Five Star AWS AI Practitioner Book ★ ★ ★ ★ ★

Within AWS generative AI services such as Amazon Bedrock, how should tokens be understood when a model processes text during training or inference?

• ❏ A. The pre-trained parameter values of a foundation model that can later be fine-tuned

• ❏ B. The smallest textual units a model reads and writes, such as words or subword pieces

• ❏ C. The dense vector representations that capture word or concept meaning

• ❏ D. Amazon Comprehend

NorthRiver Finance, a regional credit union, is building an AI-powered portfolio advisor. At times the model suggests aggressive actions that violate internal compliance policies. The team wants to constrain the model so its outputs remain within policy-approved guidance. Which prompt-engineering approach will help enforce these limits?

• ❏ A. Use zero-shot prompting to elicit more direct answers

• ❏ B. Increase the model’s input and output token limits

• ❏ C. Define explicit safety constraints and guardrails within the prompt

• ❏ D. Amazon GuardDuty

A fashion retailer uses an image diffusion model in Amazon Bedrock to create 6K product ads and social media visuals. Which considerations will most improve the image quality and brand consistency of the outputs? (Choose 2)

• ❏ A. Using a carefully labeled, high-quality training and reference image set

• ❏ B. Tracking inference latency and throughput

• ❏ C. Optimizing prompt token usage to reduce size

• ❏ D. Fine-tuning the diffusion model on brand-specific examples

• ❏ E. Expanding the model’s context window length

A regional insurer, Cedar Ridge Insurance Group, plans a 45-day pilot of Amazon Q Business to auto-summarize reports and provide cross-team insights for claims and underwriting. Because the company processes confidential policyholder data, the security office needs clarity on which administrative guardrails and response-source controls exist in Amazon Q Business to meet compliance needs. What should the teams consider about Amazon Q Business admin controls and guardrails? (Choose 2)

• ❏ A. Amazon Q Business can be configured to answer using only the model’s built-in knowledge

• ❏ B. Configure Amazon Q Business to use enterprise data only or combine enterprise data with model knowledge

• ❏ C. End users can never upload files in chat to generate answers from those files

• ❏ D. AWS WAF

• ❏ E. Amazon Q Business guardrails provide topic-level rules that define how the app responds when a blocked topic is mentioned

A regional logistics provider, Polaris Freight, is creating an AI roadmap to automate back-office tasks and enhance analytics. Executives need a clear view of how Artificial Intelligence (AI), Machine Learning (ML), Deep Learning (DL), and Generative AI (GenAI) relate so they can align budgets and teams. Which ordering correctly shows the hierarchy from the broadest discipline to the most specialized capability?

• ❏ A. Machine Learning > Deep Learning > Artificial Intelligence > Generative AI

• ❏ B. Generative AI > Deep Learning > Machine Learning > Artificial Intelligence

• ❏ C. Artificial Intelligence > Machine Learning > Deep Learning > Generative AI models

• ❏ D. Artificial Intelligence > Generative AI > Machine Learning > Deep Learning

A regional procurement agency has added a large language model to turn long vendor contracts, often exceeding 90 pages, into standardized compliance briefs. The goal is to reduce manual effort and improve consistency, but the legal review team is concerned the model could favor certain clauses or phrasing and subtly bias approval decisions. They want a low-maintenance method to assess the model for fairness and representational balance that still provides useful, repeatable insights. Which approach should they use to evaluate potential bias with minimal administrative effort?

• ❏ A. Launch a limited pilot and gather structured bias feedback through targeted surveys

• ❏ B. Amazon Bedrock model evaluation with pre-built bias and fairness prompt datasets

• ❏ C. Continuously fine-tune the model using recent responses from a diverse user group

• ❏ D. Amazon SageMaker Clarify

A corporate training provider, NovaPath Learning, plans to use foundation models to create individualized study guides and automatically draft lesson materials. The curriculum team wants a clear understanding of what these models can do so they can assess fit for their courses. Which statement accurately describes Foundation Models in generative AI?

• ❏ A. Foundation models cannot personalize outputs based on learner interactions

• ❏ B. Foundation models are pre-trained on large, diverse datasets and can be fine-tuned or guided with prompts to handle many downstream tasks

• ❏ C. Each foundation model is built for a single narrow use and cannot be adapted to other applications

• ❏ D. On Amazon Bedrock, foundation models must be retrained from scratch for every subject domain

An e-commerce marketplace called NovaGoods builds a demand forecasting model to anticipate product purchases. It reports 99% accuracy on its training data, but when evaluated on live customer orders from the next 45 days it performs poorly. What is the most likely cause of this behavior?

• ❏ A. The training dataset was missing labels

• ❏ B. Amazon Forecast

• ❏ C. The model has overfit to the training data

• ❏ D. The model is underfitting the training data

A regional insurance carrier, Northwind Mutual, is experiencing rapid growth in the volume of scanned policies, addendums, and claim files and wants to speed up review by automatically extracting key clauses, effective and renewal dates, and named entities across roughly 80,000 pages each month while maintaining high accuracy. Which options would best enable an automated solution to meet this need? (Choose 3)

• ❏ A. Amazon Polly

• ❏ B. Amazon Textract

• ❏ C. Generative AI summarization chatbot

• ❏ D. Amazon Personalize

• ❏ E. Convolutional Neural Network (CNN)

• ❏ F. Amazon Comprehend

A streaming media startup, NovaStream, is building machine learning models to study viewer engagement and improve content recommendations. Over the last 18 months, the team has ingested structured records from relational tables and unstructured assets such as captions, thumbnails, and audio stored in Amazon S3. To choose the right feature engineering steps and algorithms, how should the team distinguish between structured and unstructured data?

• ❏ A. Structured data is typically freeform text with no specific organization, while unstructured data is arranged in a fixed tabular layout

• ❏ B. Structured data must reside in Amazon RDS, and unstructured data must be stored only in Amazon S3 and cannot be queried

• ❏ C. Structured data conforms to a defined schema, often as rows and columns that are easy to query and aggregate, while unstructured data lacks a fixed model and includes items like text, images, audio, and video

• ❏ D. Structured data is only used to train models, whereas unstructured data is kept solely for archival purposes

An engineering group at BrightPixel Labs wants to try a foundation model and expose it through a private endpoint inside the team’s Amazon VPC with minimal setup in about 45 minutes. Which AWS service or feature should they use to rapidly deploy and start consuming the model from within their VPC?

• ❏ A. Amazon SageMaker endpoints

• ❏ B. Amazon Personalize

• ❏ C. Amazon SageMaker JumpStart

• ❏ D. PartyRock, an Amazon Bedrock Playground

An architecture studio plans to compare several foundation models in Amazon Bedrock to generate high-resolution marketing visuals during the next 90 days. Which evaluation criteria should they emphasize to select the most suitable model? (Choose 3)

• ❏ A. BLEU score

• ❏ B. Amazon Rekognition

• ❏ C. Model architecture and capabilities

• ❏ D. Price per image or token usage

• ❏ E. Inference latency and output quality metrics

Blue Finch Animation, a streaming content studio, uses a generative AI model to draft character bios and dialogue. After reviewing 30 recent scenes, editors notice recurring gender stereotypes in the outputs. What is the most effective first step the team should take to reduce this bias in the model’s responses?

