143 free AWS AIF-C01 practice questions from the CertStash pack of 571, each with the correct answer and a full explanation. The rule is the same for every exam: ten free on the product page, 143 free here, all 571 in the pack, and each step includes the one before it.
Get all 571 questions (US$39) · Download the first 10 as a PDF · Read the study guide
Question 1
A company makes forecasts each quarter to decide how to optimize operations to meet expected demand. The company uses ML models to make these forecasts.
An AI practitioner is writing a report about the trained ML models to provide transparency and explainability to company stakeholders.
What should the AI practitioner include in the report to meet the transparency and explainability requirements?
Show answer and explanation
Correct answer: B. Partial dependence plots (PDPs)
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.
Question 2
A law firm wants to build an AI application by using large language models (LLMs). The application will read legal documents and extract key points from the documents.
Which solution meets these requirements?
Show answer and explanation
Correct answer: C. Develop a summarization chatbot.
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.
Question 3
A company wants to classify human genes into 20 categories based on gene characteristics. The company needs an ML algorithm to document how the inner mechanism of the model affects the output.
Which ML algorithm meets these requirements?
Show answer and explanation
Correct answer: A. Decision trees
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.
Question 4
A company has built an image classification model to predict plant diseases from photos of plant leaves. The company wants to evaluate how many images the model classified correctly.
Which evaluation metric should the company use to measure the model's performance?
Show answer and explanation
Correct answer: B. Accuracy
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.
Question 5
A company is using a pre-trained large language model (LLM) to build a chatbot for product recommendations. The company needs the LLM outputs to be short and written in a specific language.
Which solution will align the LLM response quality with the company's expectations?
Show answer and explanation
Correct answer: A. Adjust the prompt.
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.
Question 6
A company uses Amazon SageMaker for its ML pipeline in a production environment. The company has large input data sizes up to 1 GB and processing times up to 1 hour. The company needs near real-time latency.
Which SageMaker inference option meets these requirements?
Show answer and explanation
Correct answer: C. Asynchronous inference
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.
Question 7
A company is using domain-specific models. The company wants to avoid creating new models from the beginning. The company instead wants to adapt pre-trained models to create models for new, related tasks.
Which ML strategy meets these requirements?
Show answer and explanation
Correct answer: B. Use transfer learning.
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.
Question 8
A company is building a solution to generate images for protective eyewear. The solution must have high accuracy and must minimize the risk of incorrect annotations.
Which solution will meet these requirements?
Show answer and explanation
Correct answer: A. Human-in-the-loop validation by using Amazon SageMaker Ground Truth Plus
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.
Question 9
A company wants to create a chatbot by using a foundation model (FM) on Amazon Bedrock. The FM needs to access encrypted data that is stored in an Amazon S3 bucket. The data is encrypted with Amazon S3 managed keys (SSE-S3).
The FM encounters a failure when attempting to access the S3 bucket data.
Which solution will meet these requirements?
Show answer and explanation
Correct answer: A. Ensure that the role that Amazon Bedrock assumes has permission to decrypt data with the correct encryption key.
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.
Question 10
A company wants to use language models to create an application for inference on edge devices. The inference must have the lowest latency possible.
Which solution will meet these requirements?
Show answer and explanation
Correct answer: A. Deploy optimized small language models (SLMs) on edge devices.
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.
What people say about the full AIF-C01 pack 5.0 from 6
Passed. The questions in this dump are very close to what I saw on exam day. If you are sitting this one just get it. The exam is not cheap and this is.
Does the job. Passed first attempt.
Got my certification today. The material was clear, concise, and exactly what I needed to get over the finish line for the AIF-C01. The format was just like the real exam, and so was the difficulty level. Cheers!
Question 11
An AI practitioner who has minimal ML knowledge wants to predict employee attrition without writing code.
Which Amazon SageMaker feature meets this requirement?
Show answer and explanation
Correct answer: A. SageMaker Canvas
Amazon SageMaker Canvas is a visual, no-code tool that enables users without ML expertise to build and train predictive models through a user-friendly interface, making it ideal for predicting employee attrition without coding.
Why the other options are wrong
- B. SageMaker Clarify focuses on model explainability and bias detection, not model building.
- C. SageMaker Model Monitor tracks deployed model performance rather than enabling model creation.
- D. SageMaker Data Wrangler prepares data for ML but requires additional ML knowledge for prediction modeling.
Question 12
What does an F1 score measure in the context of foundation model (FM) performance?
Show answer and explanation
Correct answer: A. Model precision and recall
The F1 score is the harmonic mean of precision and recall, measuring the balance between a model's ability to correctly identify positive cases (precision) and its ability to find all positive cases (recall). It is a standard evaluation metric for classification performance.
Why the other options are wrong
- B. Model speed in generating responses is measured by latency or throughput metrics, not F1 score.
- C. Financial cost of operating a model is a cost metric, not a performance metric like F1 score.
- D. Energy efficiency is measured by power consumption or carbon footprint metrics, not by F1 score.
