AWS AIF-C01 Practice Questions with Explanations

Correct answer: B

Partial dependence plots (PDPs) are a key explainability tool that show how predicted values change as a feature varies, holding other features constant. This directly demonstrates the relationship between input features and model predictions, providing transparency to stakeholders about how the model makes forecasting decisions. PDPs help non-technical audiences understand model behavior without requiring knowledge of implementation details.

Why the other options are wrong:

• A. Model training code is implementation detail, not explanation of model behavior to stakeholders.
• C. Sample training data does not explain how the model makes predictions or what it learned.
• D. Convergence tables show training progress metrics but do not explain model explainability or feature relationships.

Correct answer: C

A summarization chatbot using LLMs is the appropriate solution for extracting key points from legal documents. LLMs excel at understanding document content and generating concise summaries of important information. This approach leverages the natural language understanding capabilities of LLMs to identify and extract relevant key points from complex legal texts.

Why the other options are wrong:

• A. Named entity recognition focuses on identifying specific entity types (names, organizations) rather than extracting broader key points from documents.
• B. A recommendation engine suggests items to users based on preferences, not extracting information from documents.
• D. Translation systems convert text between languages but do not extract or summarize key points.

Correct answer: A

Decision trees are highly interpretable models that naturally document their decision- making mechanism through visual tree structures. Each node represents a feature threshold, each branch represents a decision path, and leaves represent classification outputs. This transparent structure allows stakeholders to understand exactly how the model arrives at its classification for the 20 gene categories, making it ideal for applications requiring explainability.

Why the other options are wrong:

• B. Linear regression is for regression tasks, not classification, and while interpretable, it doesn't naturally handle 20 categorical outputs as effectively.
• C. Logistic regression is less transparent than decision trees and typically used for binary classification rather than 20-category classification.
• D. Neural networks are black-box models that do not naturally document inner mechanisms without additional explainability techniques.

Correct answer: B

Accuracy is the standard metric for classification tasks, measuring the proportion of correct predictions out of total predictions. For an image classification model predicting plant diseases, accuracy directly answers the question of how many images were classified correctly, making it the appropriate evaluation metric.

Why the other options are wrong:

• A. R-squared is used for regression tasks to measure variance explained, not for classification accuracy.
• C. Root mean squared error (RMSE) measures the magnitude of prediction errors in regression problems, not classification correctness.
• D. Learning rate is a hyperparameter controlling training step size, not an evaluation metric for model performance.

Correct answer: A

Adjusting the prompt is the most direct and effective way to control LLM output characteristics including length and language. By explicitly specifying in the prompt that responses should be short and written in a specific language, the model will align its outputs with these requirements without requiring model changes or parameter adjustments.

Why the other options are wrong:

• B. Changing model size may affect performance but does not directly control output length or language specificity.
• C. Increasing temperature makes outputs more random and diverse, not shorter or more aligned with specific language requirements.
• D. Increasing Top K value affects output diversity and randomness, not output length or language specification.

Correct answer: C

Asynchronous inference is designed for scenarios with large input sizes (up to 1 GB) and long processing times (up to 1 hour). It queues requests and processes them without requiring an immediate response, enabling the system to handle resource-intensive workloads while still providing results as near real-time as possible given the constraints. This is more cost-effective than real-time inference for these specifications.

Why the other options are wrong:

• A. Real-time inference requires low latency and is not designed for 1-hour processing times or 1 GB input sizes.
• B. Serverless inference has constraints on payload size and processing duration incompatible with these requirements.
• D. Batch transform is for offline processing of large datasets, not for near real-time latency requirements.

Correct answer: B

Transfer learning is the strategy of adapting pre-trained models to new, related tasks without building models from scratch. This approach leverages knowledge learned on large datasets and applies it to domain-specific tasks, reducing training time and data requirements while maintaining performance. It is the standard technique for adapting pre- trained models to new applications.

Why the other options are wrong:

• A. Increasing epochs affects training duration but is not a strategy for adapting pre- trained models to new tasks.
• C. Decreasing epochs would reduce training time but does not address adapting pre- trained models to new tasks.
• D. Unsupervised learning does not specifically address adapting pre-trained models to new related tasks.

Correct answer: A

Amazon SageMaker Ground Truth Plus provides human-in-the-loop validation where human annotators review and validate model outputs or annotations. This directly addresses the requirement to minimize the risk of incorrect annotations by ensuring human experts verify annotations for protective eyewear images, maintaining high accuracy through manual quality control.