• ❏ A. Increase the model’s temperature setting

• ❏ B. Curate a more representative training dataset

• ❏ C. Conduct subgroup bias analysis

• ❏ D. Fine-tuning the model

NovaStream Media plans a 9-month pilot to add generative AI features to two of its applications using Amazon Bedrock. Usage could vary widely from week to week, and the team wants to avoid pre-purchasing capacity or making any long-term commitments while they experiment. Which pricing model should they select to keep costs flexible and pay only when they use the service?

• ❏ A. Provisioned throughput

• ❏ B. EC2 Spot Instances

• ❏ C. On-demand pricing

• ❏ D. EC2 Reserved Instances

An online retail startup called Northwind Insights uses Amazon Bedrock to create tailored product summaries and suggestions. The team is tuning inference settings and is testing values like Top P 0.85 and 0.35 to balance variety with accuracy. They want to understand how changing Top P affects which tokens the model can select when generating text. How does the Top P parameter influence responses during inference in Amazon Bedrock?

• ❏ A. Sets the sequences that, when produced, cause generation to halt

• ❏ B. Applies a probability threshold so the model samples from the smallest set of tokens whose cumulative probability reaches the Top P value

• ❏ C. Controls the count of top-probability candidates the model considers for the next token

• ❏ D. Limits the total number of tokens the model can generate in the response

Brightvale Furnishings prepares demand forecasts every two months to plan inventory and staffing using machine learning models. An AI practitioner must deliver a stakeholder-friendly report that emphasizes transparency and model explainability for the trained models. What should the practitioner include to best satisfy these transparency and explainability goals?

• ❏ A. Source code of the training pipeline

• ❏ B. Partial dependence plots (PDPs)

• ❏ C. A small sample of the training dataset

• ❏ D. Confusion matrix and ROC curve charts

An e-commerce analytics startup is evaluating Amazon Bedrock to build generative AI features for tailored product guidance and sales forecasting. They plan to adapt foundation models with their private catalog descriptions and support transcripts. The team expects to first expose the model to about 12 GB of domain text and later train on roughly 4,000 labeled prompt and response pairs to specialize on support workflows. They want to understand how the model customization approaches in Amazon Bedrock differ in the type of data they require. Which statement is accurate?

• ❏ A. Continued pre-training uses labeled data for pre-training and fine-tuning also uses labeled data to train a model

• ❏ B. Continued pre-training uses unlabeled data for pre-training and fine-tuning also uses unlabeled data to train a model

• ❏ C. Continued pre-training relies on unlabeled data for pre-training, while fine-tuning trains with labeled data

• ❏ D. Continued pre-training uses labeled data for pre-training, while fine-tuning trains with unlabeled data

At Luna Insights, a product owner wants a quick, no-code way to try different prompts and adjust settings such as temperature and max tokens when evaluating foundation models in Amazon Bedrock. What best describes Amazon Bedrock Playground?

• ❏ A. It captures and audits prompt activity across accounts using AWS CloudTrail

• ❏ B. A tool that creates serverless inference endpoints and manages runtime parameter caching

• ❏ C. A browser-based workspace to experiment with prompts and adjust model parameters without writing code

• ❏ D. An automated capability that fine-tunes models and promotes deployments across several AWS Regions

A meal delivery platform plans to train a model that tags each dish with exactly one cuisine, choosing one of four options such as Italian, Mexican, Thai, or Greek. Which classification technique is the most appropriate?

• ❏ A. Binary classification

• ❏ B. Single-label multiclass classification

• ❏ C. Amazon Comprehend

• ❏ D. Multi-label classification

A regional credit union is piloting an underwriting assistant on Amazon Bedrock to score small business loan risk. During a 45-day sandbox phase, the analytics team wants to validate quality, fairness, and accuracy using the right Bedrock evaluation approaches and datasets before promoting to production. Which statements about evaluating models in Amazon Bedrock are correct? (Choose 2)

• ❏ A. For human evaluation, you can use either built-in prompt datasets or your own

• ❏ B. Automated model evaluation in Bedrock generates scores using metrics such as BERTScore and F1

• ❏ C. You should use Amazon Bedrock Guardrails to compute fairness metrics and accuracy scores for model evaluation

• ❏ D. Human reviews in Bedrock are suited to qualitative judgments, while automated evaluations focus on quantitative metrics

• ❏ E. Human model evaluation provides statistical scores such as BERTScore and F1

A global e-learning platform uses Amazon Bedrock to produce localized subtitles for training videos. The translations are grammatically correct but lack regional idioms and the right tone, so viewers in some locales find them unnatural. What change should the team implement to add culturally appropriate nuance to the subtitles?

• ❏ A. Temperature adjustment

• ❏ B. Fine-tune the model with locale-specific training data

• ❏ C. BLEU score optimization

• ❏ D. Knowledge Bases for Amazon Bedrock

A retail analytics startup is piloting a foundation model to classify customer feedback as positive, neutral, or negative. Leadership wants to minimize legal exposure from unfair or biased predictions across different demographic groups. Which AWS capability should they use to evaluate and reduce bias in the data and model outputs?

• ❏ A. Model Cards

• ❏ B. Guardrails for Bedrock

• ❏ C. SageMaker Clarify

• ❏ D. Amazon Comprehend

A digital marketplace has launched a customer support assistant powered by a large language model. Adversaries attempt to slip past safety policies by mixing German phrases with escape sequences and by submitting prompts encoded in base64 or using URL-style encoding, which the input filter misses. Which techniques reflect typical methods attackers use to evade prompt restrictions? (Choose 2)

• ❏ A. Using RLHF to filter generated tokens after inference

• ❏ B. Encoding the prompt, for example in base64, to conceal harmful directives

• ❏ C. Obfuscating instructions with escape characters or by switching to another language

• ❏ D. Altering decoding settings such as temperature and maximum tokens

• ❏ E. Asking the model to obtain additional AWS IAM permissions for processing data

A product team at Aurora Retail plans to build models with Amazon SageMaker and needs a centralized way to define, version, and share model input features so multiple data science teams can reuse them consistently. Which SageMaker capability should they choose?

• ❏ A. Amazon SageMaker Model Cards

• ❏ B. Amazon SageMaker Feature Store

• ❏ C. Amazon SageMaker Data Wrangler

• ❏ D. Amazon SageMaker Clarify

A staffing agency named Horizon Talent receives tens of thousands of resumes each month as PDF files and needs to automatically extract the text so downstream systems can analyze the content at scale. Which AWS service should they use to perform this PDF text extraction?

• ❏ A. Amazon Comprehend

• ❏ B. Amazon Transcribe

• ❏ C. Amazon Textract

• ❏ D. Amazon Rekognition

Riverton Labs, a mid-size fintech startup, is evaluating Amazon Q Developer to modernize its engineering workflows over the next 90 days. The team wants help with AI-assisted code generation, automating routine tasks, and bringing machine learning guidance into their AWS development process. To confirm the fit, they need a concise summary of what the service can actually do. Which capabilities of Amazon Q Developer would meet these goals? (Choose 2)

• ❏ A. Use natural language to retrieve account-specific AWS cost insights

• ❏ B. Automatically modify AWS resources to implement cost-optimization changes

• ❏ C. Provide built-in dashboards to visualize AWS cost data inside Amazon Q Developer

• ❏ D. Deploy and provision cloud infrastructure on AWS

• ❏ E. Explore and manage your AWS environment and resources using natural language

An online marketplace called Alpine Mart plans to launch a generative AI assistant that can chat with shoppers, interpret their questions, fetch order and shipping details from an external system, and return accurate answers without human escalation. The team is considering Amazon Bedrock Agents for this automation. Which statement best describes how an Amazon Bedrock agent behaves?