Question 13
An animation company wants to provide subtitles for its content.
Which AWS service meets this requirement?
Show answer and explanation
Correct answer: C. Amazon Transcribe
Amazon Transcribe converts audio and video content into text, which is the fundamental requirement for generating subtitles. While Translate could convert text to other languages and Polly could generate speech, Transcribe is the service that performs the core function of creating text from media content that can then be used as subtitles for animations.
Why the other options are wrong
- A. Amazon Comprehend performs text analysis and sentiment analysis, not audio-to-text conversion needed for subtitles.
- B. Amazon Polly converts text to speech, which is the opposite of what is needed; the company needs to extract text from video content.
- D. Amazon Translate converts text between languages but does not extract text from media files to create subtitles in the first place.
Question 14
A company wants to use a large language model (LLM) to develop a conversational agent. The company needs to prevent the LLM from being manipulated with common prompt engineering techniques to perform undesirable actions or expose sensitive information.
Which action will reduce these risks?
Show answer and explanation
Correct answer: A. Create a prompt template that teaches the LLM to detect attack patterns.
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.
Question 15
A financial services company must ensure that its generative AI-powered chatbot provides factual responses for regulatory compliance.
Which solution prevents the underlying foundation model (FM) from hallucinating?
Show answer and explanation
Correct answer: B. Configure Amazon Bedrock Guardrails to evaluate user inputs and model responses.
Amazon Bedrock Guardrails is the direct solution for preventing hallucinations and ensuring factual responses. Guardrails can evaluate both user inputs and model outputs to filter harmful content, enforce compliance policies, and prevent the model from generating false or unsupported information, which is critical for regulatory compliance in financial services.
Why the other options are wrong
- A. AWS Config manages infrastructure compliance and configuration metadata; it does not prevent model hallucinations.
- C. Amazon Fraud Detector identifies fraudulent activities in transactions, not hallucinations in chatbot responses.
- D. AWS Audit Manager prepares compliance reports but does not actively prevent a foundation model from generating false information.
Question 16
Which option is a benefit of using infrastructure as code (IaC) in machine learning operations (MLOps)?
Show answer and explanation
Correct answer: C. IaC streamlines the deployment of scalable and consistent ML workloads in cloud environments.
Infrastructure as Code enables automated, reproducible, and consistent deployment of ML infrastructure and workloads. This streamlines scaling, reduces manual errors, enables version control, and facilitates rapid deployment of ML pipelines in cloud environments, which are core MLOps benefits.
Why the other options are wrong
- A. IaC manages infrastructure provisioning; hyperparameter tuning is a separate concern handled by ML training processes.
- B. IaC provisions infrastructure based on specifications, not necessarily powerful instances; cost and performance are configurable based on requirements.
- D. IaC enables flexible provisioning based on requirements; it doesn't inherently minimize costs by deploying only low-cost instances (that's a separate cost optimization decision).
Question 17
Which statement presents an advantage of using Retrieval Augmented Generation (RAG) for natural language processing (NLP) tasks?
Show answer and explanation
Correct answer: A. RAG can use external knowledge sources to generate more accurate and informative responses.
RAG improves NLP by retrieving relevant information from external knowledge sources and incorporating that information into the generation process, resulting in more accurate, current, and contextually informed responses. This allows language models to access up-to-date information beyond their training data, reducing hallucinations and improving factual accuracy.
Why the other options are wrong
- B. RAG is designed to improve response quality through retrieval-augmentation, not to accelerate the training process itself.
- C. RAG is an NLP technique; speech recognition is a separate domain with different architectures and approaches.
- D. RAG is specifically for NLP tasks; data augmentation in computer vision uses different techniques like image transformations.
Question 18
A company wants to identify groups for its customers based on the customers’ demographics and buying patterns.
Which algorithm should the company use to meet this requirement?
Show answer and explanation
Correct answer: B. K-means
K-means is an unsupervised clustering algorithm specifically designed to identify groups (clusters) in data based on similarities in features like demographics and buying patterns, making it ideal for customer segmentation.
Why the other options are wrong
- A. K-nearest neighbors is a supervised classification/regression algorithm, not a clustering method for grouping discovery.
- C. Decision trees are primarily used for supervised classification and regression, not unsupervised customer grouping.
- D. Support vector machines are classifiers for supervised learning, not unsupervised clustering algorithms.
Question 19
An ecommerce company wants to improve search engine recommendations by customizing the results for each user of the company’s ecommerce platform.
Which AWS service meets these requirements?
Show answer and explanation
Correct answer: A. Amazon Personalize
Amazon Personalize is specifically designed to create personalized user experiences by analyzing user behavior and preferences to provide customized recommendations. For an ecommerce platform wanting to customize search results for each user, Personalize is the ideal service. The other services serve different purposes: Kendra for enterprise search, Rekognition for image analysis, and Transcribe for speech-to-text.
Why the other options are wrong
- B. Amazon Kendra is an enterprise search service that improves search relevance across documents, not for personalizing results per individual user.