Why the other options are wrong:

• B. Data augmentation using a knowledge base increases data volume but does not minimize annotation errors or ensure accuracy.
• C. Amazon Rekognition performs image recognition but does not validate annotations or provide human review for accuracy assurance.
• D. Amazon QuickSight Q summarizes data for business intelligence but does not validate or ensure annotation accuracy.

Correct answer: A

When Amazon Bedrock attempts to access S3 data encrypted with SSE-S3, the IAM role assumed by Bedrock must have the appropriate permissions to decrypt and read the objects. SSE-S3 encryption is managed by AWS, and the role needs s3:GetObject and related permissions to access the encrypted data. Without proper IAM permissions on the role, the access will fail regardless of the encryption key itself.

Why the other options are wrong:

• B. Public access to S3 buckets would expose sensitive data and violates security best practices; it is not the solution to permission issues.
• C. Prompt engineering cannot solve access control failures; the underlying IAM permissions must be correct.
• D. The presence or absence of sensitive information in the data does not determine whether the FM can access it; permissions must still be granted.

Correct answer: A

To achieve the lowest latency for inference on edge devices, deploying optimized small language models (SLMs) directly on the edge devices eliminates network round-trip time to centralized APIs. SLMs are designed for efficiency and can run locally with minimal computational overhead, whereas large models and centralized APIs introduce network latency and dependency on external services.

Why the other options are wrong:

• B. Large language models require significantly more computational resources and memory, making them impractical for edge deployment and increasing latency.
• C. Asynchronous communication with centralized APIs introduces network latency, defeating the goal of lowest-latency inference.
• D. Centralized LLM APIs are even slower than centralized SLM APIs and introduce both network and processing latency.

Correct answer: A

Amazon SageMaker Feature Store is designed specifically to store, manage, and share machine learning features across teams and projects. It provides a centralized repository for feature definitions, versions, and lineage, enabling multiple teams to discover, reuse, and collaborate on feature development without duplication of effort.

Why the other options are wrong:

• B. Data Wrangler is a tool for data preparation and transformation, not for managing and sharing variables across teams.
• C. Clarify is used for model explainability and bias detection, not for sharing variables.
• D. Model Cards document model metadata and performance but do not manage or share development variables across teams.

Correct answer: A

Amazon Q Developer is an AI assistant designed to enhance developer productivity by generating code snippets, providing reference tracking, and identifying open source license compliance issues. These capabilities directly support software development workflows and code quality management across teams.

Why the other options are wrong:

• B. Running applications without provisioning servers is AWS Lambda's purpose, not Amazon Q Developer's function.
• C. While Amazon Q has some conversational abilities, voice commands for coding and natural language search are not its primary capabilities.
• D. Audio-to-text conversion is handled by Amazon Transcribe, not Amazon Q Developer.

Correct answer: A

AWS PrivateLink enables private connectivity between VPCs and AWS services without requiring internet access, NAT gateways, or public IP addresses. This allows a VPC with no internet traffic to securely access Amazon Bedrock endpoints while maintaining regulatory compliance and network isolation.

Why the other options are wrong:

• B. Amazon Macie is a data security and privacy service for detecting sensitive data, not for enabling private connectivity.
• C. CloudFront is a content delivery network that requires internet connectivity; it does not provide private service access.
• D. An internet gateway explicitly provides internet access, which violates the requirement that the VPC has no internet traffic.

Correct answer: C

Calculating the probability of choosing a green marble is a deterministic mathematical problem with a simple formula: favorable outcomes divided by total outcomes. Using straightforward code logic requires no model training, data collection, or operational overhead, making it the most efficient solution. This is basic arithmetic, not a machine learning problem.

Why the other options are wrong:

• A. Supervised learning requires labeled training data and model training, creating unnecessary operational complexity for a simple mathematical calculation.
• B. Reinforcement learning is designed for sequential decision-making and optimization, not for computing a static probability from a given set.
• D. Unsupervised learning for probability density estimation requires data and model training, which is inappropriate for a deterministic calculation.

Correct answer: C

Average response time directly measures how quickly an AI model returns results after receiving input, which is the primary indicator of runtime efficiency. It captures the actual operational performance experienced by users and systems in production environments.