• ❏ A. Agents convert user prompts to vector embeddings to speed up retrieval

• ❏ B. Agents perform supervised fine-tuning of foundation model weights while answering

• ❏ C. Agents coordinate the conversation with a foundation model and call external APIs to complete tasks

• ❏ D. Agents are equivalent to Amazon Lex chatbots and do not invoke external systems

A fintech risk assessment startup uses generative AI to produce tailored summaries and insights from client portfolios. The product team wants to reliably raise the quality and relevance of responses by standardizing how they write prompts. To consistently steer the model toward accurate, useful results, which components should every well-formed prompt clearly include?

• ❏ A. Instructions, Hyperparameters, Input data, Output format

• ❏ B. Amazon Bedrock

• ❏ C. Clear instructions, Relevant context, Input data, Output specification

• ❏ D. Instructions, Parameters, Input data, Output indicator

An international procurement team at a consumer electronics manufacturer processes about 18,000 supplier agreements each month for compliance and risk review. They plan to use AWS to automate intake of scanned PDFs, extract key provisions, spot missing terms, and group contracts for attorney review. In validation tests, the classifier repeatedly marks some small-business vendor agreements as high risk because the training data is skewed toward large enterprise contracts. What steps should the team take to reduce bias in the contract classification? (Choose 2)

• ❏ A. Raise the classification confidence threshold to reduce false positives

• ❏ B. Use Amazon SageMaker Clarify to detect bias and guide adjustments to data and training

• ❏ C. Retrain with a more representative dataset that spans regions, industries, company sizes, and contract types

• ❏ D. Remove human review to ensure the AI operates independently

• ❏ E. Relabel only the few misclassified samples without altering the rest of the training data

VariaPay, a regional fintech startup, is building a machine learning model and must prove that all data used for training and inference adheres to internal data governance rules and external regulatory obligations. Which approach provides the most effective foundation for data governance across the data lifecycle?

• ❏ A. Restrict developer access to training data with IAM roles

• ❏ B. Anonymize every dataset before model training

• ❏ C. Implement centralized logging, defined retention schedules, and continuous monitoring for the full data lifecycle

• ❏ D. AWS Lake Formation

A sustainability-focused apparel startup, Meridian Threads, uses Amazon Bedrock to produce seasonal advertising images for a campaign launching in 18 markets. The creative lead wants to state inside the prompt that the model must not include violent, explicit, or hateful visuals, particularly when the request is ambiguous. Which prompt-engineering approach most directly sets these disallowed elements within the prompt?

• ❏ A. Retrieval-augmented prompting with safe style examples fetched at runtime

• ❏ B. Guardrails for Amazon Bedrock

• ❏ C. Negative prompting that specifies visuals to exclude, such as explicit, violent, or hateful content

• ❏ D. Few-shot prompting using pairs of acceptable and unacceptable image descriptions

A regional public transit agency wants to create a machine learning model to predict passenger no-shows using four years of fare and trip history. The operations team has no coding experience and needs a point-and-click interface to prepare data, train, and evaluate the model without writing code. Which AWS service should they use?

• ❏ A. AWS Glue Studio

• ❏ B. Amazon QuickSight

• ❏ C. Amazon SageMaker Canvas

• ❏ D. Amazon Bedrock

An analytics group at a fintech startup is using Amazon SageMaker Autopilot to train a binary fraud detection model in which fraudulent transactions are about 3% and legitimate ones are 97%. The team wants the chosen metric to prioritize correctly identifying the minority positive class over overall correctness. Which evaluation metrics should they emphasize during model selection to address the class imbalance effectively? (Choose 2)

• ❏ A. Log loss

• ❏ B. Balanced accuracy

• ❏ C. Mean squared error (MSE)

• ❏ D. F1 score

• ❏ E. Overall accuracy

A regional e-commerce startup is using an LLM on Amazon Bedrock to label customer comments as positive, neutral, or negative. During a 30-day pilot, they want the model to return the same label whenever the same prompt is submitted across thousands of reviews. Which inference setting should they adjust to increase response determinism?

• ❏ A. Increase the temperature setting

• ❏ B. Raise the top-p value

• ❏ C. Reduce the temperature setting

• ❏ D. Increase the maximum generation length

A creative studio at Aurora Retail plans to use a diffusion-based model to generate product visuals for seasonal advertising. When executives request clear reasoning behind why certain elements appear in the images, what key drawback of this approach is most likely to cause challenges?

• ❏ A. Consistently deterministic images for the same prompt

• ❏ B. Limited ability to scale GPU training capacity on AWS

• ❏ C. Limited interpretability of the model’s image creation process

• ❏ D. Inability to use text and image modalities together

The smallest textual units a model reads and writes, such as words or subword pieces is correct because tokens are the atomic pieces of text that generative models consume and produce.

Tokens can be whole words or smaller subword or character pieces depending on the tokenizer used. Token counts are what most generative AI offerings measure for input and output volume so services like Amazon Bedrock commonly scale pricing and quotas with token usage.

The pre-trained parameter values of a foundation model that can later be fine-tuned is incorrect because that phrase describes model weights and parameters rather than the textual units a model processes.

The dense vector representations that capture word or concept meaning is incorrect because that refers to embeddings, which are numeric vectors derived from text and not the raw tokens themselves.

Amazon Comprehend is incorrect because it is an AWS service for natural language processing tasks and it is not a definition of what a token is.

When estimating cost or limits remember that AWS generative AI services usually meter usage by tokens and tokenizers often split text into subword pieces rather than whole words.

The correct option is Define explicit safety constraints and guardrails within the prompt. This approach embeds the credit union’s policy rules directly into the prompt so the model will produce recommendations that stay within approved guidance.

Encoding constraints, required formats, and example-safe responses in the prompt gives the model clear operational boundaries and practical checks it must follow. You can include forbidden actions, required disclosures, and example interactions that demonstrate compliant behavior and you can ask the model to justify how a recommendation satisfies policy so outputs are more auditable.

Use zero-shot prompting to elicit more direct answers is incorrect because zero-shot prompts do not supply the explicit rules or examples needed to constrain behavior and they tend to increase output variability when strict compliance is required.

Increase the model’s input and output token limits is incorrect because changing token limits only affects how much context the model can process and how long responses can be and it does not impose safety constraints or change the model’s adherence to policy.

Amazon GuardDuty is incorrect because GuardDuty is an AWS threat detection service for accounts and workloads and it is not a prompt-engineering or model-steering tool, so it will not enforce compliance in model outputs.

When a question asks how to limit model behavior look for language about guardrails or explicit constraints in the prompt and prefer answers that place rules and examples inside the prompt itself.

The best choices are Using a carefully labeled, high-quality training and reference image set and Fine-tuning the diffusion model on brand-specific examples. These two approaches directly increase image fidelity and produce consistent brand visuals for 6K product ads and social media content.

Using a carefully labeled, high-quality training and reference image set gives the model clear examples of lighting composition color palettes and product details that you want it to reproduce. High resolution images and consistent labels reduce artifacts and guide the model toward the specific visual features needed for high quality 6K outputs.