- C. Amazon Rekognition analyzes images and videos to identify objects and scenes, not relevant for search recommendations.
- D. Amazon Transcribe converts speech to text, unrelated to ecommerce search personalization.
Question 20
A retail company wants to build an ML model to recommend products to customers. The company wants to build the model based on responsible practices.
Which practice should the company apply when collecting data to decrease model bias?
Show answer and explanation
Correct answer: C. Ensure that the data is balanced and collected from a diverse group.
Collecting balanced, diverse data is the foundational responsible practice for reducing bias in ML models. Diverse data ensures the model learns from varied perspectives and demographics, preventing it from learning skewed patterns. Limiting to matching demographics, past purchasers, or public datasets can all introduce selection bias or underrepresentation of populations.
Why the other options are wrong
- A. Matching only similar demographics to the customer base perpetuates bias rather than reducing it.
- B. Only collecting from past purchasers excludes potential customers and introduces selection bias.
- D. Public datasets may themselves contain historical biases and are not guaranteed to be balanced.
Question 21
A company wants to develop an educational game where users answer questions such as the following: "A jar contains six red, four green, and three yellow marbles.
What is the probability of choosing a green marble from the jar?"
Which solution meets these requirements with the LEAST operational overhead?
Show answer and explanation
Correct answer: C. Use code that will calculate probability by using simple rules and computations.
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.
Question 22
A company is using custom models in Amazon Bedrock for a generative AI application. The company wants to use a company managed encryption key to encrypt the model artifacts that the model customization jobs create.
Which AWS service meets these requirements?
Show answer and explanation
Correct answer: A. AWS Key Management Service (AWS KMS)
AWS Key Management Service (AWS KMS) is the AWS service for managing encryption keys. Amazon Bedrock supports KMS integration to encrypt model customization job artifacts using customer-managed keys, providing the required control over encryption key management.
Why the other options are wrong
- B. Amazon Inspector is a vulnerability assessment service, not an encryption key management solution.
- C. Amazon Macie is a data discovery and protection service focused on detecting sensitive data, not managing encryption keys.
- D. AWS Secrets Manager stores secrets and credentials, not designed for managing encryption keys for service artifacts.
Question 23
An airline company wants to build a conversational AI assistant to answer customer questions about flight schedules, booking, and payments.
The company wants to use large language models (LLMs) and a knowledge base to create a text-based chatbot interface.
Which solution will meet these requirements with the LEAST development effort?
Show answer and explanation
Correct answer: B. Develop a Retrieval Augmented Generation (RAG) agent by using Amazon Bedrock.
Retrieval Augmented Generation (RAG) using Amazon Bedrock is the ideal solution for this use case because it combines pre-built LLMs with a knowledge base without requiring model training or fine-tuning. RAG retrieves relevant information from the knowledge base and uses it to ground the model's responses, minimizing development effort while meeting all requirements for a conversational chatbot.
Why the other options are wrong
- A. SageMaker Autopilot is designed for automated machine learning on structured data, not for building conversational AI with LLMs.
- C. Amazon Q Developer is primarily a code generation and developer assistant tool, not the optimal solution for customer-facing conversational AI.
- D. Fine-tuning models on SageMaker Jumpstart requires significant development effort and is unnecessary when using RAG with pre-built models.
Question 24
A company is developing an ML model to make loan approvals. The company must implement a solution to detect bias in the model. The company must also be able to explain the model's predictions.
Which solution will meet these requirements?
Show answer and explanation
Correct answer: A. Amazon SageMaker Clarify
Amazon SageMaker Clarify is specifically designed to detect bias in ML models and explain model predictions. It provides tools for bias detection across protected attributes, feature importance analysis, and prediction explanations, making it the comprehensive solution for both bias detection and model interpretability requirements.
Why the other options are wrong
- B. Data Wrangler is for data preparation and transformation, not for bias detection or model explanation.
- C. Model Cards document model properties but do not actively detect bias or explain predictions.
- D. AI Service Cards provide documentation about AWS AI services, not bias detection or explanation capabilities.
Question 25
A company wants to build and deploy ML models on AWS without writing any code.
Which AWS service or feature meets these requirements?
Show answer and explanation
Correct answer: A. Amazon SageMaker Canvas
Amazon SageMaker Canvas is a no-code/low-code service that allows users to build and deploy ML models without writing code. It provides a visual interface for preparing data, training models, and making predictions, making it ideal for businesses wanting to leverage ML without requiring machine learning expertise or coding skills.
Why the other options are wrong
- B. Amazon Rekognition is a pre-built service for image analysis, not a no-code model building platform.
- C. AWS DeepRacer is a reinforcement learning service focused on autonomous racing, not general-purpose no-code ML model building.
- D. Amazon Comprehend is a pre-built NLP service, not a platform for building custom ML models without code.
Question 26
A financial company is developing a generative AI application for loan approval decisions. The company needs the application output to be responsible and fair.
Which solution meets these requirements?
