-specific knowledge into the model weights, so the model may still misinterpret or fail to generate accurate responses for niche jargon, and it often requires longer prompts that increase latency. Option C is wrong because a large general-purpose LLM with zero-shot prompting has high inference latency due to its size and lacks domain-specific training, making it unsuitable for both understanding jargon and meeting low-latency constraints.

Practice this question →

36

MCQeasy

A developer is using a large language model to generate code snippets. The model often produces code that is syntactically correct but functionally incorrect. What is the most effective way to improve the functional correctness of the generated code?

A.Provide few-shot examples of correct code in the prompt.

B.Increase the temperature parameter to generate more creative solutions.

C.Ask the model to only output syntactically valid code.

D.Set max_tokens to a very high value to allow the model more room to think.

AnswerA

Few-shot examples help the model understand the expected output.

Why this answer

Providing few-shot examples of correct code in the prompt directly demonstrates the desired functional behavior to the model. This technique, known as few-shot prompting, grounds the model's output in concrete examples, significantly improving the likelihood that the generated code will be functionally correct by aligning the model's pattern completion with the intended logic, not just syntax.

Exam trap

Oracle often tests the misconception that increasing model parameters like temperature or max_tokens can improve output quality, when in fact these parameters control randomness and length, not functional correctness, which is best addressed through prompt engineering techniques like few-shot learning.

How to eliminate wrong answers

Option B is wrong because increasing the temperature parameter makes the model's output more random and creative, which typically reduces functional correctness by increasing the chance of generating plausible but incorrect logic. Option C is wrong because asking the model to only output syntactically valid code does not address functional correctness; the model already generates syntactically valid code by default, and this instruction does not guide it toward correct logic. Option D is wrong because setting max_tokens to a very high value does not improve reasoning quality; it only allows longer outputs, which can actually increase the risk of generating more irrelevant or incorrect code without improving functional correctness.

Practice this question →

37

MCQmedium

A customer support company uses Cohere Command on OCI to answer user queries. They have enabled grounding with a knowledge base of product manuals. However, for about 20% of queries, the model provides incorrect product recommendations that are not in the manuals. The team has verified the knowledge base is up to date. What is the most likely cause and solution?

A.The model's temperature is too high, causing creative responses. Lower temperature to 0.

B.The model is hallucinating; switch to a larger model.

C.The query phrasing may not match the knowledge base; improve the retrieval system or use query rewriting.

D.The grounding settings are too restrictive; increase the number of retrieved documents.

AnswerC

Correct: Query mismatch causes retrieval of irrelevant content, leading to incorrect recommendations.

Why this answer

The most likely cause is that the query phrasing does not align well with the knowledge base content, leading to retrieval of irrelevant or no documents. Even with an up-to-date knowledge base, if the query is phrased differently from the manual's text, grounding may fail to retrieve the correct information. Improving the retrieval system (e.g., using semantic search) or implementing query rewriting can bridge this gap, enabling the model to use the appropriate context and provide accurate recommendations.

Option A might help if temperature were high, but the core issue is retrieval, not creativity. Option B (switching to a larger model) does not address the retrieval problem. Option D (increasing retrieved documents) could introduce more noise if the retrieved documents are already irrelevant.

Practice this question →

38

MCQeasy

Refer to the exhibit. What is the primary reason the response is incomplete?

A.The temperature is not set.

B.The model-id is incorrect.

C.The max-tokens limit is too low.

D.The prompt is too short.

AnswerC

Setting max-tokens to 100 restricts the output length, causing truncation.

Why this answer

The response is incomplete because the max-tokens limit is too low, causing the model to truncate its output before completing the full answer. When the token budget is exhausted, the generation stops mid-sentence or mid-thought, leaving the response unfinished regardless of prompt length or other parameters.

Exam trap

Oracle often tests the distinction between parameters that affect output quality (temperature, top_p) versus those that constrain output length (max_tokens, stop sequences), and the trap here is that candidates mistake a short prompt or missing temperature for the cause of truncation when the real culprit is the token budget.

How to eliminate wrong answers

Option A is wrong because the temperature parameter controls randomness in token selection, not the length or completeness of the response; a missing temperature would default to 1.0 and still allow full output. Option B is wrong because the model-id identifies which LLM to use (e.g., gpt-3.5-turbo or cohere.command-text-v14) and does not affect whether the response is truncated; an incorrect model-id would either fail to load or produce different output, not an incomplete one. Option D is wrong because a short prompt can still yield a complete response; prompt length influences context and relevance, but the max-tokens parameter is the direct limiter of output length.

Practice this question →

39

MCQmedium

A company is deploying a large language model in a customer-facing chatbot. The model's responses must be both accurate and safe. Which combination of techniques should be employed?

A.Use only a system prompt instructing the model to be accurate and safe.

B.Use retrieval-augmented generation (RAG) for factual accuracy and a content safety filter for safe outputs.

C.Use a high temperature for creativity and a safety classifier for blocking toxic outputs.

D.Fine-tune the model on all historical chat logs and use a high temperature.

AnswerB

RAG improves accuracy; safety filter ensures safety.

Why this answer

RAG grounds the model's responses in a verified external knowledge base, reducing hallucinations and improving factual accuracy, while a content safety filter (e.g., a classifier or guardrail) actively blocks toxic or unsafe outputs before they reach the user. This combination addresses both accuracy and safety independently, unlike a single system prompt which is easily bypassed.

Exam trap

Oracle often tests the misconception that a single technique (like a system prompt or fine-tuning) can simultaneously guarantee both accuracy and safety, when in practice they require separate, complementary mechanisms.

How to eliminate wrong answers

Option A is wrong because a system prompt alone is a static instruction that can be overridden by user input or model behavior, providing no enforcement mechanism for accuracy or safety. Option C is wrong because a high temperature increases randomness and creativity, which is counterproductive for accuracy and can amplify unsafe outputs; a safety classifier is a partial solution but does not address factual grounding. Option D is wrong because fine-tuning on all historical chat logs may introduce biases, errors, or unsafe patterns from the data, and a high temperature further degrades reliability.

Practice this question →

40

MCQmedium

A team is fine-tuning an LLM on OCI Generative AI for a domain-specific task. They have a dataset of 10,000 labeled examples. What is a best practice to avoid catastrophic forgetting during fine-tuning?

A.Increase the learning rate to speed up adaptation.

B.Use only the new domain-specific data for fine-tuning.

C.Reduce the number of training epochs to the minimum.

D.Include a small percentage of general-domain data in the training mix.