Fine-tuning the diffusion model on brand-specific examples adapts the foundation model to the retailer’s exact aesthetic and product characteristics. Fine tuning reduces off brand variations and improves reproducibility across different prompts and campaign assets which leads to stronger brand consistency.

Tracking inference latency and throughput is helpful for deployment planning and meeting SLAs but it does not by itself improve the visual quality or the brand alignment of generated images.

Optimizing prompt token usage to reduce size can save cost and sometimes speed up generation but shortening or compressing prompts does not replace the need for high quality training data and model adaptation when the goal is better imagery.

Expanding the model’s context window length is a concept that mainly applies to large language models and not to image diffusion models. Increasing context window is unlikely to materially affect image fidelity or brand consistency for diffusion based generation.

On image generation questions prioritize data quality and model alignment as the primary levers for better visuals rather than operational metrics or LLM specific parameters.

The correct selections are Configure Amazon Q Business to use enterprise data only or combine enterprise data with model knowledge and Amazon Q Business guardrails provide topic-level rules that define how the app responds when a blocked topic is mentioned.

Administrators can set Configure Amazon Q Business to use enterprise data only or combine enterprise data with model knowledge as a response-source control so answers come only from your approved enterprise data or from a blend of enterprise data and the model knowledge. This setting helps meet compliance needs by limiting external model provenance when required and by allowing a mixed mode for broader answers during pilot testing.

Amazon Q Business guardrails provide topic-level rules that define how the app responds when a blocked topic is mentioned gives admins the ability to define topic-specific policies that alter behavior when certain subjects are detected. Guardrails can block content or return a controlled response so the application behaves predictably whenever a restricted or sensitive topic appears.

Amazon Q Business can be configured to answer using only the model’s built-in knowledge is incorrect because there is not a mode that limits answers to model-only knowledge in isolation. The supported administrative choices are enterprise-only or a combination of enterprise data and model knowledge.

End users can never upload files in chat to generate answers from those files is incorrect because file upload behavior is governed by admin settings. Administrators can enable or restrict uploads and control how uploaded data is used for responses and logging.

AWS WAF is incorrect because the web application firewall is not the control plane for Amazon Q Business guardrails or response-source policies. AWS WAF protects web traffic and is not used to configure Q Business response-source or topic-level guardrail settings.

Before a pilot, verify the response-source setting and the topic-level guardrails in the admin console and confirm file upload and audit logging policies to meet compliance requirements.

The correct ordering is Artificial Intelligence > Machine Learning > Deep Learning > Generative AI models. This option lists the broadest discipline first and the most specialized capability last.

Artificial Intelligence is the umbrella field that covers systems that perform tasks which normally require human intelligence and Machine Learning is a subset of AI that focuses on algorithms that learn patterns from data. Deep Learning is a further subset of ML that uses multi layer neural networks to learn hierarchical features and Generative AI models are specialized deep learning models trained to create new content such as text and images. This nesting explains why the chosen ordering narrows from general to specific.

Machine Learning > Deep Learning > Artificial Intelligence > Generative AI is incorrect because it treats Machine Learning as broader than Artificial Intelligence and that inverts the superset relationship. AI is the larger domain and ML sits within it.

Generative AI > Deep Learning > Machine Learning > Artificial Intelligence is wrong because it elevates Generative AI above its parent disciplines. Generative AI is a niche within deep learning and it cannot logically be the broadest category.

Artificial Intelligence > Generative AI > Machine Learning > Deep Learning is incorrect because it places Generative AI as broader than Machine Learning and Deep Learning. That ordering reverses the usual nesting where GenAI sits inside DL which sits inside ML which sits inside AI.

Read from general to specific and pick the choice that narrows scope from AI to ML to DL to GenAI.

Amazon Bedrock model evaluation with pre-built bias and fairness prompt datasets is the correct option because it offers a managed, low maintenance way to evaluate foundation model outputs for bias and representational balance.

Amazon Bedrock model evaluation with pre-built bias and fairness prompt datasets provides standardized prompt sets and automated metrics that surface fairness issues and representational gaps without requiring custom pipelines or bespoke survey design. Using Bedrock enables consistent, repeatable checks across many contracts and yields quantitative signals that the legal team can review with minimal administrative overhead.

Launch a limited pilot and gather structured bias feedback through targeted surveys is not ideal because survey responses are subjective and you must design, run, and analyze the surveys which reduces consistency and increases effort.

Continuously fine-tune the model using recent responses from a diverse user group describes a remediation and training approach rather than an evaluation method because it requires labeled data, retraining, and validation and it adds ongoing maintenance that the question asks to avoid.

Amazon SageMaker Clarify is focused on bias detection for datasets and supervised model predictions and it generally involves more setup and is not purpose built for prompt based evaluations of foundation model outputs.

When a question stresses minimal administrative effort and repeatable fairness checks choose managed, built in evaluation tools such as pre built prompt datasets rather than custom surveys or continuous retraining.

The correct choice is Foundation models are pre-trained on large, diverse datasets and can be fine-tuned or guided with prompts to handle many downstream tasks. These models learn broad patterns during large scale pre training and can be adapted through fine tuning or prompt engineering to produce personalized study guides and draft lesson materials.

Pre training on diverse data gives foundation models a general understanding of language and knowledge and this lets them be reused across tasks. Fine tuning or prompt guidance lets you steer outputs toward a subject domain or individual learner needs and retrieval augmented generation helps ground responses in course content and learner records.

Foundation models cannot personalize outputs based on learner interactions is false because models can accept context and user data and they can be combined with prompting and retrieval to produce tailored content.

Each foundation model is built for a single narrow use and cannot be adapted to other applications is incorrect since a key advantage of foundation models is generality and reuse across many tasks and domains.

On Amazon Bedrock, foundation models must be retrained from scratch for every subject domain is wrong because Amazon Bedrock provides managed access to multiple foundation models and supports customization methods that do not require full retraining from the ground up.

Look for choices that mention pre trained and fine tuning or broad applicability when identifying foundation models and mark statements about single purpose models or mandatory full retraining as likely incorrect.

The most likely cause is The model has overfit to the training data. This scenario explains why the model reports very high accuracy on the training set but performs poorly on live customer orders from the following 45 days.

The model has overfit to the training data means the model learned noise and idiosyncratic patterns in the training examples rather than generalizable signals that apply to new orders. When real customer behavior or product demand shifts slightly the overfit model fails to generalize and its real world performance drops, and common remedies include collecting more diverse data, using regularization, applying cross validation, and simplifying the model.

The training dataset was missing labels is unlikely because missing labels normally prevent successful supervised training or cause low training accuracy rather than producing near perfect training scores.

Amazon Forecast is simply the name of a forecasting service and it does not explain a generalization error or why a model would have high training accuracy yet fail on new data.

The model is underfitting the training data is incorrect because underfitting produces low accuracy on both the training and test sets and would not result in 99% training accuracy.

When you see very high training accuracy and much lower validation or test accuracy think overfitting and focus on techniques like more varied data, regularization, and cross validation to improve generalization.

The correct options are Amazon Textract, Amazon Comprehend, and a Generative AI summarization chatbot.

Amazon Textract performs OCR on scanned PDFs and images and extracts machine readable text, key value pairs, and tables so the 80,000 pages per month become structured data that downstream processing can use. Amazon Comprehend analyzes that text to detect named entities, key phrases, and dates so it can identify parties, effective and renewal dates, and other critical terms. A Generative AI summarization chatbot can then synthesize clauses, normalize extracted fields, and produce concise summaries that accelerate human review while preserving accuracy.