Show answer and explanation
Correct answer: A. Review the training data to check for biases. Include data from all demographics in the training data.
Reviewing training data for biases and ensuring diverse demographic representation directly addresses fairness and responsibility in AI systems. This approach mitigates algorithmic bias at the source and produces more equitable loan approval decisions.
Why the other options are wrong
- B. Deep learning with many hidden layers increases model complexity but does not inherently ensure fairness or responsibility.
- C. Keeping decision-making processes secret contradicts the requirement for responsible AI, which demands transparency and explainability.
- D. Monitoring performance on a static test dataset alone does not address bias or fairness; datasets must be updated and diverse.
Question 27
A company wants to control employee access to publicly available foundation models (FMs).
Which solution meets these requirements?
Show answer and explanation
Correct answer: C. Configure Amazon SageMaker JumpStart to restrict discoverable FMs.
Amazon SageMaker JumpStart can be configured to control which foundation models employees can discover and access. This provides the necessary governance and access control mechanism to restrict employee access to publicly available FMs based on organizational policies.
Why the other options are wrong
- A. AWS Cost Explorer analyzes spending patterns but does not control access to foundation models.
- B. AWS Artifact provides security and compliance documents for reference but does not enforce access controls to FMs.
- D. Amazon OpenSearch Service is for search and analytics capabilities, not for controlling access to foundation models.
Question 28
A company is developing a mobile ML app that uses a phone's camera to diagnose and treat insect bites. The company wants to train an image classification model by using a diverse dataset of insect bite photos from different genders, ethnicities, and geographic locations around the world.
Which principle of responsible AI does the company demonstrate in this scenario?
Show answer and explanation
Correct answer: A. Fairness
The company demonstrates fairness by intentionally using a diverse dataset that includes different genders, ethnicities, and geographic locations. Fairness in AI ensures that models perform equitably across different demographic groups and populations, preventing bias and discrimination in predictions.
Why the other options are wrong
- B. Explainability refers to making model decisions understandable, not addressed in the dataset diversity scenario.
- C. Governance relates to organizational structures and policies for AI oversight, not dataset composition.
- D. Transparency involves disclosing how AI systems work, not the representativeness of training data.
Question 29
An AI practitioner wants to predict the classification of flowers based on petal length, petal width, sepal length, and sepal width.
Which algorithm meets these requirements?
Show answer and explanation
Correct answer: A. K-nearest neighbors (k-NN)
K-nearest neighbors (k-NN) is a classification algorithm that uses feature similarity to classify instances based on their nearest neighbors. It is well-suited for flower classification problems using continuous features like petal and sepal dimensions, as demonstrated by its use in classic datasets like Iris.
Why the other options are wrong
- B. K-means is an unsupervised clustering algorithm, not a classification algorithm for predefined flower categories.
- C. ARIMA is a time series forecasting algorithm, not applicable to static feature-based classification.
- D. Linear regression is used for continuous value prediction, not for categorical flower classification.
Question 30
A company has signed up for Amazon Bedrock access to build applications. The company wants to restrict employee access to specific models available on Amazon Bedrock.
Which solution meets these requirements?
Show answer and explanation
Correct answer: A. Use AWS Identity and Access Management (IAM) policies to restrict model access.
IAM policies are the standard AWS mechanism for controlling access to specific resources and services. To restrict employee access to specific Amazon Bedrock models, you configure IAM policies that define which models users or roles can access, providing fine-grained permission control at the service level.
Why the other options are wrong
- B. AWS STS generates temporary credentials but does not provide the mechanism to restrict access to specific models; it manages credential validity rather than resource permissions.
- C. IAM service roles define permissions for AWS services, not for restricting access to specific models on behalf of employees; this would require IAM policies for users or groups.
- D. Amazon Inspector is a vulnerability assessment tool for scanning infrastructure and applications, not for controlling access to Bedrock models.
Question 31
A company is using an Amazon Bedrock base model to summarize documents for an internal use case. The company trained a custom model to improve the summarization quality.
Which action must the company take to use the custom model through Amazon Bedrock?
Show answer and explanation
Correct answer: A. Purchase Provisioned Throughput for the custom model.
A custom model created in Amazon Bedrock through fine-tuning or continued pre-training needs Provisioned Throughput before it can be used for inference. Once Provisioned Throughput is purchased, the model is invoked through the Bedrock API like other models.
Why the other options are wrong
- B. Bedrock custom models are served by Bedrock, not by deploying them to a SageMaker endpoint.
- C. The SageMaker Model Registry tracks model versions and does not make a model usable in Bedrock.
- D. Model access grants apply to base models from providers, not to running a custom model.
Question 32
What is the purpose of chunking in Retrieval Augmented Generation (RAG)?
Show answer and explanation
Correct answer: C. To improve the contextual relevancy of results retrieved from the vector index
Chunking in RAG breaks large documents into smaller segments before embedding and indexing. This improves contextual relevancy because each chunk maintains semantic coherence and can be independently embedded and retrieved. Smaller, well-defined chunks ensure that retrieved results are topically focused and contextually relevant to user queries, rather than returning overly large or loosely related content.