Why the other options are wrong:

• A. Customer satisfaction score measures user perception and experience quality, not the technical runtime efficiency of model operations.
• B. Training time per epoch measures the model development phase efficiency, not the runtime efficiency of deployed, operating models.
• D. The number of training instances relates to training infrastructure scale, not the efficiency of operating a deployed model.

Correct answer: B

Amazon Transcribe converts audio recordings into text transcripts, which is the essential first step for analyzing and extracting information from customer call audio. Once transcribed, the text can be further processed with other services for sentiment analysis, entity extraction, or classification as needed.

Why the other options are wrong:

• A. Amazon Lex is for building conversational chatbots, not for analyzing existing call recordings.
• C. SageMaker Model Monitor is used for monitoring model performance in production, not for extracting information from audio data.
• D. Amazon Comprehend can classify text, but it cannot directly extract information from audio; transcription must occur first.

Correct answer: B

The company has unlabeled customer data and wants to classify customers into tiers without predefined categories. This is a classic unsupervised learning scenario where clustering algorithms can automatically group customers based on patterns in the data. Supervised learning requires labeled training data, which the company doesn't have. Reinforcement learning and RLHF are designed for sequential decision-making and model alignment, not customer segmentation.

Why the other options are wrong:

• A. Supervised learning requires labeled data with known customer tier classifications, which the company does not have.
• C. Reinforcement learning is for training agents through trial and error with rewards, not for customer classification.
• D. RLHF is used for aligning model outputs with human preferences, not for customer tier clustering.

Correct answer: A

A multi-modal embedding model can process both text and images and convert them into vector representations that can be compared for semantic similarity in a search application. This allows the search engine to understand queries containing both modalities. Text embedding models only handle text, image generation models create new images rather than search, and multi-modal generation models are designed to generate content rather than perform semantic search and retrieval.

Why the other options are wrong:

• B. Text embedding models cannot process image data, which is required for the application's hybrid text-image queries.
• C. Multi-modal generation models are designed to generate new content, not to power semantic search functionality.
• D. Image generation models create new images rather than enabling search across existing text and image content.

Correct answer: A

Amazon Bedrock's fine-tuning capability requires labeled training data formatted with a prompt field and a completion field. This structured format allows the model to learn the relationship between inputs and desired outputs. A CSV text file format is not the proper method for Bedrock fine-tuning. Provisioned Throughput affects inference performance, not fine-tuning capability. Training on general sources like journals and textbooks is pre- training, not the fine-tuning of an existing foundation model with company-specific data.

Why the other options are wrong:

• B. While CSV format can store data, Bedrock fine-tuning specifically requires the prompt/completion field structure, not plain CSV lines.
• C. Provisioned Throughput is for inference performance optimization, not for enabling or improving fine-tuning capabilities.
• D. Training on general knowledge sources is pre-training; fine-tuning uses company- specific labeled data to adapt an existing model.

Correct answer: C

Anomaly detection systems are designed to identify unusual patterns and outliers in data, which directly applies to identifying suspicious IP addresses that deviate from normal network behavior. Speech recognition is unrelated to network security. Named entity recognition extracts named entities from text, not network behavior patterns. Fraud forecasting predicts future fraud events rather than detecting anomalous IP sources in real-time.

Why the other options are wrong:

• A. Speech recognition systems process audio and are completely unrelated to IP address threat detection.
• B. NLP named entity recognition extracts people, places, and organizations from text, not network threat indicators.
• D. Fraud forecasting predicts potential fraud through historical trends, whereas the requirement is to detect suspicious sources now.

Correct answer: C

Amazon OpenSearch Service supports vector indexing and nearest neighbor search, which are the core capabilities needed to build vector database applications. These features enable semantic search and similarity matching on embeddings. S3 integration is for object storage rather than vector operations. Geospatial indexing handles location- based queries. Real-time streaming analysis is unrelated to vector database functionality.

Why the other options are wrong:

• A. Integration with S3 provides object storage capabilities, not vector database functionality.
• B. Geospatial indexing supports location-based queries, not vector similarity search operations.
• D. Real-time streaming analysis is a separate capability distinct from vector database and nearest neighbor search functionality.

Correct answer: B

Creating photorealistic images from text descriptions is a direct use case for generative AI models like image generation models. These models generate new, original content based on textual prompts. Network security and intrusion detection use classification and anomaly detection (non-generative). Database performance optimization is a systems administration task. Stock market forecasting is predictive analytics, not generative AI.