AnswerD

General data acts as a regularizer to maintain base knowledge.

Why this answer

Catastrophic forgetting occurs when a fine-tuned model loses previously learned general knowledge. By including a small percentage (e.g., 5–10%) of general-domain data in the training mix, the model retains its broad capabilities while adapting to the new domain-specific task. This technique, often called 'replay' or 'experience replay,' is a standard practice in continual learning for LLMs.

Exam trap

Oracle often tests the misconception that fine-tuning should exclusively use the new dataset, whereas the best practice is to blend in general data to preserve prior knowledge.

How to eliminate wrong answers

Option A is wrong because increasing the learning rate can cause the model to overfit to the new domain data and accelerate forgetting, not prevent it. Option B is wrong because using only domain-specific data removes all exposure to general knowledge, which is the primary cause of catastrophic forgetting. Option C is wrong because reducing epochs to the minimum may prevent the model from learning the new task adequately, but it does not address the retention of general knowledge; the model can still forget if the new data dominates the gradient updates.

Practice this question →

41

Multi-Selecteasy

Which TWO are advantages of using LoRA for fine-tuning?

Select 2 answers

A.Requires less GPU memory

B.Guarantees higher accuracy

C.Reduces number of trainable parameters

D.Increases model size

E.Improves inference speed

AnswersA, C

Fewer trainable parameters means lower memory usage during training.

Why this answer

LoRA (Low-Rank Adaptation) reduces GPU memory requirements because it freezes the original model weights and injects trainable low-rank matrices into specific layers. This means only a tiny fraction of parameters need gradients and optimizer states, drastically lowering memory consumption during fine-tuning compared to full fine-tuning.

Exam trap

Oracle often tests the misconception that reducing trainable parameters automatically improves inference speed, but LoRA's memory and parameter savings apply only to training, not to inference latency.

Practice this question →

42

MCQmedium

An LLM-based application must comply with data privacy regulations by not memorizing personally identifiable information (PII). Which technique best reduces memorization of PII?

A.Use a larger model with more parameters

B.Decrease the temperature during inference

C.Train with differential privacy

D.Increase the number of training epochs

AnswerC

Differential privacy bounds the influence of any single data point, reducing memorization.

Why this answer

Differential privacy (DP) is the correct technique because it directly limits the model's ability to memorize training data, including PII, by adding calibrated noise to the gradient updates during training. This ensures that the model's parameters do not encode specific individual records, providing a formal mathematical guarantee against memorization. Other options like model size, temperature, or training epochs do not address the root cause of memorization in the training process.

Exam trap

Oracle often tests the misconception that inference-time parameters like temperature or model size affect training data memorization, when in fact memorization is a training-phase phenomenon that must be addressed during training itself.

How to eliminate wrong answers

Option A is wrong because increasing model parameters generally increases the model's capacity to memorize training data, making PII leakage more likely, not less. Option B is wrong because temperature controls the randomness of output token sampling during inference, not the memorization of training data; it has no effect on whether PII is stored in the model weights. Option D is wrong because increasing training epochs typically leads to overfitting and greater memorization of training examples, including PII, as the model sees the data more times.

Practice this question →

43

MCQmedium

A company wants to deploy a private instance of a large language model on OCI for sensitive data processing. What is the recommended approach?

A.Use OCI Data Science with a publicly accessible model.

B.Use the OCI Generative AI public endpoint with data encryption.

C.Use OCI Dedicated AI Clusters for a private endpoint.

D.Use third-party model hosting outside OCI.

AnswerC

Dedicated AI Clusters offer isolated compute with private networking, meeting security and compliance needs.

Why this answer

OCI Dedicated AI Clusters provide a fully isolated, private endpoint for deploying large language models, ensuring that sensitive data never traverses the public internet. This approach meets the requirement for private inference with no data leaving the customer's tenancy, unlike public endpoints or third-party hosting.

Exam trap

Oracle often tests the misconception that encryption alone (Option B) is sufficient for private deployment, but the trap is that encryption does not eliminate the need for network isolation when processing sensitive data in a shared infrastructure environment.

How to eliminate wrong answers

Option A is wrong because OCI Data Science with a publicly accessible model exposes the model endpoint to the internet, violating the requirement for private, sensitive data processing. Option B is wrong because the OCI Generative AI public endpoint, even with data encryption, still routes traffic through OCI's shared infrastructure and public IP space, which does not guarantee the isolation needed for sensitive data. Option D is wrong because third-party model hosting outside OCI would require data to leave the OCI tenancy, breaking the requirement for a private deployment within OCI.

Practice this question →

44

Multi-Selecthard

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

Select 3 answers

A.Bias in training data perpetuating stereotypes

B.High computational cost for inference

C.Hallucination of plausible but incorrect information

D.Fast inference speed due to parallelization

E.Low memory footprint

AnswersA, B, C

Models can reflect and amplify biases from training data.

Why this answer

Large language models (LLMs) are trained on vast, unfiltered internet text corpora that inherently contain societal biases. These biases are learned and can be amplified during inference, leading to outputs that perpetuate harmful stereotypes, which is a well-documented production challenge.

Exam trap

Oracle often tests the distinction between known challenges (bias, cost, hallucination) and desirable properties (fast inference, low memory) that are actually false for LLMs, trapping candidates who confuse optimization goals with current limitations.

Practice this question →

45

MCQhard

A development team wants to generate code snippets from natural language. Which model strategy should they adopt?

A.Use a code-specific model like Code Llama.

B.Use a general-purpose LLM like Llama 2.

C.Use a multimodal model.

D.Use an embedding model for text.

AnswerA

Correct: Code-specific models are fine-tuned for code generation.

Why this answer

Code Llama is a specialized variant of Llama 2 that has been fine-tuned on code datasets, enabling it to generate syntactically and semantically correct code from natural language prompts. This makes it the optimal choice for code generation tasks, as general-purpose LLMs lack the targeted training on code structures and programming languages.

Exam trap

Oracle often tests the distinction between general-purpose and domain-specific models, and the trap here is that candidates assume any large language model can handle code generation equally well, overlooking the critical fine-tuning on code corpora that makes Code Llama superior for this task.

How to eliminate wrong answers

Option B is wrong because a general-purpose LLM like Llama 2 is trained on diverse text but not specifically optimized for code, leading to higher rates of syntax errors and logical inconsistencies in generated code. Option C is wrong because multimodal models process images, audio, and text, but code generation from natural language does not require multiple modalities and would add unnecessary complexity without improving code quality. Option D is wrong because embedding models are designed to convert text into vector representations for similarity search or clustering, not for generating new text or code snippets.