Why the other options are wrong
- A. While chunking can help with storage, that is not its primary purpose in RAG; vector databases handle large volumes efficiently.
- B. Chunking does not avoid the need for embeddings; each chunk must still be converted to vector embeddings for retrieval.
- D. Storage cost reduction is not a primary objective of chunking in RAG architectures.
Question 33
A company is building a customer service chatbot. The company wants the chatbot to improve its responses by learning from past interactions and online resources.
Which AI learning strategy provides this self-improvement capability?
Show answer and explanation
Correct answer: B. Reinforcement learning with rewards for positive customer feedback
Reinforcement learning with rewards for positive customer feedback directly enables self-improvement. The model learns by receiving signals (rewards) when it performs well and penalties when it performs poorly, allowing it to continuously improve its responses based on actual customer interactions. This mirrors how the chatbot learns from experience.
Why the other options are wrong
- A. Supervised learning with manual curation is static; it requires human intervention to improve, not self-improvement.
- C. Unsupervised learning finds patterns but doesn't inherently provide feedback-driven improvement mechanisms.
- D. Supervised learning with an updated FAQ is still manual and doesn't enable autonomous self-improvement from interactions.
Question 34
Which scenario describes a potential risk and limitation of prompt engineering in the context of a generative AI model?
Show answer and explanation
Correct answer: B. Prompt engineering could expose the model to vulnerabilities such as prompt injection attacks.
Prompt injection attacks are a significant vulnerability where attackers craft malicious prompts to manipulate model behavior, bypass safety measures, or extract confidential information. Prompt engineering itself can inadvertently expose models to these injection-based attacks, representing a real security risk in generative AI applications.
Why the other options are wrong
- A. While true that prompt engineering doesn't guarantee consistency, this describes a limitation rather than a specific risk vulnerability.
- C. Data poisoning and model hijacking are pre-deployment risks unrelated to prompt engineering vulnerabilities.
- D. While accurate that prompts don't guarantee reliability, this describes a general limitation rather than a specific prompt engineering risk.
Question 35
A company is integrating AI into its employee recruitment and hiring solution. The company wants to mitigate bias risks and ensure responsible AI practices while prioritizing equitable hiring decisions.
Which core dimensions of responsible AI should the company consider? (Choose two.)
Show answer and explanation
Correct answer: A, E. Fairness; Transparency
Fairness and Transparency are core dimensions of responsible AI in hiring. Fairness ensures equitable treatment across demographics and reduces bias in recruitment decisions. Transparency ensures the company can explain how the AI model makes hiring recommendations, which is critical for regulatory compliance and building trust with candidates. These two dimensions directly address the company's stated goals of mitigating bias and ensuring responsible AI practices.
Why the other options are wrong
- B. Tolerance is not a core dimension of responsible AI frameworks; it is not a standard pillar in AI ethics.
- C. Flexibility is not a core dimension of responsible AI; it refers to system adaptability rather than ethical principles.
- D. Open source is a development approach, not a core dimension of responsible AI principles.
Question 36
A company is using a large language model (LLM) on Amazon Bedrock to build a chatbot. The chatbot processes customer support requests. To resolve a request, the customer and the chatbot must interact a few times.
Which solution gives the LLM the ability to use content from previous customer messages?
Show answer and explanation
Correct answer: B. Add messages to the model prompt.
To enable an LLM to reference previous customer messages in a multi-turn conversation, those messages must be included in the prompt sent to the model. Amazon Bedrock's LLMs are stateless and do not retain conversation history between API calls, so the application must explicitly add prior messages to each new prompt to maintain context.
Why the other options are wrong
- A. Model invocation logging collects messages for auditing purposes but does not provide them back to the model for context.
- C. Amazon Personalize is designed for recommendation systems, not for managing conversation history for LLMs.
- D. Provisioned Throughput controls capacity and pricing but has no effect on the model's ability to access previous messages.
Question 37
A company wants to build an interactive application for children that generates new stories based on classic stories. The company wants to use Amazon Bedrock and needs to ensure that the results and topics are appropriate for children.
Which AWS service or feature will meet these requirements?
Show answer and explanation
Correct answer: C. Guardrails for Amazon Bedrock
Guardrails for Amazon Bedrock is specifically designed to filter and control AI-generated content to ensure it meets safety requirements. This feature allows the company to set policies that block inappropriate topics and ensure results are suitable for children, directly meeting the requirement to keep content appropriate.
Why the other options are wrong
- A. Amazon Rekognition is for image and video analysis, not text generation or content filtering.
- B. Bedrock playgrounds are interactive testing environments but lack built-in safety controls for child-appropriate content.
- D. Agents for Amazon Bedrock enable autonomous task execution but don't provide content filtering or safety guardrails.
Question 38
A company wants to use a large language model (LLM) on Amazon Bedrock for sentiment analysis. The company needs the LLM to produce more consistent responses to the same input prompt.