Why the other options are wrong:

• A. Intrusion detection uses classification and anomaly detection, not generative AI model capabilities.
• C. Database optimization is a systems engineering task unrelated to generative AI content creation.
• D. Financial forecasting is predictive analytics that uses time-series models, not generative AI for content creation.

Correct answer: B

The context window determines the maximum amount of input and output tokens that can fit in a single prompt, directly addressing how much information can be included. Temperature controls output randomness/creativity. Batch size affects how many requests are processed together, not prompt capacity. Model size relates to overall model complexity but not specifically to prompt length limits.

Why the other options are wrong:

• A. Temperature controls the randomness of responses, not the amount of information that can fit in a prompt.
• C. Batch size determines how many inference requests are processed together, not the capacity of individual prompts.
• D. Model size refers to parameter count and computational requirements, not the prompt information capacity limit.

Correct answer: C

Experimenting and refining prompts is the standard approach to guide a foundation model toward producing outputs that match desired tone and style. Prompts can include tone specifications and examples that shape the model's behavior. Token limits restrict length, not tone. Batch inferencing is for processing efficiency, not response quality. High temperature increases randomness, which contradicts the need for consistent company tone adherence.

Why the other options are wrong:

• A. Limiting token output constrains response length but does not control tone or style of the content produced.
• B. Batch inferencing is an optimization technique for processing multiple requests efficiently, not for controlling response tone.
• D. Higher temperature increases randomness and unpredictability, making it harder to maintain consistent company tone.

Correct answer: A

Few-shot learning with examples is a proven prompt engineering strategy that improves LLM performance on classification tasks. By providing labeled examples of positive and negative sentiment text passages followed by the new text to classify, the model learns the pattern and applies it accurately to new data. This approach is significantly more effective than providing no context or unrelated examples.

Why the other options are wrong:

• B. Explaining how sentiment analysis and LLMs work doesn't teach the model the specific classification pattern needed for this task.
• C. Providing text without examples or context gives the model no guidance on how to structure the sentiment classification output.
• D. Examples of unrelated tasks like summarization or question answering introduce noise and confuse the model rather than guide it toward sentiment classification.

Correct answer: B

AWS CloudTrail records all API calls made to AWS services, including Amazon Bedrock invocations. It captures who made the request, when it was made, and whether it succeeded or failed, making it the ideal service for identifying unauthorized access attempts. CloudTrail logs are essential for security auditing and informing IAM policy decisions.

Why the other options are wrong:

• A. AWS Audit Manager focuses on compliance audits and control assessments rather than real-time unauthorized access detection.
• C. Amazon Fraud Detector is designed to identify fraudulent transactions and activities in business data, not API access attempts to AWS services.
• D. AWS Trusted Advisor provides best practice recommendations and checks for cost optimization, not unauthorized access attempt detection.

Correct answer: A

Amazon SageMaker Serverless Inference is specifically designed to deploy ML models without managing infrastructure. It automatically scales based on demand, requires no capacity planning, and eliminates the need for infrastructure management while serving predictions through a managed endpoint.

Why the other options are wrong:

• B. Amazon CloudFront is a content delivery network designed for distributing static and dynamic content, not for hosting and serving ML model predictions.
• C. Amazon API Gateway is used to create and manage APIs but does not provide the infrastructure to host and execute ML models for inference.
• D. AWS Batch is designed for batch processing of large computing workloads, not for real-time model serving and prediction requests.

Correct answer: B

AWS Artifact is the service that provides on-demand access to AWS compliance reports and certifications. It also allows customers to set up email notifications when new compliance reports from ISVs and third parties become available, meeting the requirement for receiving notifications about compliance report availability.

Why the other options are wrong:

• A. AWS Audit Manager helps manage internal audit and compliance processes but is not the primary service for ISV compliance report notifications.
• C. AWS Trusted Advisor provides recommendations for optimizing AWS usage and security, not third-party compliance report notifications.
• D. AWS Data Exchange is a marketplace for third-party data products, not specifically designed for compliance report notifications.

Correct answer: A

Creating a system prompt or prompt template that teaches the LLM to recognize and defend against common prompt injection attack patterns is an effective defense mechanism. By embedding instruction to detect adversarial prompts and malicious requests, the model can refuse dangerous inputs and avoid exposing sensitive information or performing undesirable actions.

Why the other options are wrong:

• B. Increasing temperature makes the model's outputs more random and unpredictable, which increases rather than decreases the risk of uncontrolled behavior.
• C. Restricting to SageMaker-listed LLMs does not inherently provide prompt injection protection; the defense must be built into the system design.
• D. Decreasing input tokens limits context but does not specifically address prompt engineering attacks or security vulnerabilities.

