An organization is deploying a summarization model on Vertex AI and needs to ensure that the model's responses are consistent and avoid hallucinations. They have a labeled dataset of source documents and human-written summaries. Which approach would best align the model with their quality requirements?
Supervised fine-tuning adjusts the model's weights using paired source documents and human-written summaries, directly teaching the desired output distribution. This satisfies the consistency and hallucination-reduction constraint, unlike prompt engineering or retrieval, which cannot reliably shape response style.
Why this answer
Supervised fine-tuning (D) directly optimizes the model's weights using the labeled dataset of source documents and human-written summaries, which teaches the model to produce consistent, factual outputs and reduces hallucinations. This approach aligns the model's behavior with the specific quality requirements by learning from ground-truth examples, unlike prompt engineering or parameter adjustments that do not modify the underlying model.
Exam trap
The trap here is that candidates often overestimate the power of prompt engineering or parameter tweaks, mistakenly believing they can achieve the same level of reliability as fine-tuning, when in fact only supervised fine-tuning directly modifies the model to internalize the labeled data's quality standards.
How to eliminate wrong answers
Option A is wrong because increasing max_output_tokens only extends the length limit of the response, which does not improve consistency or reduce hallucinations—it may actually allow the model to generate more unverified content. Option B is wrong because prompt engineering with few-shot examples provides in-context guidance but does not permanently adjust the model's weights; it is less reliable for ensuring consistent, hallucination-free outputs across diverse inputs compared to fine-tuning. Option C is wrong because increasing temperature to 0.9 increases randomness in token selection, which amplifies variability and the risk of hallucinations, directly opposing the goal of consistency.