Practice this question →

46

MCQmedium

Refer to the exhibit. A developer ran the OCI CLI command shown and received the JSON output. What does the output indicate about the model's confidence and why?

A.The model is uncertain because all scores are roughly equal.

B.The model is neutral because the neutral score is lowest.

C.The model is unsure because the scores are probabilities that sum to 1.

D.The model is highly confident the text is positive, as indicated by the 0.98 score.

AnswerD

0.98 is very close to 1, indicating high confidence.

Why this answer

The JSON output shows a sentiment score of 0.98 for 'positive', which is very close to 1.0, indicating the model is highly confident that the text is positive. In sentiment analysis models, scores represent probabilities for each class, and a value near 1.0 for one class with much lower scores for others reflects strong confidence.

Exam trap

Oracle often tests the distinction between the sum of probabilities equaling 1 (a mathematical property) and the actual confidence level indicated by the distribution of those probabilities, leading candidates to mistakenly choose option C.

How to eliminate wrong answers

Option A is wrong because the scores are not roughly equal; the positive score (0.98) is significantly higher than the negative (0.01) and neutral (0.01) scores, indicating high confidence, not uncertainty. Option B is wrong because the neutral score being lowest does not imply neutrality; the model is highly confident the text is positive, not neutral. Option C is wrong because while the scores do sum to 1 (as probabilities should), this fact alone does not indicate uncertainty; the distribution of probabilities matters, and here the high positive score shows confidence.

Practice this question →

47

MCQmedium

A data scientist is using OCI Data Science with the Generative AI service to fine-tune a Cohere Command model on a custom dataset of customer support tickets. After training, the model produces poor, irrelevant responses. What is the most likely cause?

A.Incorrect tokenizer configuration

B.Insufficient training data quality or quantity

C.Too many epochs causing overfitting

D.Model architecture mismatch between fine-tuned and base model

AnswerB

Cohere models need clean, diverse, and task-relevant data; poor data leads to poor fine-tuning.

Why this answer

Insufficient training data quality or quantity is the most likely cause because fine-tuning a Cohere Command model on a custom dataset of customer support tickets requires a sufficiently large and representative dataset to teach the model domain-specific patterns. If the dataset is too small, noisy, or lacks diversity, the model will fail to generalize and produce irrelevant responses, even with correct tokenization and training hyperparameters.

Exam trap

Oracle often tests the misconception that overfitting (Option C) is the primary cause of poor model output after fine-tuning, but in this scenario the irrelevance points to data insufficiency rather than memorization of training examples.

How to eliminate wrong answers

Option A is wrong because incorrect tokenizer configuration would typically cause tokenization errors or mismatched vocabulary, not poor semantic relevance; the Cohere Command model uses a fixed tokenizer that is automatically applied during fine-tuning in OCI Data Science. Option C is wrong because too many epochs causing overfitting would result in the model memorizing training examples and producing overly specific or repetitive responses, not generally irrelevant ones; overfitting typically degrades performance on unseen data but does not cause broad irrelevance. Option D is wrong because model architecture mismatch between fine-tuned and base model is not possible in OCI Data Science's Generative AI service, as the fine-tuning process uses the same architecture as the base model; the service enforces compatibility.

Practice this question →

48

Multi-Selecthard

Which three characteristics of LLMs can lead to hallucinations? (Select THREE)

Select 3 answers

A.Overconfidence in predictions

B.Ability to generate plausible-sounding text

C.Lack of real-world grounding

D.Gaps in training data coverage

E.Large vocabulary size

AnswersB, C, D

Correct: Fluency can mask inaccuracies.

Why this answer

LLMs are trained to generate text that is statistically plausible and coherent, but they lack mechanisms to verify factual accuracy. This means they can produce sentences that sound convincing and grammatically correct while being entirely false, which is a direct cause of hallucinations.

Exam trap

Oracle often tests the distinction between symptoms and root causes, so the trap here is that candidates might confuse 'overconfidence in predictions' (a symptom) with a direct cause of hallucinations, or mistakenly think 'large vocabulary size' contributes to hallucinations when it is merely an enabler of the model's generative capability.

Practice this question →

49

MCQeasy

A data scientist is using a large language model to summarize customer support tickets. The model occasionally generates summaries that include hallucinated details not present in the original ticket. Which technique would best reduce hallucinations while maintaining summary quality?

A.Implement retrieval-augmented generation (RAG) to ground the model in relevant documents.

B.Use a longer system prompt instructing the model to be factual.

C.Fine-tune the model on a large corpus of general text to improve its knowledge.

D.Increase the temperature parameter to 0.9 to encourage more deterministic outputs.

AnswerA

RAG provides factual context, reducing hallucinations.

Why this answer

Retrieval-Augmented Generation (RAG) reduces hallucinations by grounding the model's output in external, verifiable documents retrieved from a knowledge base. Instead of relying solely on the model's parametric memory, RAG fetches relevant context (e.g., the original ticket) at inference time, ensuring the summary is factually aligned with the source. This maintains summary quality because the model can still generate fluent text while being constrained to the retrieved evidence.

Exam trap

Oracle often tests the misconception that simply instructing the model to be factual (Option B) or fine-tuning (Option C) can eliminate hallucinations, when in reality grounding via retrieval (RAG) is the only technique that directly supplies external evidence to constrain generation.

How to eliminate wrong answers

Option B is wrong because a longer system prompt instructing the model to be factual does not provide new factual data; it only changes the model's behavior via instruction tuning, which cannot correct hallucinations stemming from missing or incorrect parametric knowledge. Option C is wrong because fine-tuning on a large corpus of general text would not specifically address hallucinations in customer support tickets; it might even dilute domain-specific accuracy and does not provide a retrieval mechanism to verify facts. Option D is wrong because increasing the temperature parameter to 0.9 actually increases randomness and creativity, making outputs less deterministic and more prone to hallucination, not less.

Practice this question →

50

MCQhard

A company is using OCI GenAI with a Dedicated AI Cluster to serve a large language model for real-time chat applications. They notice high inference latency (average 2 seconds per response) and want to reduce it to under 500 milliseconds without significantly degrading the quality of responses. The cluster is configured with NVIDIA A100 GPUs. The model is the base Cohere Command model (52B parameters). They have explored increasing batch size, but that increases latency for interactive use cases. Which action should they take?

A.Deploy the model with inference optimization frameworks like vLLM, TensorRT, or ONNX Runtime.