Correct answer: D

Building and training a generative AI model from scratch using proprietary data gives the company maximum ownership of security responsibilities. The company controls the entire pipeline including data collection, model architecture, training, validation, and deployment, meaning it bears responsibility for all security aspects throughout the process.

Why the other options are wrong:

• A. Using a third-party enterprise application minimizes the company's security responsibilities as the vendor maintains most infrastructure and model security.
• B. Building applications using an existing third-party FM means the vendor controls the base model security while the company handles application-level security.
• C. Fine-tuning a third-party FM reduces security responsibility compared to building from scratch, as the base model and its training data remain vendor-controlled.

Correct answer: A

Object detection is the computer vision technique specifically designed to identify and localize objects (in this case, animals) within images. It goes beyond simple classification by detecting what objects are present and where they are located, enabling automatic categorization of animals without manual effort.

Why the other options are wrong:

• B. Anomaly detection identifies unusual or outlier data points rather than categorizing normal objects like animals in photos.
• C. Named entity recognition is an NLP task for identifying named entities in text, not for identifying animals in images.
• D. Inpainting is a technique for reconstructing or filling in missing parts of images, not for identifying and categorizing objects present in photos.

Correct answer: A

Amazon Bedrock's On-Demand pricing model charges based on actual usage with no minimum commitments or long-term contracts. This provides maximum flexibility and is ideal for companies with limited budgets that prefer to scale without upfront financial obligations, paying only for what they consume.

Why the other options are wrong:

• B. Model customization focuses on fine-tuning capabilities rather than a pricing model option.
• C. Provisioned Throughput requires committing to reserved capacity and longer-term contracts, which contradicts the preference for flexibility without long-term commitment.
• D. Spot Instance is not an Amazon Bedrock pricing option; it is relevant to EC2 services and does not apply to foundation model usage.

Correct answer: B

Amazon SageMaker JumpStart is specifically designed to help teams quickly discover, evaluate, and deploy foundation models within their own VPC environment. It provides pre-trained models and one-click deployment capabilities, making it ideal for rapid FM consumption in a secure, isolated network. It offers VPC deployment options that keep models within the organization's network boundaries.

Why the other options are wrong:

• A. Amazon Personalize is for building recommendation engines, not for deploying foundation models.
• C. PartyRock is a playground for experimenting with Amazon Bedrock models in a public interface, not for VPC deployment within a team's infrastructure.
• D. SageMaker endpoints are a deployment mechanism but don't specifically address the quick discovery and deployment of pre-built foundation models like JumpStart does.

Correct answer: A

Secure LLM usage on Amazon Bedrock combines two critical practices: designing clear and specific prompts reduces ambiguity and prevents unintended outputs, while configuring IAM roles and policies with least privilege access ensures only authorized users and applications can invoke models and access data. This defense-in-depth approach addresses both prompt-level security and infrastructure-level access control.

Why the other options are wrong:

• B. AWS Audit Manager is for compliance auditing, not for model evaluation or securing LLM usage.
• C. Amazon Bedrock's automatic model evaluation jobs assess model performance, not security posture.
• D. CloudWatch Logs are useful for monitoring but don't inherently make models more explainable or secure; they're observability tools, not security mechanisms.

Correct answer: A

Generative Pre-trained Transformers (GPT) are the optimal choice for converting natural language input into SQL queries. GPTs excel at understanding human language and generating structured outputs like SQL commands, making them ideal for non-technical employees to describe what data they need in plain English. The model can learn the company's database schema and translate user intent directly into functional queries.

Why the other options are wrong:

• B. Residual neural networks are primarily used for image classification and computer vision tasks, not natural language processing.
• C. Support vector machines are supervised learning classifiers for structured classification problems, not generative models for text-to-SQL translation.
• D. WaveNet is a generative model designed for audio synthesis and time-series data, not for natural language understanding or SQL generation.

Correct answer: B

Inference is the process where a trained model makes predictions on new, unseen data. When the deployed object detection model analyzes a new image to identify objects, it is performing inference-applying the learned patterns to generate outputs without any training or model updates occurring.

Why the other options are wrong:

• A. Training is the process of teaching the model using labeled data; this already occurred before production deployment.
• C. Model deployment is the act of putting the model into production; it is the prerequisite for inference, not the process of analyzing images.
• D. Bias correction is a data preprocessing or model adjustment technique, not the process of analyzing images with a trained model.

