Comprehensive and Detailed In-Depth Explanation:

A presence penalty reduces the probability of tokens that have already appeared in the output, applying the penalty each time they reoccur after their first use, to discourage repetition. This makes Option D correct. Option A (equal penalties) ignores prior appearance. Option B is the opposite—penalizing unused tokens isn’t the intent. Option C (more than twice) adds an arbitrary threshold not typically used. Presence penalty enhances output variety.

OCI 2025 Generative AI documentation likely details presence penalty under generation control parameters.

Comprehensive and Detailed In-Depth Explanation:

Fine-tuning is suitable when an LLM underperforms on a specific task and prompt engineering alone isn’t feasible due to large, task-specific data that can’t be efficiently included in prompts. This adjusts the model’s weights, making Option B correct. Option A suggests no customization is needed. Option C favors RAG for latest data, not fine-tuning. Option D is vague—fine-tuning requires data and goals, not just optimization without direction. Fine-tuning excels with substantial task-specific data.

OCI 2025 Generative AI documentation likely outlines fine-tuning use cases under customization strategies.

Comprehensive and Detailed In-Depth Explanation:

In systems like RAG, the Ranker evaluates and sorts the information retrieved by the Retriever (e.g., documents or snippets) based on relevance to the query, ensuring the most pertinent data is passed to the Generator. This makes Option C correct. Option A is the Generator’s role. Option B describes the Retriever. Option D is unrelated, as the Ranker doesn’t interact with users but processes retrieved data. The Ranker enhances output quality by prioritizing relevant content.

OCI 2025 Generative AI documentation likely details the Ranker under RAG pipeline components.

Comprehensive and Detailed In-Depth Explanation:

Loss measures the discrepancy between a model’s predictions and true values, with lower values indicating better fit—Option D is correct. Option A (accuracy difference) isn’t loss—it’s a derived metric. Option B (error percentage) is closer to error rate, not loss. Option C (accuracy improvement) is a training outcome, not loss’s definition. Loss is a fundamental training signal.

OCI 2025 Generative AI documentation likely defines loss under fine-tuning metrics.

Comprehensive and Detailed In-Depth Explanation:

In RAG, the Ranker evaluates and prioritizes retrieved information (e.g., documents) based on relevance to the query, refining what the Retriever fetches—Option D is correct. The Retriever (A) fetches data, not ranks it. Encoder-Decoder (B) isn’t a distinct RAG component—it’s part of the LLM. The Generator (C) produces text, not prioritizes. Ranking ensures high-quality inputs for generation.

OCI 2025 Generative AI documentation likely details the Ranker under RAG pipeline components.

Comprehensive and Detailed In-Depth Explanation:

Chain-of-Thought (CoT) prompting encourages an LLM to emit intermediate reasoning steps before providing a final answer, improving performance on complex tasks by mimicking human reasoning. This matches the scenario, making Option D correct. Option A (In-context Learning) involves learning from examples in the prompt, not necessarily reasoning steps. Option B (Step-Back Prompting) involves reframing the problem, not emitting steps. Option C (Least-to-Most Prompting) breaks tasks into subtasks but doesn’t focus on intermediate reasoning explicitly. CoT is widely recognized for reasoning tasks.

OCI 2025 Generative AI documentation likely covers Chain-of-Thought under advanced prompting techniques.