B.Increase batch size to process multiple queries at once.

C.Swap the model to a smaller variant, such as Cohere Command Light (6B).

D.Enable model quantization (e.g., int8) to reduce memory and computation.

AnswerA

These frameworks optimize GPU utilization and reduce latency without changing the model.

Why this answer

Inference optimization frameworks like vLLM, TensorRT, and ONNX Runtime are specifically designed to reduce latency for large language models on NVIDIA A100 GPUs. These frameworks use techniques such as PagedAttention (vLLM), kernel fusion, and graph optimization to significantly lower per-request latency without degrading output quality, making them ideal for real-time chat applications where sub-500ms responses are required.

Exam trap

Oracle often tests the misconception that model quantization or smaller models are the only ways to reduce latency, but the trap here is that inference optimization frameworks can achieve dramatic latency reductions without sacrificing model quality or capability.

How to eliminate wrong answers

Option B is wrong because increasing batch size improves throughput but increases per-request latency, which is counterproductive for interactive use cases requiring low latency. Option C is wrong because swapping to a smaller model (Cohere Command Light, 6B) would reduce latency but also significantly degrade response quality and capability, which the question explicitly wants to avoid. Option D is wrong because enabling model quantization (e.g., int8) reduces memory and computation, which can lower latency, but it often introduces a trade-off in model accuracy and may not achieve the target latency on its own without combining with inference optimization frameworks; the question asks for the best single action, and optimization frameworks directly target latency reduction more effectively.

Practice this question →

51

MCQmedium

A company wants to create a chatbot that answers questions based on a large internal document set that is updated weekly. They have limited ML expertise. Which approach is recommended?

A.Fine-tune a model on the entire document set.

B.Train a custom model from scratch.

C.Include all documents in the system prompt.

D.Use retrieval-augmented generation (RAG) with a vector database.

AnswerD

Correct: RAG handles dynamic data without retraining.

Why this answer

Retrieval-Augmented Generation (RAG) with a vector database is the recommended approach because it allows the chatbot to answer questions based on a large, frequently updated document set without requiring model retraining. RAG retrieves relevant document chunks at query time using vector similarity search, then passes them as context to the LLM, ensuring up-to-date answers with minimal ML expertise.

Exam trap

The exam often tests the misconception that fine-tuning is the only way to incorporate custom data, when in fact RAG is the preferred method for dynamic, large-scale document sets due to its cost-effectiveness, ease of updates, and lower ML expertise requirements.

How to eliminate wrong answers

Option A is wrong because fine-tuning a model on the entire document set would require significant ML expertise, computational resources, and would need to be repeated weekly to incorporate updates, making it impractical for a dynamically changing corpus. Option B is wrong because training a custom model from scratch is extremely resource-intensive, requires deep ML expertise, and is unnecessary when pre-trained LLMs can be leveraged with RAG. Option C is wrong because including all documents in the system prompt would exceed the LLM's context window limits (typically 4K-128K tokens), causing truncation, high latency, and increased cost, while also failing to scale with a large document set.

Practice this question →

52

MCQmedium

A company uses OCI GenAI to build a content moderation system that filters toxic language in user-generated comments. They have a small labeled dataset of 1,000 comments (500 toxic, 500 non-toxic) and need an efficient solution that balances accuracy, cost, and latency. They are considering different model options: fine-tuning a large LLM (e.g., Cohere Command), using a pre-trained LLM with prompting, fine-tuning a smaller BERT-based classifier, or building a rule-based system. The team has moderate ML experience and wants to deploy using OCI Data Science. Which approach is most efficient for this binary classification task?

A.Fine-tune a BERT-based classifier (e.g., 'bert-base-uncased') on the dataset.

B.Develop a rule-based system using regular expressions and keyword lists.

C.Use a pre-trained LLM with a toxic/non-toxic prompt.

D.Fine-tune the Cohere Command model on the labeled dataset.

AnswerA

BERT is efficient for classification, fine-tunes quickly on small data, and has low inference cost.

Why this answer

Fine-tuning a BERT-based classifier (e.g., 'bert-base-uncased') is the most efficient approach because BERT is specifically designed for text classification tasks, requiring far fewer computational resources and lower latency than large LLMs. With only 1,000 labeled samples, BERT can achieve high accuracy through transfer learning, while keeping inference costs minimal—ideal for a production content moderation system on OCI Data Science.

Exam trap

Oracle often tests the misconception that larger LLMs (like Cohere Command) are always superior for classification tasks, ignoring the practical constraints of small datasets, cost, and latency that make fine-tuned BERT models the optimal choice for binary classification.

How to eliminate wrong answers

Option B is wrong because rule-based systems using regex and keyword lists cannot generalize to nuanced toxic language (e.g., sarcasm, misspellings, or context-dependent toxicity) and require constant manual maintenance, leading to poor accuracy and high operational overhead. Option C is wrong because using a pre-trained LLM with prompting (e.g., Cohere Command) incurs high per-token inference costs and latency, and with only 1,000 examples, few-shot prompting may not reliably capture the specific toxicity patterns in the dataset. Option D is wrong because fine-tuning a large LLM like Cohere Command on a tiny dataset of 1,000 samples risks catastrophic forgetting and overfitting, while also being computationally expensive and slower for real-time moderation compared to a smaller BERT model.

Practice this question →

53

Multi-Selecteasy

Which THREE are essential steps in the prompt engineering process for an LLM?

Select 3 answers

A.Test the prompt with a variety of input examples

B.Fine-tune the model on a domain corpus

C.Define the desired output format and constraints

D.Quantize the model to INT8

E.Iteratively refine the prompt based on model responses

AnswersA, C, E

Testing ensures robustness across different inputs.

Why this answer

Testing the prompt with a variety of input examples is essential to evaluate the LLM's generalization, robustness, and sensitivity to different phrasing or contexts. This step helps identify edge cases, biases, or inconsistencies in the model's responses before deployment.

Exam trap

Oracle often tests the distinction between prompt engineering (input-side optimization) and model modification (fine-tuning, quantization) to trap candidates who confuse these fundamentally different processes.

Practice this question →

54

MCQmedium

A developer uses OCI Generative AI's chat endpoint with a system message placed after user messages. The model ignores the system message. What is the most likely reason?

A.The system message is too long

B.Temperature is set too high

C.The model has not been fine-tuned for instruction following

D.The system message is placed after user messages

AnswerD

The standard order is system first, then user; otherwise the model may misinterpret.