Correct answer: A

Data augmentation for imbalanced classes directly addresses bias in training data by creating additional synthetic samples to balance underrepresented groups and attributes. When the input data shows bias toward certain attributes or professions, augmenting the dataset to include more diverse and balanced examples ensures the model learns to generate images without systematic bias toward specific characteristics.

Why the other options are wrong:

• B. Model monitoring for class distribution is a post-deployment monitoring technique, not a solution for fixing biased input data before training.
• C. Retrieval Augmented Generation (RAG) is used to supplement LLMs with external knowledge bases; it doesn't address image generation bias in the training data.
• D. Watermark detection is a technique for identifying image ownership or authenticity, completely unrelated to reducing bias in generative image models.

Correct answer: C

Amazon Bedrock knowledge bases enable organizations to supplement foundation models with private, proprietary data sources. This solution allows the Titan FM to access and augment its responses with relevant company-specific information through a managed retrieval mechanism, without fine-tuning the model itself.

Why the other options are wrong:

• A. Using a different FM doesn't solve the requirement to augment the current Titan model with private data.
• B. Temperature is a hyperparameter controlling response randomness; lowering it doesn't provide access to private data sources.
• D. Invocation logging is an observability feature for tracking model usage, not a mechanism for supplementing the model with private data.

Correct answer: B

Amazon SageMaker Clarify is specifically designed to generate interpretable metrics, reports, and examples that explain model behavior and decisions. For medical diagnostic applications requiring regulatory compliance and transparency, Clarify provides feature importance analysis, model explanations, and bias detection-directly addressing the need for explainable and transparent AI systems.

Why the other options are wrong:

• A. Amazon Inspector is a vulnerability assessment tool for infrastructure security, not for making models transparent or explainable.
• C. Amazon Macie is a data discovery and protection service focused on securing sensitive data; it doesn't provide model explainability.
• D. Gathering more data and using Rekognition for labeling improves model training but doesn't inherently make the model transparent or explainable for regulatory purposes.

Correct answer: B

Increasing the number of epochs allows the model to train longer and learn more from the training data, which improves accuracy up to a point of convergence. Decreasing batch size can add noise and instability, decreasing epochs reduces learning, and increasing temperature increases randomness in outputs-none of these improve accuracy for foundation model training.

Why the other options are wrong:

• A. Decreasing batch size introduces more noise and gradient instability, which typically worsens model performance.
• C. Decreasing epochs reduces the number of training iterations, limiting the model's opportunity to learn and reducing accuracy.
• D. Increasing temperature makes the model's outputs more random and less deterministic, which decreases accuracy.

Correct answer: B

Average call duration directly measures the efficiency of the chatbot in reducing the workload on call center employees. Shorter call durations indicate that the chatbot is handling more customer questions efficiently, reducing the actions employees need to take. Website engagement, CSR, and compliance are not directly tied to this specific objective.

Why the other options are wrong:

• A. Website engagement rate measures user interaction with web content, not the effectiveness of a call center chatbot in reducing employee workload.
• C. Corporate social responsibility is not related to measuring the operational efficiency of a customer service chatbot.
• D. Regulatory compliance is a governance concern, not a business metric for evaluating chatbot efficiency in reducing employee actions.

Correct answer: D

Amazon SageMaker Clarify is specifically designed to detect and mitigate bias in machine learning models during both the data preparation and model evaluation stages. It analyzes datasets and model predictions to identify potential bias against specific groups. While it provides explainability features, its primary purpose is bias detection and mitigation.

Why the other options are wrong:

• A. RAG workflow integration is handled by other services like Amazon Bedrock and knowledge base features, not SageMaker Clarify.
• B. Model monitoring in production is the role of Amazon SageMaker Model Monitor, not SageMaker Clarify.
• C. ML model documentation is handled by SageMaker Model Registry and other MLOps tools, not SageMaker Clarify's primary function.

Correct answer: C

The significant performance drop from training to production indicates overfitting, where the model learned training data patterns that don't generalize well. Increasing the volume of training data provides more diverse examples and reduces overfitting by giving the model better generalization capability. Reducing data worsens overfitting, hyperparameters alone don't address data inadequacy, and longer training time compounds overfitting.