Which adjustment to an inference parameter should the company make to meet these requirements?
Show answer and explanation
Correct answer: A. Decrease the temperature value.
Temperature controls the randomness of model outputs. Decreasing temperature makes the model more deterministic and produces more consistent responses to the same input, while increasing temperature adds more randomness and variability. For consistent responses, lowering temperature is the correct adjustment.
Why the other options are wrong
- B. Increasing temperature increases randomness, which produces less consistent responses.
- C. Output token length affects response length, not consistency of responses.
- D. Maximum generation length affects response length, not consistency of responses.
Question 39
Which AWS service or feature can help an AI development team quickly deploy and consume a foundation model (FM) within the team's VPC?
Show answer and explanation
Correct answer: B. Amazon SageMaker JumpStart
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.
Question 40
A hospital developed an AI system to provide personalized treatment recommendations for patients. The AI system must provide the rationale behind the recommendations and make the insights accessible to doctors and patients.
Which human-centered design principle does this scenario present?
Show answer and explanation
Correct answer: A. Explainability
Explainability is the principle that AI systems should provide clear, understandable reasoning for their decisions. This scenario explicitly requires the AI to provide rationale behind recommendations and make insights accessible to both doctors and patients, which is the core definition of explainability in human-centered AI design.
Why the other options are wrong
- B. Privacy and security concern data protection and confidentiality, not the transparency of AI reasoning.
- C. Fairness addresses bias and equal treatment across different groups, not the explanation of individual recommendations.
- D. Data governance involves managing data lifecycle and quality, not explaining AI system outputs to end users.
Question 41
A company is using the Generative AI Security Scoping Matrix to assess security responsibilities for its solutions. The company has identified four different solution scopes based on the matrix.
Which solution scope gives the company the MOST ownership of security responsibilities?
Show answer and explanation
Correct answer: D. Building and training a generative AI model from scratch by using specific data that a customer owns.
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.
Question 42
A large retail bank wants to develop an ML system to help the risk management team decide on loan allocations for different demographics.
What must the bank do to develop an unbiased ML model?
Show answer and explanation
Correct answer: D. Measure class imbalance on the training dataset. Adapt the training process accordingly.
Measuring class imbalance and adapting the training process accordingly is a fundamental technique for developing unbiased ML models. Class imbalance often leads to biased predictions favoring majority demographics. By identifying and correcting this imbalance through techniques like reweighting, resampling, or adjusted loss functions, the model becomes more equitable across demographic groups.
Why the other options are wrong
- A. Reducing dataset size generally increases bias rather than reducing it.
- B. Ensuring consistency with historical results would perpetuate existing historical biases rather than create an unbiased model.
- C. Creating separate models for each demographic group does not address bias; it fragments the solution and can reinforce demographic divisions.
Question 43
A company uses Amazon Bedrock for its generative AI application. The company wants to use Amazon Bedrock Guardrails to detect and filter harmful user inputs and model-generated outputs.
Which content categories can the guardrails filter? (Choose two.)
Show answer and explanation
Correct answer: A, C. Hate; Violence
Amazon Bedrock Guardrails can filter content categories including Hate and Violence. These are standard harmful content categories that guardrails are designed to detect and filter from both user inputs and model outputs to prevent harmful content generation.
Why the other options are wrong
- B. Politics is not a standard harmful content category that Bedrock Guardrails filter.
- D. Gambling is not a standard harmful content category that Bedrock Guardrails filter.
- E. Religion is not a standard harmful content category that Bedrock Guardrails filter.
Question 44
A company needs to log all requests made to its Amazon Bedrock API. The company must retain the logs securely for 5 years at the lowest possible cost.
Which combination of AWS service and storage class meets these requirements? (Choose two.)
Show answer and explanation
Correct answer: A, D. AWS CloudTrail; Amazon S3 Intelligent-Tiering
AWS CloudTrail captures all API calls made to AWS services including Amazon Bedrock, providing comprehensive audit logs required for compliance. Amazon S3 Intelligent-Tiering automatically optimizes storage costs by moving objects between access tiers based on usage patterns, making it ideal for 5-year retention at the lowest possible cost. CloudTrail integrates seamlessly with S3 Intelligent-Tiering to meet logging and cost requirements.
Why the other options are wrong
- B. CloudWatch is for metrics and monitoring, not for API request logging; it would not retain logs for 5 years cost-effectively.
- C. AWS Audit Manager is for compliance assessment and audit management, not for capturing API request logs.
- E. Amazon S3 Standard keeps all data in the frequent access tier, resulting in higher costs compared to Intelligent-Tiering for long-term retention.
Question 45
A company has built a solution by using generative AI. The solution uses large language models (LLMs) to translate training manuals from English into other languages. The company wants to evaluate the accuracy of the solution by examining the text generated for the manuals.
Which model evaluation strategy meets these requirements?