Why this answer

In OCI Generative AI's chat endpoint, the system message must be placed before user messages to establish the model's behavior and context. When placed after user messages, the model treats it as part of the conversation history rather than a directive, causing it to be ignored. This ordering is a fundamental requirement for the chat API's message structure.

Exam trap

Oracle often tests the specific API message ordering requirement, where candidates mistakenly attribute the failure to model limitations or hyperparameters rather than the structural placement of the system message.

How to eliminate wrong answers

Option A is wrong because the system message being too long would cause a token limit error or truncation, not silent ignoring. Option B is wrong because temperature controls randomness in output, not whether instructions are followed; a high temperature might produce varied responses but does not cause the model to ignore the system message. Option C is wrong because OCI Generative AI models are pre-trained for instruction following without requiring fine-tuning; the issue is purely about message ordering, not model capability.

Practice this question →

55

MCQmedium

A company uses OCI Generative AI to power a chatbot for customer support. They notice that the model's responses sometimes contain factual inaccuracies. Which strategy would best reduce hallucination?

A.Implementing Retrieval-Augmented Generation (RAG).

B.Increasing the temperature parameter.

C.Reducing the max token limit.

D.Fine-tuning the model on a larger general corpus.

AnswerA

RAG retrieves relevant facts from a knowledge base, grounding the output and reducing hallucination.

Why this answer

Retrieval-Augmented Generation (RAG) grounds the model's responses in retrieved factual information, directly reducing hallucination. Increasing temperature increases randomness, fine-tuning on a larger corpus may not fix factual accuracy, and reducing max tokens does not affect correctness.

Practice this question →

56

Multi-Selectmedium

Which TWO techniques are commonly used to reduce the memory footprint of LLM inference?

Select 2 answers

A.Quantization

B.Increasing batch size

C.KV cache optimization

D.Gradient checkpointing

E.Using full precision (FP32)

AnswersA, C

Reduces memory by using lower precision weights.

Why this answer

Quantization reduces the memory footprint by lowering the precision of model weights and activations from FP32 to lower bit-widths like INT8 or FP16, which directly decreases the memory required to store and compute with the model. KV cache optimization reduces memory usage by efficiently managing the key-value cache during autoregressive decoding, often through techniques like shared memory, pruning, or compression, which is critical for long-context inference.

Exam trap

Oracle often tests the distinction between training and inference techniques, so candidates mistakenly apply gradient checkpointing (a training memory saver) to inference, or confuse batch size scaling with memory reduction.

Practice this question →

57

Multi-Selectmedium

Which TWO factors are most likely to cause hallucinations in LLMs?

Select 2 answers

A.High temperature

B.Short context window

C.Excessive fine-tuning

D.Low top-p

E.Inadequate training data

AnswersA, E

High temperature increases randomness, leading to less factual outputs.

Why this answer

A high temperature setting increases the randomness of token sampling, making the model more likely to generate plausible-sounding but factually incorrect or nonsensical outputs. This directly contributes to hallucinations by encouraging the model to deviate from the most probable, grounded responses.

Exam trap

Oracle often tests the misconception that low top-p or short context windows are primary causes of hallucinations, when in fact high temperature and insufficient training data are the two most direct factors that increase the likelihood of generating false or fabricated content.

Practice this question →

58

Multi-Selecteasy

Which TWO techniques can help reduce bias in LLM outputs?

Select 2 answers

A.Setting temperature to 0

B.Using only English data

C.Using diverse training data

D.Increasing model size

E.Applying adversarial debiasing

AnswersC, E

Diverse data reduces representation bias.

Why this answer

Using diverse training data helps the model learn from a wide range of perspectives, reducing the risk of over-representing any single group or viewpoint. This directly mitigates bias by ensuring the training distribution is more representative of the real world, rather than skewed toward a dominant demographic or cultural norm.

Exam trap

Oracle often tests the misconception that lowering temperature or increasing model size can fix bias, when in reality these parameters affect randomness and capacity, not the underlying distributional fairness of the training data.

Practice this question →

59

Multi-Selecthard

A developer is evaluating OCI GenAI model families. Which three are correct characteristics of the available models? (Choose three.)

Select 3 answers

A.Llama models are open-source and available for fine-tuning

B.All models support real-time streaming of tokens

C.Cohere embedding models produce vector representations

D.OCI GenAI provides both hosted and dedicated deployment options

E.Cohere Command models are optimized for multilingual tasks

AnswersA, C, D

Meta's Llama models are open-source and supported by OCI GenAI for fine-tuning.

Why this answer

Llama models, such as Llama 2 and Llama 3, are open-source large language models originally developed by Meta. OCI GenAI provides them as pre-built models that developers can fine-tune using their own datasets, enabling customization for domain-specific tasks without training from scratch.

Exam trap

Oracle often tests the misconception that all models in a platform share the same capabilities, such as streaming or multilingual optimization, when in reality each model family (e.g., Llama, Cohere Command, Cohere Embed) has distinct design goals and feature sets.

Practice this question →

60

MCQhard

A company fine-tunes an LLM on internal support tickets. After deployment, the model hallucinates company-specific product names. What is the most effective mitigation?

A.Switch to a smaller model to reduce hallucination risk

B.Use prompt engineering to remind the model to be accurate

C.Implement RAG with a verified product database

D.Fine-tune further with more ticket data

AnswerC

RAG provides factual grounding, reducing hallucinations.

Why this answer

RAG (Retrieval-Augmented Generation) grounds the LLM's output in a verified product database, providing factual context that prevents hallucination of company-specific product names. Unlike fine-tuning, which only adjusts model weights and can still produce plausible but incorrect names, RAG retrieves exact records at inference time, ensuring accuracy for proprietary terminology.

Exam trap

Oracle often tests the misconception that fine-tuning alone can fix factual accuracy for domain-specific entities, when in reality RAG is required to ground outputs in a verifiable external knowledge source.

How to eliminate wrong answers

Option A is wrong because switching to a smaller model reduces capacity and often increases hallucination risk due to lower parameter count and less memorization ability. Option B is wrong because prompt engineering is a fragile, surface-level fix that cannot enforce factual accuracy for specific product names; the model may still generate plausible but incorrect names. Option D is wrong because further fine-tuning with more ticket data risks overfitting and does not guarantee elimination of hallucinated product names, as the model can still invent names not present in the training distribution.