Why the other options are wrong:

• A. Reducing training data volume exacerbates overfitting and makes generalization worse.
• B. Adding hyperparameters does not address the root cause of overfitting caused by insufficient training data diversity.
• D. Increasing training time on the same limited dataset reinforces overfitting rather than mitigating it.

Correct answer: D

Using Amazon Bedrock knowledge bases is the most cost-effective solution because it implements Retrieval Augmented Generation (RAG), which retrieves only relevant PDF content for each query rather than including all documents in every prompt. This minimizes token usage and associated costs. Options A and B waste tokens by including unnecessary context, while fine-tuning incurs significant training costs for a use case better served by RAG.

Why the other options are wrong:

• A. Adding one PDF as context wastes tokens on irrelevant information and requires manual selection.
• B. Adding all PDFs to every prompt maximizes token consumption and dramatically increases costs, especially with large document sets.
• C. Fine-tuning a model incurs substantial training costs and is unnecessary when RAG can efficiently retrieve relevant context on-demand.

Correct answer: D

Benchmark datasets are pre-curated, standardized datasets specifically designed to evaluate model performance and bias systematically with minimal effort. They come with established evaluation frameworks and metrics, requiring the least administrative overhead. User-generated content, moderation logs, and guidelines would all require significant manual curation, annotation, and setup to be suitable for bias evaluation.

Why the other options are wrong:

• A. User-generated content requires extensive curation, labeling, and validation before it can reliably evaluate bias.
• B. Moderation logs require manual review and annotation to identify bias patterns, demanding significant administrative effort.
• C. Content guidelines are normative documents; they define standards but don't provide evaluation data to assess bias in model outputs.

Correct answer: C

Effective prompt engineering with clear instructions and context directly guides the pre- trained model to generate content aligned with the company's brand voice and messaging. This approach requires no model modifications or retraining. Optimizing architecture and adding layers require model retraining, and pre-training a new model is unnecessary when a pre-trained model is already available.

Why the other options are wrong:

• A. Optimizing architecture and hyperparameters requires retraining, which is unnecessary when the goal is to align a pre-trained model's outputs.
• B. Adding layers increases model complexity and requires full retraining without addressing the content alignment requirement.
• D. Pre-training a new model is expensive and unnecessary when the requirement can be met by prompting an existing pre-trained model.

Correct answer: B

Amazon Bedrock currently does not support custom models directly within the service. To use a custom model alongside Bedrock, the company must deploy it in an Amazon SageMaker endpoint for real-time inference. This allows the custom model to run independently while the company continues using Bedrock for base models.

Why the other options are wrong:

• A. Provisioned Throughput is used for Bedrock base models, not for integrating custom models into Bedrock workflows.
• C. Registering with the SageMaker Model Registry is a tracking mechanism but does not enable use through Bedrock.
• D. Amazon Bedrock does not have a mechanism to register or grant access to custom- trained models within the service itself.

Correct answer: B

To identify a model that generates responses in a style employees prefer, the company must evaluate models using a human workforce and custom prompt datasets. This approach allows actual employees to assess outputs against their specific stylistic preferences and use cases, which is the only way to determine subjective quality like preferred writing style.

Why the other options are wrong:

• A. Built-in prompt datasets are generic and won't reflect the company's specific style preferences.
• C. Public leaderboards measure benchmarks like accuracy or speed, not stylistic preferences or internal use-case fit.
• D. InvocationLatency is a performance metric and does not evaluate response quality or style.

Correct answer: C

The scenario describes students copying content from generative AI to write essays without attribution, which is plagiarism-presenting someone else's work (or in this case, AI-generated work) as one's own. This directly represents the plagiarism challenge in responsible AI.

Why the other options are wrong:

• A. Toxicity refers to harmful or offensive language in AI outputs, not relevant to content copying.
• B. Hallucinations are false or fabricated information generated by AI models, not the issue here.
• D. Privacy concerns data handling and personal information protection, not unauthorized content reuse.

Correct answer: D

The EC2 Trn series instances are purpose-built for deep learning training with optimized power efficiency and lower energy consumption compared to other instance types. They deliver better environmental performance per unit of compute, making them the least environmentally impactful choice for training LLMs.

Why the other options are wrong:

• A. C series instances are compute-optimized for general purposes, not specialized for training and less efficient than Trn.
• B. G series instances are GPU-based for graphics and ML inference, less efficient for training workloads than Trn.
• C. P series instances are older GPU-based instances with higher power consumption than the newer Trn series.