Generative AI Leader Practice Question: A data science team wants to build a custom model…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Fine-tune a foundation model (e.g., PaLM 2) using Vertex AI Model Garden

  Why this is correct

  Fine-tuning adapts a foundation model's weights to the brand's tone using the 5,000 labelled examples, achieving higher fidelity than prompting alone at far lower cost than training from scratch. Vertex AI Model Garden provides the managed pipeline for this.

• ✗

  Use a pre-built API with prompt engineering and few-shot examples

  Why it's wrong here

  Few-shot prompting cannot alter a pre-built API's underlying weights, so brand-specific phrasing and tone remain constrained by the base model rather than learned from the 5,000 examples. It suits rapid prototyping where general capability suffices; fine-tuning is required when consistent adherence to bespoke guidelines matters.

• ✗

  Use Vertex AI Agent Builder with a custom prompt

  Why it's wrong here

  Agent Builder with a custom prompt orchestrates and grounds existing foundation models; it does not fine-tune weights on the 5,000 brand examples, so guideline adherence stays prompt-dependent. It is tempting because Agent Builder is correct when the task is retrieval, tool orchestration or conversational grounding rather than style adaptation.

• ✗

  Train a model from scratch using TensorFlow on Vertex AI

  Why it's wrong here

  Training from scratch requires far more than 5,000 examples and massive compute, producing a weaker model at prohibitive cost. It is tempting because from-scratch training is correct when no foundation model exists for the domain, such as novel molecular or proprietary signal data.

Go deeper

Related to this question