Practice this question →

61

MCQhard

A healthcare startup is building a chatbot to answer patient inquiries using a large language model (LLM) deployed on OCI Data Science AI Quick Actions. The chatbot must comply with HIPAA regulations, so all patient data must remain within the OCI tenancy and never be sent to third-party APIs. The team has fine-tuned a Llama 2 7B model on de-identified medical records using OCI Data Science notebooks. The model is deployed as a managed endpoint via AI Quick Actions. Early testing shows that the chatbot sometimes generates responses containing specific patient names or dates of birth that were present in the fine-tuning dataset. Moreover, the model occasionally hallucinates medication dosages that are not medically accurate. Which course of action should the team take to address both issues while maintaining HIPAA compliance?

A.Deploy a rule-based post-processing script that checks each response against a list of known patient names and medication dosages, and rejects any response containing them.

B.Switch to a larger model (e.g., Llama 2 70B) to improve accuracy and reduce hallucinations, and apply output filtering to remove any detected PII from responses.

C.Increase the fine-tuning dataset size with more varied de-identified records to reduce overfitting, and apply a temperature setting of 0 to make outputs deterministic.

D.Re-fine-tune the model using differential privacy to limit memorization of training data, and implement retrieval-augmented generation (RAG) with a curated medical knowledge base to ground medication-related responses.

AnswerD

Differential privacy during training reduces the risk of memorizing private data, and RAG grounds responses in a trusted knowledge base, reducing hallucinations. This combination addresses both issues effectively.

Why this answer

It addresses both memorization of PII and hallucination of medication dosages while maintaining HIPAA compliance. Differential privacy during fine-tuning limits the model's ability to memorize specific patient data, and retrieval-augmented generation (RAG) grounds responses in a curated medical knowledge base, reducing hallucinations without sending data outside the OCI tenancy.

Exam trap

Oracle often tests the misconception that simply filtering outputs or increasing model size can solve memorization and hallucination issues, when in fact only training-time techniques like differential privacy and inference-time grounding like RAG address the root causes.

How to eliminate wrong answers

Option A is wrong because a rule-based post-processing script cannot catch all variations of patient names or hallucinated dosages (e.g., misspellings, new names), and rejecting responses containing known names does not prevent the model from generating them in the first place. Option B is wrong because switching to a larger model (Llama 2 70B) does not inherently reduce memorization of training data or hallucinations; it may even increase both, and output filtering alone cannot guarantee removal of all PII without risking false positives or missing subtle leaks. Option C is wrong because increasing dataset size does not guarantee reduced overfitting or memorization, and setting temperature to 0 makes outputs deterministic but does not prevent the model from reproducing memorized PII or hallucinating dosages; it only removes randomness.

Practice this question →

62

MCQeasy

An OCI GenAI practitioner wants to deploy a model that can generate code from natural language descriptions. Which type of model is most suitable?

A.T5

B.ResNet

C.BERT

D.GPT

AnswerD

GPT (decoder-only) excels at autoregressive text generation, ideal for code generation.

Why this answer

GPT (Generative Pre-trained Transformer) is the most suitable model for code generation from natural language because it is an autoregressive language model optimized for text generation tasks. Unlike encoder-only models, GPT generates coherent, contextually relevant sequences of tokens, making it ideal for producing code based on descriptive prompts.

Exam trap

OCI GenAI exams often test the distinction between encoder-only (BERT) and decoder-only (GPT) architectures, leading candidates to mistakenly choose BERT for generation tasks because they associate it with language understanding, not realizing it cannot generate sequences.

How to eliminate wrong answers

Option A is wrong because T5 is a text-to-text transformer that, while capable of generation, is primarily designed for translation, summarization, and classification tasks, not specifically optimized for autoregressive code generation like GPT. Option B is wrong because ResNet is a convolutional neural network (CNN) architecture for image recognition and computer vision tasks, not for natural language processing or code generation. Option C is wrong because BERT is an encoder-only transformer model designed for understanding tasks (e.g., classification, question answering) and cannot generate coherent sequences of text or code due to its bidirectional, non-autoregressive architecture.

Practice this question →

63

Multi-Selecteasy

Which two factors are essential for calculating the cost of using OCI Generative AI for text generation? (Choose two.)

Select 2 answers

A.Model architecture (encoder-only vs decoder-only).

B.Number of API calls per minute.

C.Temperature setting.

D.Number of input tokens.

E.Number of output tokens.

AnswersD, E

Input tokens are a direct factor in cost calculation.

Why this answer

The cost of using OCI Generative AI for text generation is primarily determined by the number of input tokens (the prompt you send) and the number of output tokens (the generated response). OCI charges per token processed, making these two factors essential for cost calculation.

Exam trap

Oracle often tests the misconception that API call frequency or model architecture parameters directly influence cost, when in reality only token counts (input and output) are the billing units.

Practice this question →

64

Multi-Selecthard

Which THREE factors should be considered when choosing between a fine-tuning and a prompt engineering approach?

Select 3 answers

A.Latency requirements

B.Need for model personalization

C.Availability of foundation model in OCI

D.Amount of labeled data available

E.Budget for GPU compute

AnswersB, D, E

Fine-tuning is necessary for deep personalization.

Why this answer

Model personalization is a key driver for choosing fine-tuning over prompt engineering. Fine-tuning modifies the model's weights to adapt it to a specific domain or task, enabling deeper customization that prompt engineering alone cannot achieve, especially when the desired behavior requires learning new patterns or knowledge not present in the base model.

Exam trap

Oracle often tests the misconception that latency or model availability are primary differentiators, when in fact the core trade-off is between the need for deep personalization (fine-tuning) versus the ease and speed of prompt engineering, with labeled data and compute budget being practical constraints.

Practice this question →

65

MCQhard

Refer to the exhibit. A developer encounters this error. Which action should they take to resolve the issue?

A.Wait and retry after some time.

B.Change the model to cohere.command-light.

C.Increase the max-tokens value.

D.Decrease the temperature to 0.0.

AnswerA

Rate limit errors require waiting for the quota to reset, typically after a short period. Automatic retries with backoff are recommended.

Why this answer

The error indicates a rate limit or throttling issue, typically returned by the OCI Generative AI service when the API request quota is exceeded. Waiting and retrying after the cooldown period allows the rate limit to reset, which is the correct resolution for transient throttling errors.

Exam trap

Oracle often tests the misconception that model parameters (like temperature or max-tokens) can resolve API-level errors, when in fact throttling errors require waiting or implementing retry logic with backoff.

