Generative AI Leader Practice Test 28 — 15 Questions

Question 1

A developer receives the above JSON response from a Vertex AI language model. The output content is correct, but the developer expected the model to not answer geography questions. What should the developer do to prevent the model from responding to geography queries?

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

Accepted Answer

Add a system instruction to not answer geography questions. Vertex AI does not have a predefined safety filter category for 'Geography'. To prevent the model from answering geography questions, the developer should use a system instruction that explicitly tells the model not to respond to geography queries. System instructions are the appropriate mechanism for custom content restrictions, while safety filters are limited to predefined categories such as toxic, harassment, etc.

Answer

Adjust the safety filter thresholds for the 'Toxic' category

Answer

Enable Vertex AI Grounding with a geography knowledge base

Answer

Configure a safety filter for the 'Geography' category

Question 2

After fine-tuning a foundation model on company emails, the model outputs confidential information. What is the most likely cause?

Accepted Answer

The fine-tuning dataset was not anonymized. The most likely cause of a fine-tuned model outputting confidential information is that the fine-tuning dataset contained sensitive data that was not anonymized. During fine-tuning, the model learns patterns and can memorize specific sequences, including confidential details like names, addresses, or proprietary information, which it then reproduces in responses. This is a well-known data leakage risk in fine-tuning workflows.

Answer

The prompt is too vague

Answer

The model is too large

Answer

Overfitting to the training data leading to memorization

Question 3

A data scientist is using Vertex AI Model-as-a-Service (MaaS) to deploy a fine-tuned open-source model. They notice high latency during inference. What is the most likely cause?

Accepted Answer

The endpoint is set to autoscaling with a low minimum node count. High latency during inference on Vertex AI MaaS is most often caused by the endpoint scaling configuration. When autoscaling is enabled with a low minimum node count, the system may need to provision additional nodes to handle the request load, which introduces cold-start latency. This is especially pronounced for fine-tuned open-source models, which can be large and take time to load onto new nodes.

Answer

The model is too large for the hardware

Answer

The model is not quantized

Answer

The region is incorrect

Question 4

A media company is using Vertex AI Imagen to generate marketing images. The output frequently contains unrealistic artifacts, especially in human faces. The team has fine-tuned the model using their brand assets. What is the most likely cause and recommended fix?

Accepted Answer

The fine-tuning dataset is too small or too homogeneous; augment and diversify the training data.. Unrealistic artifacts in fine-tuned generative models, especially in human faces, typically stem from a training dataset that is too small or lacks diversity. When the dataset is homogeneous, the model overfits to limited patterns and fails to generalize, leading to distorted outputs. Augmenting and diversifying the training data with varied poses, lighting, and ethnicities helps the model learn robust facial features.

Answer

Safety filters are too aggressive; reduce them.

Answer

Negative prompts are missing; always include 'unrealistic'.

Answer

Inference steps are too low; increase to 100.

Question 5

A developer wants to use Gemini 1.5 Pro to analyze hour-long video content and generate a summary. Which feature of Gemini 1.5 Pro is most suitable for this task?

Accepted Answer

Long context window (up to 1 million tokens). Gemini 1.5 Pro's long context window of up to 1 million tokens allows it to process and analyze hour-long video content in a single pass, including both audio and visual frames. This capability is essential for generating a coherent summary of long-form video, as it can retain and reason over the entire video's context without needing to chunk or downsample the content.

Answer

Multimodal generation from text and images

Answer

Code generation and debugging

Answer

Function calling to retrieve external data

Question 6

A team uses a generative model to summarize lengthy legal documents. The summaries are accurate but often exceed the target length of 200 words, varying widely. Which simple adjustment should be applied to ensure consistent output length?

Accepted Answer

Set the max output tokens parameter to 200.. Setting the max output tokens parameter to 200 directly caps the number of tokens the model can generate, ensuring the summary cannot exceed the target length. This is a deterministic, model-level constraint that works regardless of prompt phrasing or training data, making it the most reliable adjustment for consistent output length.

Answer

Fine-tune the model on summaries that are exactly 200 words.

Answer

Add a system prompt that says 'Summarize in exactly 200 words.'

Answer

Lower the temperature to reduce variability in word choices.

Question 7

Which THREE factors should be considered when selecting a foundation model for a generative AI application in a regulated industry?

Accepted Answer

Transparency of the model's training data and sources. In regulated industries (e.g., healthcare, finance), transparency of training data and sources is critical for compliance with regulations like GDPR or HIPAA. Without knowing the provenance and composition of the training data, an organization cannot audit for prohibited content, verify consent, or ensure the model does not inadvertently expose sensitive information. This transparency directly impacts the ability to perform due diligence and meet legal obligations for data usage.

Answer

Latency and throughput requirements

Answer

Size of the model in terms of parameters

Question 8

Which THREE approaches are effective for reducing bias in generative model outputs? (Choose three.)