Amazon Polly is a text to speech service so it converts text into audio and does not perform OCR or extract entities or clauses from documents.

Amazon Personalize focuses on personalization and recommendations and it does not provide document parsing or NLP extraction for named entities and dates.

Convolutional Neural Network (CNN) refers to a model family used for vision tasks and it is not a managed, end to end AWS service for document information extraction so it is not the best choice for this use case.

For scanned documents use Amazon Textract for OCR, then apply Amazon Comprehend for entity and date extraction, and consider a generative AI summarization chatbot to normalize and summarize results for reviewers.

The correct choice is Structured data conforms to a defined schema, often as rows and columns that are easy to query and aggregate, while unstructured data lacks a fixed model and includes items like text, images, audio, and video. In this scenario the relational records represent schema based data and the captions thumbnails and audio are schema free assets that need different preprocessing.

Structured data adheres to a known schema so you can filter join and aggregate it with SQL style operations to build features directly from columns and rows. Unstructured data does not follow a rigid model so it usually requires feature extraction such as text tokenization and embeddings or image and audio feature engineering before it becomes suitable for model training.

Structured data is typically freeform text with no specific organization, while unstructured data is arranged in a fixed tabular layout is incorrect because it reverses the correct definitions of structured and unstructured data.

Structured data must reside in Amazon RDS, and unstructured data must be stored only in Amazon S3 and cannot be queried is incorrect because both data types can be stored and queried across AWS services and S3 objects can be queried with tools like Amazon Athena and cataloged with AWS Glue.

Structured data is only used to train models, whereas unstructured data is kept solely for archival purposes is incorrect because both structured and unstructured data are commonly used for analytics and machine learning workflows after appropriate preprocessing.

Remember that schema driven data maps to SQL style feature engineering and schema free content needs feature extraction such as NLP or embeddings before model training.

The correct choice is Amazon SageMaker JumpStart. It provides a catalog of prebuilt foundation models and guided workflows to deploy them rapidly with minimal setup and to expose the model through a private endpoint inside your team VPC.

Amazon SageMaker JumpStart supplies deployment templates and example artifacts that let you select a foundation model and launch a real time endpoint quickly. The JumpStart workflows automate model selection and common configuration steps and they support configuring the endpoint networking to use your subnets and security groups so the endpoint can be private to your VPC. This combination of a model catalog and guided deployment is what enables a rapid private deployment in the stated time frame.

Amazon SageMaker endpoints are the hosting and inference mechanism that run models in SageMaker but they do not include a built in catalog of foundation models or the guided deployment templates that JumpStart provides. Using only Amazon SageMaker endpoints requires you to bring or train a model and then configure the endpoint which increases setup time and effort.

Amazon Personalize is a managed service focused on personalization and recommendation use cases and it is not a general foundation model deployment solution for arbitrary models inside a VPC.

PartyRock, an Amazon Bedrock Playground is an educational sandbox and it is not intended for VPC integrated private inference or production deployments. That makes it unsuitable when the requirement is a private team endpoint inside a VPC.

When a question emphasizes fast private deployment from within a VPC choose a solution that bundles a model catalog with guided deployment such as JumpStart.

Model architecture and capabilities, Inference latency and output quality metrics, and Price per image or token usage are the correct criteria to emphasize when comparing foundation models in Amazon Bedrock for high resolution marketing visuals.

Model architecture and capabilities determine whether a model can produce photorealistic images and support the style controls and guidance features you need. You should evaluate support for high output resolution prompt conditioning fine tuning or adapters and any built in tools for controlling composition and color because these factors directly affect the final marketing imagery.

Inference latency and output quality metrics show how the model performs in real world workflows and how visually convincing the images are. Measure latency under expected load and track image quality using automated metrics such as FID and complemented by human review because automated scores do not always capture aesthetic suitability for marketing use.

Price per image or token usage matters for a 90 day comparison because cost scales with volume and with provider specific pricing in Bedrock. Compare estimated costs for your expected output rates and include any extra costs for higher resolutions or multiple generation passes so you avoid surprises when you scale.

BLEU score is incorrect because it is a text based metric for machine translation and other NLP tasks and it does not assess image quality or visual fidelity.

Amazon Rekognition is incorrect because it is an image analysis service rather than a generative foundation model and it is not an evaluation criterion for selecting which Bedrock image model will create marketing visuals.

When you run the model comparison create a small representative image set and measure visual quality with automated metrics and human review while recording latency and cost per image to reflect production constraints.

Curate a more representative training dataset is the most effective first step because generative models learn statistical patterns from their training data and improving data representation reduces the chance the model will reproduce gender stereotypes.

Ensuring balanced and diverse examples and filtering harmful patterns addresses the root cause of biased outputs and gives you a solid foundation for later actions. Curating the dataset reduces harmful associations and makes subsequent steps such as fine tuning and evaluation more reliable.

Increase the model’s temperature setting only alters randomness and lexical variety and does not change the model’s learned associations so the same stereotypes can still appear.

Conduct subgroup bias analysis is useful for detecting and monitoring disparities across demographics but it measures the problem rather than remediating the training data that causes biased generation.

Fine-tuning the model can be effective when performed with a curated and inclusive dataset and with careful evaluation but fine tuning without improving data quality risks reinforcing existing biases and should follow data remediation and testing.

Prioritize data quality and representation before adjusting model parameters and use bias analysis to measure progress and guardrails for runtime controls.

On-demand pricing is the correct choice because it allows the team to pay only for actual Amazon Bedrock usage during the nine month pilot and it avoids any upfront capacity purchases or long term commitments.

The On-demand pricing model charges per usage so costs scale with activity and it fits workloads that vary widely from week to week and it enables experimentation without capacity planning.

Provisioned throughput is meant for steady predictable workloads because it reserves capacity ahead of time and it can waste money when usage is sporadic.

EC2 Spot Instances provide discounted EC2 compute that can be interrupted and they are unrelated to Bedrock pricing so they are not applicable for Bedrock usage.

EC2 Reserved Instances require one or three year commitments to reduce EC2 compute costs and they do not apply to Bedrock pricing so they are not suitable for a short flexible pilot.

Match pricing to variability and choose options that let you pay only for what you use when demand is unpredictable.

Applies a probability threshold so the model samples from the smallest set of tokens whose cumulative probability reaches the Top P value is correct. This describes how Top P implements nucleus sampling and limits selection to the smallest group of tokens whose probabilities sum to the given threshold.

When you raise Top P the model can draw from a larger cumulative probability mass and that usually increases variety in outputs while still excluding very low probability tokens. When you lower Top P the model restricts sampling to a smaller high probability set and that tends to make responses more conservative and predictable.

Sets the sequences that, when produced, cause generation to halt is incorrect because that behavior belongs to stop sequences and it stops output rather than changing which candidate tokens are considered.

Controls the count of top-probability candidates the model considers for the next token is incorrect because that describes Top K which fixes a number of candidates rather than using a cumulative probability cutoff.

Limits the total number of tokens the model can generate in the response is incorrect because that describes max tokens which restricts output length and does not affect the sampling pool for each token.

When tuning sampling remember that Top P controls cumulative probability mass and Top K controls a fixed candidate count. Try a small grid of values to find the balance of diversity and accuracy that fits your use case.

