1Z0-1127-25 Fundamentals of Large Language Models Practice Test 9 — 15 Questions

Question 1

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?

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

Accepted Answer

Insufficient training data quality or quantity. 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.

Answer

Incorrect tokenizer configuration

Answer

Too many epochs causing overfitting

Answer

Model architecture mismatch between fine-tuned and base model

Question 2

An enterprise wants to deploy a large language model for processing sensitive internal documents. They must ensure that data does not leave their OCI tenancy. Which OCI GenAI deployment option meets this requirement?

Accepted Answer

Dedicated AI Cluster with on-demand model hosting. A Dedicated AI Cluster with on-demand model hosting provides a fully isolated, single-tenant infrastructure within your OCI tenancy. This ensures that all data processing and model inference occur on dedicated hardware that never leaves your tenancy, meeting strict data residency and security requirements for sensitive internal documents.

Answer

Using a third-party model via OCI Marketplace

Answer

Accessing models through OCI Console only

Answer

Using the OCI GenAI API with the default endpoint

Question 3

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

Accepted Answer

Use semantic search (embeddings) for document retrieval. Option C is correct because semantic search using embeddings is the core retrieval mechanism in a RAG pipeline: documents and the user query are converted into vector representations, and similarity search (e.g., cosine similarity or dot product over a vector index) retrieves the most contextually relevant chunks to feed the LLM, which dramatically improves answer grounding compared to keyword-only matching. Option D is correct because caching responses for frequently asked questions reduces redundant embedding lookups, vector searches, and LLM inference calls, lowering latency and cost while maintaining consistent answers for repeated queries. Option A is incorrect because the embedding model is used only for retrieval and indexing, while generation is performed by a separate large language model; they are distinct model types and need not match. Option B is incorrect because dumping all documents into a single large index degrades retrieval precision and scalability; best practice is chunking documents and using partitioned or filtered indexes. Option E is incorrect because summarization (or context compression) is a useful technique to fit relevant content within the model's context window and improve generation quality, so disabling it is not a best practice.

Answer

Use the same embedding model for both retrieval and generation

Answer

Store all documents in a single large index

Answer

Disable summarization to save inference costs

Question 4

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

Accepted Answer

Llama models are open-source and available for fine-tuning. Option A is correct because Meta's Llama family in OCI Generative AI is offered as open-source models that can be fine-tuned on custom datasets via dedicated AI clusters. Option C is correct because Cohere embedding models (e.g., embed-english-v3.0, embed-multilingual-v3.0) generate dense vector representations used for semantic search and RAG. Option D is correct because OCI Generative AI offers on-demand hosted inference endpoints as well as dedicated AI cluster deployment for fine-tuning and hosting custom models. Option B is not marked correct because not every model in the catalog supports token streaming (availability varies by model and endpoint type). Option E is not marked correct because the multilingual optimization claim is not a defining characteristic of Cohere Command models in this context; Command models are primarily optimized for instruction-following and RAG-style tasks.

Answer

All models support real-time streaming of tokens

Answer

Cohere Command models are optimized for multilingual tasks

Question 5

A healthcare company is using OCI GenAI to generate patient summaries from clinical notes. The model output sometimes includes hallucinated medical facts, such as incorrect dosages or diagnoses, which could be dangerous. The team needs to improve factual accuracy while maintaining data privacy. They have a large collection of internal medical knowledge bases (clinical guidelines, drug databases) that are stored in OCI Object Storage. The current implementation uses a zero-shot prompt with the base Cohere Command model. The data science team has limited GPU resources and wants to avoid building a complex pipeline. Which course of action best addresses the hallucination problem?

Accepted Answer

Implement a RAG pipeline that retrieves relevant documents from the internal knowledge bases and includes them in the prompt.. A Retrieval-Augmented Generation (RAG) pipeline directly addresses hallucination by grounding the model's output in verified, internal medical knowledge bases stored in OCI Object Storage. This approach retrieves relevant clinical guidelines or drug database entries and includes them in the prompt, providing factual context without requiring fine-tuning or complex GPU-intensive pipelines. It also preserves data privacy by keeping sensitive medical data within OCI and avoids exposing it to external model training.

Answer

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

Answer

Use prompt engineering to add 'Only provide facts that are absolutely certain.'

Answer

Fine-tune the Cohere model on a publicly available medical dataset like PubMed.

Question 6

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?