Accepted Answer

Use a balanced training dataset.. A balanced training dataset reduces the risk of the model learning spurious correlations or skewed distributions that lead to biased outputs. By ensuring that all demographic groups, topics, or perspectives are represented proportionally, the model's learned probability distribution is less likely to favor one group over another, directly mitigating representation bias at the data level.

Answer

Set temperature to a very high value.

Answer

Use adversarial training.

Question 9

A company is developing a generative AI application that must comply with GDPR. Which TWO actions are required? (Select two.)

Accepted Answer

Provide a mechanism for users to request deletion of their personal data. GDPR requires a lawful basis for processing personal data, such as user consent, and the right to erasure (right to be forgotten). Data subjects must be able to request deletion of their data.

Answer

Anonymize all training data to prevent any re-identification

Answer

Use a model that has been pre-trained only on non-EU data

Answer

Store all user data in a centralized location for easier management

Question 10

A team is building a generative AI model for customer support. They notice the model often produces overly polite but unhelpful responses. Which technique would best improve response quality without sacrificing helpfulness?

Accepted Answer

Apply reinforcement learning from human feedback (RLHF). RLHF directly addresses the misalignment between the model's training objective (e.g., predicting the next token) and the desired outcome (helpful, not just polite). By using human feedback to train a reward model, the system learns to optimize for response quality and helpfulness, reducing sycophantic or overly polite but uninformative outputs.

Answer

Increase the amount of training data

Answer

Lower the top_k sampling value

Answer

Increase the temperature parameter

Question 11

A financial institution wants to use Gemini to analyze customer support transcripts and generate summaries. They need to ensure that personally identifiable information (PII) is not included in the summaries. Which approach should they take?

Accepted Answer

Preprocess the transcripts with Cloud DLP API to redact PII before sending to Gemini. The Cloud Data Loss Prevention (DLP) API provides a purpose-built, scalable service to detect and redact PII from text before it reaches the Gemini model. This ensures that sensitive data is removed at the source, preventing any possibility of leakage in the generated summary, regardless of model behavior or prompt engineering.

Answer

Use a carefully engineered prompt instructing Gemini not to include PII

Answer

Post‑process the generated summaries with a regex filter to remove PII

Answer

Fine‑tune Gemini to avoid generating PII

Question 12

Refer to the exhibit. An administrator creates this IAM policy for a Vertex AI project. What is the effect of this policy?

Accepted Answer

Alice can deploy pre-trained models; Bob can create and manage custom model code. The IAM policy grants Alice the `aiplatform.models.get` permission (allowing her to view and deploy pre-trained models) and grants Bob the `aiplatform.models.create` and `aiplatform.models.update` permissions (allowing him to create and manage custom model code). The policy uses separate bindings for each user, with specific roles that align with these actions.

Answer

Alice can view models; Bob can delete models

Answer

Both have full access to all Vertex AI resources

Answer

Alice can train models; Bob can deploy models

Question 13

A retail company is building a product description generator using a large language model on Vertex AI. They need to ensure the generated descriptions do not contain offensive language. Which strategy should they implement?

Accepted Answer

Implement a content moderation filter (e.g., Perspective API) as a post-processing step. Content moderation filters like Perspective API act as a post-processing safeguard that can catch offensive language the model might generate despite prompt engineering or fine-tuning. This approach provides a deterministic, rule-based or ML-based check that is independent of the model's training, ensuring compliance with content policies in production. It is a standard practice for deploying LLMs in customer-facing applications where safety is critical.

Answer

Fine-tune the model on a dataset of clean product descriptions

Answer

Use Vertex AI Model Monitoring to detect anomalies in model predictions

Answer

Include explicit instructions in the prompt to avoid offensive language

Question 14

Which of the following is a best practice when using Vertex AI for prompt engineering?

Accepted Answer

Use consistent formatting and delimiters. Consistent formatting and delimiters (e.g., using triple backticks, XML tags, or clear section headers) help the model parse instructions and context reliably, reducing ambiguity and improving output quality. This is a core best practice in prompt engineering on Vertex AI because it leverages the model's attention mechanisms to focus on distinct prompt segments, leading to more predictable and accurate responses.

Answer

Always set temperature to 0

Answer

Avoid using examples in the prompt

Answer

Use very long prompts to include all possible instructions

Question 15

A hospital deploys a generative AI model to assist in diagnosing rare diseases. The model sometimes suggests treatments that are not widely accepted. To align with Google's AI Principle of being accountable to people, what should the hospital implement?

Accepted Answer

Implement a human override mechanism that allows clinicians to reject AI suggestions. Google's AI Principle of being accountable to people requires that AI systems remain under human control, especially in high-stakes domains like healthcare. A human override mechanism ensures that clinicians can reject AI suggestions that are not widely accepted, directly addressing the principle's emphasis on human oversight and responsibility for outcomes.

Answer

Provide confidence scores for each suggestion to help clinicians assess reliability

Answer

Limit the model to only suggest treatments that appear in a predefined list

Answer

Conduct a retrospective audit of all AI suggestions monthly