The correct option is Partial dependence plots (PDPs). Partial dependence plots (PDPs) illustrate the marginal effect of individual features on the model�s output and make it easier for stakeholders to see how inputs drive predictions across their ranges.

PDPs are useful for regression and time series forecasting because they show how changing one feature while averaging out others impacts predicted values. They are model agnostic and produce visual, stakeholder friendly explanations that highlight feature influence and expected direction of effect.

Source code of the training pipeline is aimed at developers and exposes implementation details rather than providing concise, interpretable summaries of how features affect predictions. Including code can confuse business stakeholders and it does not directly show feature influence.

A small sample of the training dataset can create privacy and confidentiality risks and it only provides example inputs rather than explaining the model�s reasoning about feature effects. A data sample does not convey how changes in a feature change model outputs across the input space.

Confusion matrix and ROC curve charts are classification focused metrics and visualizations and they do not explain feature�s marginal effects for regression or forecasting tasks. These charts describe predictive performance rather than the relationship between inputs and predictions.

For stakeholder reports use visual explanations that highlight feature effects such as Partial dependence plots or SHAP values and avoid raw code or raw data samples that are technical or risky.

Continued pre-training relies on unlabeled data for pre-training, while fine-tuning trains with labeled data. This answer fits the startup scenario where about 12 GB of domain text can be used to adapt a foundation model and the later set of roughly 4,000 labeled prompt and response pairs can be used to specialize the model for support workflows.

In Amazon Bedrock continued pre-training adapts a foundation model to the startup’s domain by ingesting large volumes of unlabeled text and updating model parameters so the model better represents the domain language and style. Fine-tuning then trains with labeled input and output pairs so the model learns the specific mapping required for tasks such as guided responses or ticket resolution.

Continued pre-training uses labeled data for pre-training and fine-tuning also uses labeled data to train a model is incorrect because continued pre-training is typically self supervised and does not require labeled examples. The primary purpose of continued pre-training is domain adaptation using unlabeled corpora.

Continued pre-training uses unlabeled data for pre-training and fine-tuning also uses unlabeled data to train a model is incorrect because fine-tuning relies on labeled examples to teach the model the desired input to output behavior. Using only unlabeled data for the fine-tuning step would not provide the supervised signal needed for task specialization.

Continued pre-training uses labeled data for pre-training, while fine-tuning trains with unlabeled data is incorrect because it reverses the real roles of the two approaches. Continued pre-training is the unlabeled domain adaptation step and fine-tuning is the supervised task specialization step.

Keep in mind that continued pre-training is best for adapting models with large unlabeled corpora and fine-tuning is used when you have labeled examples that define the desired task behavior.

A browser-based workspace to experiment with prompts and adjust model parameters without writing code is correct. It best describes Amazon Bedrock Playground.

The A browser-based workspace to experiment with prompts and adjust model parameters without writing code gives product owners a no-code web interface to try prompts, tweak parameters such as temperature and max tokens, and compare outputs before integrating models into applications. The Playground is meant for quick experimentation and side by side comparisons rather than deployment or audit logging.

It captures and audits prompt activity across accounts using AWS CloudTrail is incorrect because AWS CloudTrail captures API calls and audit logs while the Playground is an interactive console feature for trying prompts.

A tool that creates serverless inference endpoints and manages runtime parameter caching is incorrect because provisioning endpoints and managing runtime caching are deployment concerns handled by inference services and APIs, not the Playground interface.

An automated capability that fine-tunes models and promotes deployments across several AWS Regions is incorrect because the Playground does not perform automated fine tuning or manage multi region rollouts.

When a question mentions no-code or prompt experimentation think of the Playground and not deployment or logging features.

The correct choice is Single-label multiclass classification. This approach fits because the problem requires assigning exactly one cuisine from a fixed set of four options to each dish.

Single-label multiclass classification trains a model to map each input to one and only one class when there are three or more possible classes. In practice you label each dish with a single cuisine and train with a categorical cross entropy loss and a softmax output so the model predicts one class per example.

Binary classification is not appropriate because it only models two possible outcomes and cannot directly handle a choice among four cuisines.

Amazon Comprehend is an AWS natural language processing service rather than a conceptual model type. It can perform text classification as a service but it is not the classification technique that defines single versus multi label setups.

Multi-label classification allows multiple labels per example and therefore contradicts the requirement that each dish receive exactly one cuisine.

Match the required label cardinality to the technique and pick single-label multiclass when each instance must have exactly one label.

Automated model evaluation in Bedrock generates scores using metrics such as BERTScore and F1 and Human reviews in Bedrock are suited to qualitative judgments, while automated evaluations focus on quantitative metrics are the correct statements.

Automated model evaluation in Bedrock generates scores using metrics such as BERTScore and F1 is correct because Bedrock can run automatic evaluation jobs that compute standardized, quantitative metrics for supported tasks and produce objective report cards that help compare model outputs at scale.

Human reviews in Bedrock are suited to qualitative judgments, while automated evaluations focus on quantitative metrics is correct because human review workflows capture subjective assessments such as coherence, relevance, and usefulness and they surface issues like contextual bias and nuance that numeric metrics may miss.

For human evaluation, you can use either built-in prompt datasets or your own is incorrect because built-in prompt datasets are provided for automated evaluation jobs and human review workflows typically require you to supply or ingest your own review dataset.

You should use Amazon Bedrock Guardrails to compute fairness metrics and accuracy scores for model evaluation is incorrect because Bedrock Guardrails enforce safety and policy constraints and they do not compute evaluation metrics such as BERTScore or F1 for model comparison.

Human model evaluation provides statistical scores such as BERTScore and F1 is incorrect because those statistical scores are produced by automated evaluation processes rather than by human raters who provide qualitative judgments and structured labels.

Use automated evaluation for standardized metrics and speed and use human review for subjective quality and bias checks. Remember that built-in datasets apply to automated evaluation jobs only.

The correct choice is Fine-tune the model with locale-specific training data. Fine tuning on curated subtitle corpora for each region teaches idioms and the preferred tone so translations feel natural to local viewers.

By applying fine-tuning the model learns recurring regional expressions and style from examples so it can generate consistent idiomatic phrasing rather than only literal translations. This approach lets you control register and cultural references by including annotated, locale specific examples during training.

Temperature adjustment only changes output randomness and variety and does not teach the model regional idioms or a consistent tone. It will not reliably produce culturally appropriate phrasing.

BLEU score optimization is a metric and a training objective that measures overlap with references. It does not itself add idiomatic style or cultural nuance to outputs.

Knowledge Bases for Amazon Bedrock can provide factual context via retrieval but they do not inherently change the model s language style or introduce idiomatic local phrasing unless you also customize the model with locale specific data.

If translations are correct but sound unnatural use fine-tuning with curated, locale specific subtitle examples to capture tone and idioms rather than changing randomness

The correct choice is SageMaker Clarify. This service is designed to compute dataset and prediction bias metrics and to provide explainability that helps detect and mitigate unfair outcomes and reduce legal and compliance risk.

SageMaker Clarify can measure disparities across demographic groups and produce fairness metrics and explanations that show which features influence predictions. Teams can use the metrics and explanations to compare groups, identify biased behavior, and apply preprocessing or postprocessing mitigation techniques to improve fairness.

Model Cards focus on documenting model lineage, intended use, and evaluation metrics for governance and transparency, and they do not perform statistical bias analysis or mitigation by themselves.