Show answer and explanation
Correct answer: A. Bilingual Evaluation Understudy (BLEU)
BLEU (Bilingual Evaluation Understudy) is the standard metric for evaluating machine translation quality by comparing generated translations against reference translations. It measures n-gram overlap and is specifically designed for translation tasks. RMSE is for regression problems, ROUGE is designed for summarization evaluation, and F1 score is for classification tasks; none are appropriate for evaluating translation accuracy.
Why the other options are wrong
- B. RMSE measures prediction error in regression tasks, not translation quality.
- C. ROUGE evaluates summarization quality, not translation accuracy.
- D. F1 score is for classification tasks, not suitable for translation evaluation.
Question 46
An AI practitioner is using a large language model (LLM) to create content for marketing campaigns. The generated content sounds plausible and factual but is incorrect.
Which problem is the LLM having?
Show answer and explanation
Correct answer: B. Hallucination
Hallucination is the phenomenon where LLMs generate plausible-sounding but factually incorrect content. This is exactly what the question describes: content that sounds credible but is incorrect. Data leakage involves unintended exposure of training data, overfitting occurs when models perform poorly on unseen data, and underfitting occurs when models fail to learn patterns; none of these describe generating confident but false information.
Why the other options are wrong
- A. Data leakage refers to unintended exposure of training data in outputs, not generating false facts.
- C. Overfitting causes poor generalization to new data, not confident generation of false information.
- D. Underfitting causes poor performance due to insufficient learning, not confident false generation.
Question 47
A documentary filmmaker wants to reach more viewers. The filmmaker wants to automatically add subtitles and voice-overs in multiple languages to their films.
Which combination of steps will meet these requirements? (Choose two.)
Show answer and explanation
Correct answer: A, C. Use Amazon Transcribe and Amazon Translate to generate subtitles in other languages; Use Amazon Polly to generate voice-overs in other languages.
Both options A and C are required. Amazon Transcribe converts the film's audio to text, Amazon Translate then translates those subtitles into multiple languages (option A). Amazon Polly generates voice-overs by synthesizing speech in multiple languages from translated text (option C). Textract is for document text extraction not applicable to films, and Amazon Translate only handles text, not voice generation.
Why the other options are wrong
- B. Amazon Textract extracts text from documents/images and is not applicable to video film content.
- D. Amazon Translate only translates text; it cannot generate voice-overs or synthesize speech.
- E. Amazon Textract is for extracting text from documents and cannot generate voice-overs.
Question 48
Which functionality does Amazon SageMaker Clarify provide?
Show answer and explanation
Correct answer: D. Identifies potential bias during data preparation
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.
Question 49
Which option is a characteristic of AI governance frameworks for building trust and deploying human-centered AI technologies?
Show answer and explanation
Correct answer: D. Developing policies and guidelines for data, transparency, responsible AI, and compliance
A comprehensive AI governance framework fundamentally requires developing clear policies and guidelines covering data management, transparency, responsible AI practices, and regulatory compliance. This ensures human-centered AI development and builds trust through accountability and ethical standards.
Why the other options are wrong
- A. While business expansion is valuable, it's a business outcome rather than a characteristic of governance frameworks themselves.
- B. Alignment with business standards is important but is a narrower concern than the broader governance framework requirements.
- C. Overcoming challenges and driving transformation are outcomes or consequences of governance, not defining characteristics of the framework.
Question 50
A company has developed an ML model for image classification. The company wants to deploy the model to production so that a web application can use the model.
The company needs to implement a solution to host the model and serve predictions without managing any of the underlying infrastructure.
Which solution will meet these requirements?
Show answer and explanation
Correct answer: A. Use Amazon SageMaker Serverless Inference to deploy the model.
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.
Question 51
Which prompting technique can protect against prompt injection attacks?
Show answer and explanation
Correct answer: A. Adversarial prompting
Adversarial prompting is a technique that tests and protects against prompt injection attacks by deliberately attempting malicious prompts and refining defenses based on the results. By understanding how injection attacks work, systems can be hardened against them. This proactive testing approach is how organizations identify and mitigate prompt injection vulnerabilities.
Why the other options are wrong
- B. Zero-shot prompting is a technique for generating responses without examples, not a defense mechanism against injection attacks.
- C. Least-to-most prompting breaks problems into simpler steps but does not specifically protect against injection attacks.
- D. Chain-of-thought prompting improves reasoning but is not designed to defend against prompt injection attacks.
Question 52
A student at a university is copying content from generative AI to write essays.
Which challenge of responsible generative AI does this scenario represent?
Show answer and explanation
Correct answer: C. Plagiarism
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.
Question 53
A research company implemented a chatbot by using a foundation model (FM) from Amazon Bedrock. The chatbot searches for answers to questions from a large database of research papers.
After multiple prompt engineering attempts, the company notices that the FM is performing poorly because of the complex scientific terms in the research papers.
How can the company improve the performance of the chatbot?
Show answer and explanation
Correct answer: B. Use domain adaptation fine-tuning to adapt the FM to complex scientific terms.