Accepted Answer

Fine-tuning the model on the curated dataset of question-answer pairs.. Fine-tuning the model on the curated dataset of question-answer pairs is the most effective approach to enforce strict consistency with approved responses. Fine-tuning adjusts the model's weights to learn the specific patterns and content of the curated answers, making it highly likely to reproduce them. This is more reliable than prompt-based methods for ensuring adherence to a fixed set of responses.

Answer

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

Answer

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

Answer

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

Question 7

A research team is using OCI Data Science and OCI GenAI to build a multilingual chatbot for customer service. They have training data in English, Spanish, and French. The model currently struggles with code-switching—users often mix languages in a single query (e.g., 'Quiero cancel my order'), and the model responds inconsistently, sometimes in English, sometimes mixing incorrectly. The team wants to improve performance on code-switching while maintaining fluency in each language. They have limited compute resources and cannot deploy separate models per language. Which approach should they take?

Accepted Answer

Fine-tune a multilingual model on a combined dataset that includes code-switching examples.. The correct option is B: fine-tuning a multilingual model on a combined dataset that includes code-switching examples. This directly addresses the scenario because the model must handle intra-sentence language mixing, and exposing it to code-switched training data teaches it to map mixed-language inputs to consistent, fluent responses without needing separate per-language models. It also fits the limited-compute constraint since one multilingual model is adapted rather than deploying multiple models. Option A is wrong because separate per-language models and routing cannot handle code-switching within a single query and would require more deployment resources. Option C is wrong because translating and routing loses code-switch context and can produce inconsistent responses. Option D is wrong because retrieval embeddings improve context lookup, not the model's generation behavior on mixed-language queries.

Answer

Train separate fine-tuned models for each language and route queries based on detected language.

Answer

Use language detection to route the query to a specific language model, then translate the response.

Answer

Use a multilingual embedding model for retrieval to improve context understanding.

Question 8

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?

Accepted Answer

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

Answer

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

Answer

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

Answer

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

Question 9

Which TWO statements about large language model (LLM) capabilities are correct?

Accepted Answer

LLMs can perform zero-shot learning without any task-specific training.. Option B is correct because LLMs are pretrained on broad corpora and can generalize to new tasks from a prompt alone, which is the definition of zero-shot learning without task-specific training. Option C is correct because LLMs are trained on large amounts of source code alongside natural language, enabling them to parse, generate, and reason about code constructs such as functions, syntax, and APIs. Option A is wrong because context windows are not fixed permanently; they can be extended through techniques like positional interpolation, longer-context training, or retrieval augmentation. Option D is wrong because LLMs can hallucinate and produce plausible but factually incorrect outputs. Option E is wrong because fine-tuning is optional, not required for every new task, since prompting and in-context learning often suffice.

Answer

LLMs have a fixed context window that cannot be extended.

Answer

LLMs always produce factually accurate outputs.

Answer

LLMs require fine-tuning for every new task.

Question 10

Refer to the exhibit. A data scientist runs this inference request and receives a response that is incomplete and seems to stop mid-sentence. Which parameter should be adjusted to allow the model to generate longer outputs?

Accepted Answer

maxTokens. The `maxTokens` parameter controls the maximum number of tokens (words or subwords) the model can generate in a single response. When a response stops mid-sentence, it indicates the token limit was reached before the model could complete its output. Increasing `maxTokens` allows the model to generate longer, complete responses.

Answer

temperature

Answer

topP

Answer

frequencyPenalty

Answer

presencePenalty

Question 11

A multinational corporation is deploying a generative AI chatbot for customer support using Oracle Cloud Infrastructure's Generative AI service. The chatbot is powered by a large language model (LLM) accessed via the on-demand serving mode. During initial testing, the chatbot provides accurate answers for well-known products but frequently hallucinates or gives incorrect specifications for niche products. The company maintains a comprehensive internal database of product specifications, updated daily. The support team prefers not to fine-tune the LLM due to cost and maintenance overhead. Additionally, the chatbot must respond within 2 seconds to maintain a good customer experience. The team considers several approaches: A. Increasing the 'temperature' parameter to make the model more creative, hoping it will generate more accurate responses when unsure. B. Using few-shot prompting with three manually curated examples of correct product specifications included in every prompt. C. Implementing a Retrieval Augmented Generation (RAG) pipeline that retrieves relevant product documents from the internal database and prepends them to the prompt before inference. D. Reducing the 'topP' parameter to 0.1 to force the model to sample only from the highest probability tokens, thereby reducing randomness. Which approach best meets the requirements of improving factual accuracy while maintaining low latency?