How to eliminate wrong answers

Option B is wrong because changing the model to cohere.command-light does not address rate limiting; it only changes the underlying LLM, which may have different quotas but does not resolve the current throttling error. Option C is wrong because increasing max-tokens affects the length of generated responses, not the request rate or quota limits. Option D is wrong because decreasing temperature to 0.0 controls output randomness and determinism, not API request throttling or rate limits.

Practice this question →

66

MCQeasy

A data scientist is using OCI Data Science to fine-tune a Cohere command model on domain-specific documents. They observe that the fine-tuned model generates repetitive text. What is the most likely cause?

A.The number of epochs was insufficient.

B.The training dataset lacked diversity.

C.The learning rate was too high.

D.The batch size was too small.

AnswerB

Lack of diversity in training data leads to overfitting and repetitive outputs.

Why this answer

Repetitive text in fine-tuned models is a classic symptom of overfitting to a narrow or homogeneous training dataset. When the domain-specific documents lack diversity in phrasing, topics, or contexts, the model learns to latch onto the most common patterns and repeats them, rather than generalizing. This is not a hyperparameter tuning issue but a data quality issue.

Exam trap

The trap here is that candidates often blame hyperparameters (epochs, learning rate, batch size) for overfitting symptoms, but The 1Z0-1127 exam specifically tests the understanding that data diversity is the root cause of repetitive generation in fine-tuned LLMs.

How to eliminate wrong answers

Option A is wrong because insufficient epochs typically cause underfitting, not repetitive text; the model would fail to learn patterns at all. Option C is wrong because a learning rate that is too high usually leads to training instability or divergence, not repetitive outputs. Option D is wrong because a batch size that is too small increases gradient noise and can slow convergence, but it does not directly cause repetitive text generation.

Practice this question →

67

MCQeasy

Which of the following best describes the role of attention in transformer models?

A.It assigns equal weight to all words in the input.

B.It is used only during training, not inference.

C.It allows the model to focus on relevant parts of the input sequence when generating output.

D.It replaces the need for positional encoding.

AnswerC

This is the core function of attention: it enables the model to selectively attend to important input parts.

Why this answer

The attention mechanism in transformer models dynamically computes a weighted sum of all input tokens, allowing the model to focus on the most relevant parts of the input sequence when generating each output token. This is achieved through scaled dot-product attention, which assigns higher weights to tokens that are more contextually important, enabling the model to capture long-range dependencies effectively.

Exam trap

Oracle often tests the misconception that attention is only for training or that it replaces positional encoding, so candidates must remember that attention is inherently order-agnostic and requires positional encoding to capture sequence order, and that it is used in both training and inference phases.

How to eliminate wrong answers

Option A is wrong because attention does not assign equal weight to all words; instead, it computes a distribution of weights (attention scores) that vary based on the relevance of each token to the current query, with some tokens receiving much higher weights than others. Option B is wrong because attention is used during both training and inference; during inference, the model still computes attention over the input sequence to generate each output token, though the key-value cache may be used for efficiency. Option D is wrong because attention does not replace the need for positional encoding; the self-attention operation is permutation-invariant (it treats the input as a set), so positional encodings are required to inject information about the order of tokens in the sequence.

Practice this question →

68

MCQhard

A company uses RAG (Retrieval-Augmented Generation) with OCI OpenSearch and OCI Generative AI. The system retrieves irrelevant documents. What is the first step to debug?

A.Use a different LLM

B.Increase the number of retrieved documents

C.Check the embeddings quality

D.Lower the temperature

AnswerC

Embeddings directly impact retrieval relevance; low-quality embeddings cause irrelevant results.

Why this answer

When RAG retrieves irrelevant documents, the most common root cause is poor embedding quality—vectors that fail to capture semantic similarity between the query and the documents. Checking embeddings (e.g., cosine similarity scores, dimensionality, or model used) is the logical first step before adjusting retrieval parameters or the LLM itself.

Exam trap

Oracle OCI GenAI exams often test the misconception that retrieval issues are caused by the LLM or its parameters, when in fact the root cause is almost always in the embedding or indexing pipeline.

How to eliminate wrong answers

Option A is wrong because swapping the LLM changes only the generation step, not the retrieval step; irrelevant documents will still be fed to any LLM. Option B is wrong because increasing the number of retrieved documents only adds more noise if the embeddings are poor, worsening the problem. Option D is wrong because lowering the temperature affects the randomness of the LLM's output, not the relevance of retrieved documents.

Practice this question →

69

MCQhard

An AI assistant needs to solve complex math word problems step by step. Which prompting technique is most suitable?

A.Chain-of-thought prompting with few-shot examples.

B.Zero-shot prompting with the problem only.

C.Prompting with a high temperature setting.

D.Using a model with a larger context window.

AnswerA

Correct: CoT with examples guides reasoning.

Why this answer

Chain-of-thought prompting with few-shot examples is most suitable because it guides the LLM to break down complex math word problems into intermediate reasoning steps, mimicking human problem-solving. Few-shot examples provide a template for the desired reasoning structure, which significantly improves accuracy on multi-step arithmetic tasks compared to direct answer generation.

Exam trap

Oracle often tests the misconception that simply increasing model capacity (context window) or randomness (temperature) can substitute for structured reasoning, when in fact the prompting strategy itself is the critical factor for multi-step tasks.

How to eliminate wrong answers

Option B is wrong because zero-shot prompting lacks the explicit reasoning structure needed for multi-step math problems, often leading to incorrect or incomplete answers. Option C is wrong because a high temperature setting increases randomness in token selection, which is counterproductive for deterministic math tasks requiring precise calculations. Option D is wrong because a larger context window does not inherently improve reasoning quality; it only allows more input tokens, but without structured prompting the model may still fail to perform step-by-step logic.

Practice this question →

70

MCQmedium

Refer to the exhibit. A developer runs this command and sees that the 'cohere.embed-english-v3.0' model is INACTIVE. What is the most likely cause?

A.The model is not supported in the current region.

B.The API call lacks the required OCI policy for the model.

C.The model has been deprecated and is no longer available.

D.The compartment does not have access to the model.

AnswerC

An INACTIVE state indicates the model has been deprecated or retired, making it unavailable for new inference requests.

Why this answer

The 'cohere.embed-english-v3.0' model is listed as INACTIVE because Oracle Cloud Infrastructure (OCI) has deprecated it, meaning it is no longer available for inference. When a model is deprecated, its status changes to INACTIVE, and any attempt to invoke it will fail, even if the region, policies, and compartment permissions are correctly configured.

Exam trap

Oracle often tests the distinction between model lifecycle states (INACTIVE vs. ACTIVE) and common operational errors (policy, region, compartment), leading candidates to confuse a deprecation event with a configuration or permission issue.