Guardrails for Bedrock provides content safety and output filtering for foundation model interactions, and it is not a substitute for the statistical fairness testing and explainability that Clarify provides.

Amazon Comprehend offers managed NLP features such as sentiment analysis and entity recognition, and it does not provide the bias measurement and mitigation tooling needed to evaluate model fairness across demographic groups.

When a question mentions fairness bias metrics or explainability think of SageMaker Clarify for measurement and mitigation and think of Guardrails for Bedrock for content safety and filtering.

Encoding the prompt, for example in base64, to conceal harmful directives and Obfuscating instructions with escape characters or by switching to another language are the correct options because they describe common obfuscation methods attackers use to hide intent from simple keyword filters.

Encoding the prompt is effective when input sanitizers do not decode incoming text before checking it so the hidden directives pass through initial filters and can be revealed or executed later in the processing pipeline.

Obfuscating instructions leverages escape characters or language switching to break recognizable patterns so that keyword based detectors miss them while the model can still interpret the instruction when it encounters the obfuscated sequence.

Using RLHF to filter generated tokens after inference is not an attacker technique because RLHF is a developer driven alignment approach used during training and evaluation and it is not a method for bypassing input filters.

Altering decoding settings such as temperature and maximum tokens affects response variability and length and does not conceal prompt content from input sanitization or change how filters parse incoming text.

Asking the model to obtain additional AWS IAM permissions for processing data is irrelevant because IAM access is controlled by the cloud identity and access management system and the model cannot grant or acquire those permissions on its own.

When you see encoding or escape characters in a scenario think obfuscation and rule out choices about model tuning or cloud IAM.

The correct choice is Amazon SageMaker Feature Store. It provides a centralized and versioned repository for defining storing and sharing model input features so multiple data science teams can reuse them consistently.

Amazon SageMaker Feature Store separates feature engineering from training and serving and supports both offline and online stores. It offers feature versioning access controls and integration with SageMaker training pipelines so teams can discover reuse and maintain consistent feature definitions across experiments.

Amazon SageMaker Model Cards captures documentation and governance details about models but it does not manage reusable feature data.

Amazon SageMaker Data Wrangler simplifies data preparation and transformations but it is not a centralized feature registry for cross team reuse.

Amazon SageMaker Clarify focuses on bias detection and explainability and it does not provide storage or sharing of features for reuse.

When a question lists centralized feature definitions versioning team reuse or both online and offline access think Feature Store.

The correct option is Amazon Textract. It is the service built to perform OCR on document files such as PDFs so resumes can be converted to searchable text and structured data for downstream analysis.

Amazon Textract extracts plain text and structured elements like forms and tables and it supports large scale and automated workflows so a staffing agency can process tens of thousands of resumes each month.

Amazon Comprehend is incorrect because it performs natural language processing on text that is already available and it is intended to be used after OCR rather than for PDF ingestion.

Amazon Transcribe is incorrect because it converts spoken audio into text and it is not designed to extract text from PDF documents.

Amazon Rekognition is incorrect because it focuses on image and video analysis and it is not the appropriate service for extracting text and document structure from PDFs.

When a question asks about extracting text and structure from PDFs choose Amazon Textract and remember that Comprehend analyzes text after extraction while Transcribe handles audio.

Use natural language to retrieve account-specific AWS cost insights and Explore and manage your AWS environment and resources using natural language are correct because Amazon Q Developer can answer account specific cost and usage questions and can query and describe resources in your AWS account using plain language.

Use natural language to retrieve account-specific AWS cost insights is supported through integration with AWS Cost Explorer so Q Developer can surface cost and usage details and explain billing trends and anomalies for your account in conversational form.

Explore and manage your AWS environment and resources using natural language lets you ask about resources and get listings, descriptions, and console deep links so your team can understand resource configurations and find the right places to act from within the developer experience.

Automatically modify AWS resources to implement cost-optimization changes is incorrect because Q Developer can recommend actions but it does not automatically change or apply modifications to your resources for you.

Provide built-in dashboards to visualize AWS cost data inside Amazon Q Developer is incorrect because visualizations and dashboards are provided by AWS Cost Explorer and not rendered as built in dashboards inside Q Developer.

Deploy and provision cloud infrastructure on AWS is incorrect because provisioning and deployment are handled by services like AWS CloudFormation and the AWS Cloud Development Kit and not by Q Developer directly.

Think of Amazon Q Developer as an AI assistant that can query, explain, and link to account cost and resource information but it will not deploy or automatically change your infrastructure.

The correct choice is Agents coordinate the conversation with a foundation model and call external APIs to complete tasks. This option matches the described Alpine Mart assistant that needs to chat with shoppers, interpret questions, fetch order and shipping details from external systems, and respond without human escalation.

Agents orchestrate a foundation model to reason about user intent and they invoke configured actions or API calls to retrieve real data and complete multi step workflows. This design combines model understanding with external system integration so the assistant can produce accurate, up to date answers and perform tasks such as looking up orders and shipping status.

Agents convert user prompts to vector embeddings to speed up retrieval is incorrect because generating embeddings and performing vector search are retrieval mechanisms and they do not provide the orchestration or external API invocation that agents perform. Embeddings help find relevant content but they are not the orchestration layer.

Agents perform supervised fine-tuning of foundation model weights while answering is incorrect because agents use pre trained foundation models at inference and they do not fine tune model weights during runtime. Fine tuning is an offline process and it is separate from agent execution.

Agents are equivalent to Amazon Lex chatbots and do not invoke external systems is incorrect because Amazon Lex is a separate conversational service and it does not natively provide the same agent orchestration for invoking external actions. Agents are built to call APIs and integrate with backend systems which goes beyond a basic Lex chatbot capability.

When a question emphasizes coordinating steps, invoking external APIs, and using a foundation model to reason choose Agents rather than options about embeddings or runtime fine tuning.

Clear instructions, Relevant context, Input data, Output specification is the correct option because it lists the four parts that make prompts clear and repeatable.

A well formed prompt begins with Clear instructions that define the task and any constraints. It then provides Relevant context that grounds the model with background or domain details. The prompt must include the Input data that the model should analyze or summarize. Finally the prompt specifies the Output specification to define the expected structure and format of the response.

Instructions, Hyperparameters, Input data, Output format is incorrect because hyperparameters are configuration knobs for training or inference and they are not part of prompt wording. The wording output format is similar to output specification but the inclusion of hyperparameters makes this choice wrong.

Amazon Bedrock is incorrect because it names an AWS service for hosting and accessing models rather than a set of prompt components. It is not a prompt structure and it is therefore not the right answer for prompt composition.

Instructions, Parameters, Input data, Output indicator is incorrect because parameters refers to internal learned model values that you do not include in a prompt. The term output indicator is also vague compared with a clear Output specification and so this option is not as accurate.

Memorize the four prompt parts and watch for distractors that mention parameters or hyperparameters since those refer to model settings rather than prompt text.

Use Amazon SageMaker Clarify to detect bias and guide adjustments to data and training and Retrain with a more representative dataset that spans regions, industries, company sizes, and contract types are correct because they target the root causes of biased classifications rather than only changing decision thresholds or removing oversight.

Use Amazon SageMaker Clarify to detect bias and guide adjustments to data and training supplies dataset and model bias metrics and explainability so the team can identify which features and vendor segments cause false high risk labels and then update labeling, sampling, or feature engineering to mitigate those effects.