Accepted Answer

Implement a Retrieval Augmented Generation (RAG) pipeline that retrieves relevant product documents from the internal database and prepends them to the prompt before inference.. Retrieval Augmented Generation (RAG) directly addresses the hallucination problem by providing the LLM with up-to-date, factual product specifications from the internal database as context in the prompt. This approach improves factual accuracy without fine-tuning, and because retrieval can be optimized (e.g., using vector search with approximate nearest neighbor algorithms), it can meet the 2-second latency requirement. RAG leverages the LLM's existing knowledge while grounding responses in authoritative external data.

Answer

Reduce the 'topP' parameter to 0.1 to force the model to sample only from the highest probability tokens, thereby reducing randomness.

Answer

Use few-shot prompting with three manually curated examples of correct product specifications included in every prompt.

Answer

Increase the 'temperature' parameter to make the model more creative, hoping it will generate more accurate responses when unsure.

Question 12

An ML engineer is preparing a fine-tuning dataset for an OCI Generative AI custom model that must answer questions about a proprietary 40-page employee handbook. The engineer has 300 question-answer pairs but no long-form documents in the training file. The evaluation shows the model answers simple factual questions correctly but fails on questions requiring synthesis across multiple sections. Which dataset design change best addresses this limitation?

Accepted Answer

Add training examples where each answer requires combining information from two or more distinct handbook sections, with the source sections cited.. The evaluation shows the fine-tuned model can retrieve isolated facts but cannot integrate evidence across sections, which points to a training-signal structure problem rather than a volume problem. Adding multi-section synthesis examples, especially with citations, teaches the model the reasoning pattern the evaluation is testing. Duplicating data, isolating answers to single chunks, or swapping the corpus does not create that multi-hop signal. Dataset design, not dataset size, is the lever that closes this gap.

Answer

Replace the handbook with a public HR policy corpus so the model has more diverse training material.

Answer

Duplicate the 300 question-answer pairs three times to increase the effective dataset size to 900 examples.

Answer

Convert the handbook into chunked passages and pair each chunk with a question whose answer appears only in that chunk.

Question 13

A machine learning engineer is configuring inference for a text generation model on OCI Generative AI. They need the model to consistently select the most probable next token at each step to produce deterministic outputs for a testing pipeline. Which decoding strategy should they use?

Accepted Answer

Greedy decoding. Greedy decoding always picks the highest-probability token, yielding deterministic outputs. For a testing pipeline that requires reproducibility, this is the correct choice. Sampling methods like top-k and top-p introduce randomness, and beam search, while more deterministic than sampling, can still vary due to tie-breaking and is not guaranteed to be identical across runs.

Answer

Top-k sampling with k=50

Answer

Beam search with beam width 5

Answer

Top-p sampling with p=0.9

Question 14

A data scientist is fine-tuning a base model on OCI Generative AI using a dataset of customer support conversations. The dataset contains many examples where the assistant's response is a single word like 'Yes' or 'No'. After fine-tuning, the model frequently produces overly terse answers even for complex queries. Which data characteristic is the most likely cause?

Accepted Answer

The dataset lacks sufficient examples of longer, more detailed responses, causing the model to learn a bias toward brevity.. Fine-tuning a model on a dataset where most assistant responses are very short teaches the model to mimic that brevity. The model adjusts its weights to minimize loss on the training data, so it learns that short answers are appropriate. To correct this, the dataset should include a balanced mix of short and detailed responses, or the training data should be filtered to remove trivial examples.

Answer

The learning rate used during fine-tuning was too high, causing the model to forget how to generate long sequences.

Answer

The base model's tokenizer is incompatible with the dataset's vocabulary, causing truncation of longer responses.

Answer

The training dataset was not shuffled before fine-tuning, so the model only saw short examples in the final epochs.

Question 15

A solutions architect is designing an OCI Generative AI application that must respond to user prompts within 1 second. The architect is evaluating the Llama 3.1 405B model but is concerned about inference latency. Which factor most directly affects the model's inference speed during generation?

Accepted Answer

The number of parameters in the model. Inference latency is primarily determined by the model's parameter count, as each generated token requires a forward pass through all parameters. Larger models like 405B are slower than smaller ones. While other factors like hardware and batching matter, the parameter count is the fundamental driver of per-token computation time.

Answer

The tokenizer's vocabulary size

Answer

The size of the training dataset used to pretrain the model

Answer

The temperature setting used during inference