A machine learning engineer is curating examples for either fine-tuning or few-shot instruction for a RAG-based summarization application. This system is designed to take 3-5 segments of text and produce a brief, coherent summary that addresses a user's inquiry by integrating relevant details. From the potential prompt-response samples below, which one is most suitable for training this specific model capability?
Answer & explanation
Correct answer: E
WHY E: A training or few-shot example should mirror what the model sees at inference: a user question, several retrieved segments, and an instruction to stay within them, answered by one concise paragraph that combines the segments. This pair teaches exactly that context-grounded synthesis.
WHY NOT A: A closed-book factual pair trains the model to answer from its own parametric knowledge, which is the opposite of summarizing from supplied text segments. It contains no context and no synthesis.
WHY NOT B: Keyword input with a label output is a classification-style pair. It does not teach fluent paragraph generation, grounding in context, or combining several segments.
WHY NOT C: Translation is a sentence-to-sentence rewrite task, not multi-segment synthesis. Including it dilutes the task focus and can confuse the model about the expected output format.
WHY NOT D: Long essay pairs teach verbose, multi-paragraph output with no supplied context, which pulls the model away from the brief, grounded summary the application needs.