Retrain with a more representative dataset that spans regions, industries, company sizes, and contract types addresses representation gaps so the classifier learns patterns that generalize to small business agreements and reduces systematic misclassification that stems from training mostly on large enterprise contracts.

Raise the classification confidence threshold to reduce false positives only shifts the decision boundary and may lower some false positives but it does not remove biased signals in the data and can create more false negatives for genuinely risky contracts.

Remove human review to ensure the AI operates independently eliminates important governance and feedback loops. Human-in-the-loop review is necessary to validate outputs and to provide the corrective labels and guidance that improve fairness over time.

Relabel only the few misclassified samples without altering the rest of the training data treats symptoms rather than causes. Fixing a small set of labels will not correct systemic dataset imbalance and is unlikely to stop recurring bias against underrepresented vendor types.

Prioritize improving data representativeness and using tools that report bias metrics and explainability when a question mentions fairness and bias.

Implement centralized logging, defined retention schedules, and continuous monitoring for the full data lifecycle is the correct option because it establishes an end to end governance foundation that supports auditability and regulatory proof across collection, storage, processing, model training, and deletion.

Implement centralized logging, defined retention schedules, and continuous monitoring for the full data lifecycle creates immutable audit trails and consistent retention enforcement, and it enables automated alerts for misuse or policy drift so the organization can demonstrate compliance during audits and investigations.

Restrict developer access to training data with IAM roles is important for applying least privilege, but it does not by itself provide lifecycle controls, comprehensive logging, or organization wide retention policies that prove how long data was kept and when it was removed.

Anonymize every dataset before model training can reduce exposure of sensitive attributes, but anonymization alone does not provide data lineage, retention records, or continuous monitoring and some models require partially identifiable features for correct behavior.

AWS Lake Formation helps with permissions and cataloging inside data lakes, but it is a component level service and does not by itself deliver centralized, organization wide logging, retention scheduling, and continuous monitoring across all data sources and services.

Look for answers that cover end to end controls such as logging, retention, and monitoring across the full lifecycle rather than single layer fixes.

The correct approach is Negative prompting that specifies visuals to exclude, such as explicit, violent, or hateful content. This option most directly embeds explicit do not include instructions inside the prompt so the model is instructed not to generate those elements.

Negative prompting works by listing disallowed content inside the prompt text and it provides a clear and immediate boundary the model can follow when a request is ambiguous or broad. Placing exclusions in the prompt is the most direct way to ensure the generation avoids explicit, violent, or hateful imagery for a multinational campaign.

Retrieval-augmented prompting with safe style examples fetched at runtime can help ground the model with desirable styles and references but it does not itself place explicit prohibitions inside the prompt. This makes it less direct for enforcing disallowed content.

Few-shot prompting using pairs of acceptable and unacceptable image descriptions can guide the model by example, yet it relies on examples that may still leave room for ambiguity and it is not as explicit as stating exclusions directly in the prompt.

Guardrails for Amazon Bedrock are valuable as platform level safety and policy controls and they help enforce rules across requests, but they do not satisfy the requirement to set the boundaries inside the prompt itself. Guardrails complement negative prompting but they are not the in prompt mechanism the question asks for.

When a question asks about placing restrictions inside the prompt prefer answers that explicitly embed exclusions and avoid options that only provide examples or platform level enforcement.

The correct option is Amazon SageMaker Canvas. It matches the requirement for a point and click, no code interface that lets non technical operations staff prepare data, train models, evaluate accuracy, and generate predictions for tabular fare and trip history.

Amazon SageMaker Canvas offers a guided visual workflow where users can import data from files or AWS sources, run automated modeling, compare model metrics, and export predictions without writing code. The service is aimed at business users and it integrates with the broader SageMaker ecosystem when more advanced work is needed.

AWS Glue Studio focuses on extract transform load development and job orchestration for data pipelines. It is not intended as a no code environment for training and evaluating supervised machine learning models so it does not meet the requirement.

Amazon QuickSight provides dashboards and visual analytics and it can surface ML powered insights but it does not provide a full point and click workflow to build and evaluate custom predictive models on tabular data so it is not the right choice.

Amazon Bedrock is targeted at generative AI and foundation models and it is not designed for no code supervised modeling of structured datasets so it does not satisfy the scenario.

When a question highlights no code and a point and click UI for predictive models on structured data choose the managed no code ML tool rather than ETL, BI, or foundation model services.

The correct choices are Balanced accuracy and F1 score.

Balanced accuracy averages recall across the positive and negative classes which ensures the minority fraud class is weighted equally with the majority legitimate class and prevents the 97 to 3 majority from dominating the evaluation.

F1 score is the harmonic mean of precision and recall and it emphasizes detecting rare fraudulent transactions while also controlling false positives which helps when you need both good detection and manageable alert volume.

Overall accuracy is misleading on severely imbalanced datasets because predicting the majority class can yield a high accuracy while missing most minority positives.

Log loss measures probability calibration and penalizes confident wrong predictions but it does not directly prioritize recall for the minority class and a model can have acceptable log loss while still failing to find many fraud cases.

Mean squared error (MSE) is a regression metric and is not appropriate for evaluating binary classification performance under class imbalance.

When classes are highly imbalanced focus on recall for the fraud class and use F1 and balanced accuracy to compare models while also checking the confusion matrix and precision to control false positives.

Reduce the temperature setting is the correct inference setting to adjust to increase response determinism for repeated prompts because temperature controls the sampling randomness and lowering it makes the model favor higher probability tokens so outputs become more stable across identical inputs.

Lowering the temperature moves the model toward greedy or high probability token selection and reduces variation in responses. For a simple classification task like labeling comments as positive neutral or negative this makes the model more likely to return the same label each time the same prompt is submitted, and it affects token selection rather than output length.

Increase the temperature setting is incorrect because raising temperature injects more randomness into token choice and it will make repeated outputs less consistent rather than more consistent.

Raise the top-p value is incorrect because increasing top-p enlarges the candidate token pool for sampling and can increase variability, so it does not improve determinism when you need repeatable labels.

Increase the maximum generation length is incorrect because changing the allowed token length only affects how long the output can be and it does not reduce sampling randomness, so it will not make short classification labels more repeatable.

When you need repeatable outputs keep the temperature low and avoid increasing top-p. Tune length separately if you need more or fewer tokens.

The correct option is Limited interpretability of the model’s image creation process. This limitation is most likely to cause challenges when executives ask for clear reasoning behind why certain elements appear in generated visuals.

Diffusion models construct images by iteratively denoising random noise and the model’s internal decisions are distributed across many sampling steps. That process is powerful yet opaque and it is hard to provide a simple step by step explanation for how a specific attribute emerged which complicates governance and brand accountability.

Consistently deterministic images for the same prompt is incorrect because diffusion models are inherently stochastic and the same prompt can produce different outputs depending on sampling and random seeds.

Limited ability to scale GPU training capacity on AWS is incorrect because cloud providers such as AWS offer scalable GPU resources including Amazon SageMaker and EC2 P family instances that support large scale training and inference.

Inability to use text and image modalities together is incorrect because many diffusion based systems support multimodal conditioning and text to image workflows.

When a question targets generative model weaknesses focus on explainability, bias, hallucination, or compute. For diffusion models remember the iterative denoising process makes outputs hard to explain.