How to eliminate wrong answers

Option A is wrong because if the model were unsupported in the current region, the command would typically return a 'not found' or 'unsupported' error, not an INACTIVE status. Option B is wrong because a missing OCI policy would result in a 403 Forbidden or authorization error, not an INACTIVE model status. Option D is wrong because compartment access issues would produce a permissions error, not an INACTIVE status; the model's availability is independent of compartment-level access.

Practice this question →

71

Multi-Selectmedium

Which TWO of the following are benefits of using OCI Generative AI service compared to self-hosting an LLM?

Select 2 answers

A.Lower latency always

B.No data egress costs

C.Built-in content safety filters

D.Automatic scaling

E.Full control over model weights

AnswersC, D

OCI Generative AI includes safety filters.

Why this answer

OCI Generative AI service includes built-in content safety filters that automatically detect and block harmful or inappropriate content (e.g., hate speech, violence, sexual content) without requiring manual configuration. This is a key advantage over self-hosting, where you must implement and maintain your own content moderation pipeline.

Exam trap

The trap here is that candidates often assume a managed service like OCI Generative AI always provides lower latency or eliminates all data transfer costs, but in reality, self-hosting can be optimized for latency and egress costs depend on network architecture, not just the service model.

Practice this question →

72

MCQhard

A data scientist is using the OCI Generative AI SDK to create embeddings for a large corpus of legal documents. They want to perform semantic search. Which endpoint should they use?

A./v1/classify

B./v1/embed

C./v1/generate

D./v1/chat

AnswerB

The /v1/embed endpoint returns embeddings that can be stored in a vector database and used for semantic search.

Why this answer

The /v1/embed endpoint is specifically designed to generate vector embeddings from input text, which are numerical representations that capture semantic meaning. For semantic search over a large corpus of legal documents, embeddings must be created to enable similarity comparisons, making this the correct choice.

Exam trap

Oracle often tests the distinction between embedding endpoints and generation/classification endpoints, trapping candidates who confuse the purpose of semantic search (which requires embeddings) with text generation or classification tasks.

How to eliminate wrong answers

Option A is wrong because /v1/classify is used for text classification tasks (e.g., sentiment analysis or topic labeling), not for generating embeddings. Option C is wrong because /v1/generate is for text generation (e.g., completing a prompt or producing new content), not for creating vector representations. Option D is wrong because /v1/chat is designed for conversational interactions with a chat model, not for producing embeddings for semantic search.

Practice this question →

73

MCQeasy

A developer is using the OCI Generative AI API to generate text. The responses are often too short and incomplete. Which parameter adjustment is most likely to produce longer, more complete responses?

A.Decrease the max_tokens parameter.

B.Increase the max_tokens parameter.

C.Increase the top_p parameter.

D.Decrease the frequency_penalty parameter.

AnswerB

Increasing max_tokens gives the model more room to generate a complete response, directly addressing the issue of short outputs.

Why this answer

The max_tokens parameter controls the maximum number of tokens (words or subwords) the model can generate in a single response. By increasing max_tokens, the model is allowed to produce longer sequences, which directly addresses the issue of responses being too short and incomplete. In the OCI Generative AI API, this is the primary parameter for capping output length.

Exam trap

Oracle often tests the distinction between parameters that control output length (max_tokens) versus those that control output diversity or repetition (top_p, frequency_penalty), leading candidates to confuse 'more complete' with 'more creative' or 'less repetitive'.

How to eliminate wrong answers

Option A is wrong because decreasing max_tokens would further restrict the output length, making responses even shorter and more incomplete. Option C is wrong because increasing top_p adjusts nucleus sampling (the cumulative probability threshold for token selection) to control randomness and diversity, not the length of the output. Option D is wrong because decreasing frequency_penalty reduces the penalty for repeating tokens, which may increase repetition but does not directly extend the overall length or completeness of the response.

Practice this question →

74

Multi-Selecthard

Which THREE of the following are known limitations of large language models that practitioners must consider?

Select 3 answers

A.Hallucination of facts not present in the input.

B.Generation of toxic or harmful language.

C.Limited to processing only one language at a time.

D.Bias amplification from training data.

E.Inability to process inputs longer than a few hundred tokens.

AnswersA, B, D

LLMs often generate plausible but false information.

Why this answer

Large language models (LLMs) are prone to hallucination, where they generate plausible-sounding but factually incorrect information that was not present in the input. This occurs because LLMs are next-token predictors without a built-in fact-checking mechanism, and they can invent details, citations, or events to maintain coherence. Practitioners must implement retrieval-augmented generation (RAG) or external verification to mitigate this risk.

Exam trap

Oracle often tests the misconception that LLMs have a hard token limit of a few hundred tokens, but the trap is that modern models have large context windows (e.g., 128K tokens) and the real limitation is the quadratic computational cost of attention, not a strict inability to process longer inputs.

Practice this question →

75

MCQhard

An architect needs to ensure that an LLM deployed in OCI does not reveal sensitive information in its outputs. Which technique should be used?

A.Limiting max tokens

B.OCI Data Safe masking

C.Output filtering via custom inference wrapper

D.Input sanitization

AnswerC

A custom wrapper can filter outputs to remove sensitive information.

Why this answer

Output filtering via a custom inference wrapper allows the architect to inspect and sanitize the model's generated text before it reaches the user, preventing the leakage of sensitive information such as PII, credentials, or internal data. This technique operates at the application layer, intercepting the LLM's response and applying rules or regex patterns to redact or block prohibited content, which is essential for compliance and data security in production deployments.

Exam trap

Oracle often tests the distinction between input-side controls (like sanitization) and output-side controls (like filtering), and the trap here is that candidates confuse input sanitization with output filtering, assuming that cleaning the input is sufficient to prevent data leakage from the model's training or internal knowledge.

How to eliminate wrong answers

Option A is wrong because limiting max tokens only restricts the length of the output, not its content, and does nothing to prevent sensitive information from appearing within the allowed token count. Option B is wrong because OCI Data Safe masking is designed for structured databases and relational data, not for unstructured text generated by an LLM; it cannot be applied to model outputs in real-time. Option D is wrong because input sanitization focuses on cleaning user prompts before they reach the model, which is important for prompt injection prevention but does not control what the model generates in its response.

Practice this question →

Page 1 of 2 · 105 questions totalNext →

Ready to test yourself?

Try a timed practice session using only Fundamentals of Large Language Models questions.

Start 20-question session