Domain adaptation fine-tuning is the appropriate solution when a foundation model struggles with specialized vocabulary and concepts. By fine-tuning the FM on research papers with complex scientific terms, the model learns domain-specific language patterns and improves its ability to understand and respond to scientific queries.
Why the other options are wrong
- A. Few-shot prompting provides in-context examples but doesn't teach the model the complex scientific terminology itself.
- C. Inference parameters control output generation style and diversity, not the model's understanding of domain-specific concepts.
- D. Removing complex scientific terms defeats the purpose; the model needs to learn these terms to handle real research papers effectively.
Question 54
A grocery store wants to create a chatbot to help customers find products in the store. The chatbot must check the inventory in real time and provide the product location in the store.
Which prompt engineering technique should the store use to build the chatbot?
Show answer and explanation
Correct answer: D. Reasoning and acting (ReAct) prompting
ReAct (Reasoning and Acting) prompting is the correct technique for this scenario because it enables the chatbot to reason about user queries and then take actions (like querying the inventory database and retrieving product locations). ReAct prompting allows the model to interleave reasoning steps with action steps, making it ideal for tasks that require real-time information retrieval and decision-making based on external data sources.
Why the other options are wrong
- A. Zero-shot prompting provides no examples and cannot reliably trigger the complex behavior needed to query inventory systems and provide accurate location data.
- B. Few-shot prompting demonstrates how to answer questions but does not inherently enable the model to take actions or access external systems like inventory databases.
- C. Least-to-most prompting breaks down complex problems into simpler subproblems but is not specifically designed for scenarios requiring real-time data retrieval and actionable responses.
Question 55
A company is building a generative AI tool. The company will use internal documents to customize a foundation model (FM).
Which approach will meet this requirement?
Show answer and explanation
Correct answer: B. Continued pre-training
Continued pre-training is the approach that customizes a foundation model using internal documents by further training the model on domain-specific data. This allows the model to learn from and adapt to the company's specific documents and use cases. Classification and regression are supervised learning tasks for making predictions, while distillation is about compressing models, none of which customize a foundation model with internal documents.
Why the other options are wrong
- A. Classification is a supervised learning task for categorizing data, not for customizing foundation models with documents.
- C. Distillation is a technique for compressing large models into smaller ones, not for customizing with internal documents.
- D. Regression is a supervised learning task for predicting continuous values, not for adapting foundation models to internal data.
Question 56
A company wants to identify harmful language in the comments section of social media posts by using an ML model. The company will not use labeled data to train the model.
Which strategy should the company use to identify harmful language?
Show answer and explanation
Correct answer: B. Use Amazon Comprehend toxicity detection.
Amazon Comprehend toxicity detection is a pre-trained, managed service that identifies harmful language without requiring labeled training data. It operates as a ready-to-use API that analyzes text for toxic content. Rekognition handles image moderation, SageMaker requires training data to build custom models, and Polly is for text-to-speech synthesis, not comment monitoring.
Why the other options are wrong
- A. Amazon Rekognition is designed for image and video moderation, not text-based comment analysis.
- C. SageMaker built-in algorithms require labeled training data, which the company explicitly will not use.
- D. Amazon Polly converts text to speech; it is not designed for content moderation or harmful language detection.
Question 57
A company has petabytes of unlabeled customer data to use for an advertisement campaign. The company wants to classify its customers into tiers to advertise and promote the company's products.
Which methodology should the company use to meet these requirements?
Show answer and explanation
Correct answer: B. Unsupervised learning
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.
Question 58
An AI practitioner wants to generate more diverse and more creative outputs from a large language model (LLM).
How should the AI practitioner adjust the inference parameter?
Show answer and explanation
Correct answer: A. Increase the temperature value.
Increasing the temperature value makes the model's output more random and diverse by increasing the probability of selecting less likely tokens. Higher temperature (e.g., 0.8-1.0) promotes creativity and diversity in outputs, while lower temperature (e.g., 0.1-0.3) makes outputs more deterministic and focused.
Why the other options are wrong
- B. Decreasing Top K restricts the model to fewer candidate tokens, reducing diversity rather than increasing it.
- C. Response length controls how long the output is but does not directly affect the diversity or creativity of the generated content.
- D. Prompt length does not impact the randomness or creativity of the model's inference behavior.
Question 59
A company designed an AI-powered agent to answer customer inquiries based on product manuals.
Which strategy can improve customer confidence levels in the AI-powered agent's responses?
Show answer and explanation
Correct answer: B. Including referenced product manual links in the response
Including referenced product manual links in responses improves customer confidence by providing verifiable sources that customers can check independently. This transparency and traceability demonstrate that the AI's recommendations are grounded in authoritative documentation, building trust through accountability and evidence-based responses.
Why the other options are wrong
- A. Simply writing a confidence level without supporting evidence does not genuinely improve confidence; it may actually undermine it if the reasoning is unclear.
- C. A computer-like avatar is a visual design choice that does not address the substantive concern of response reliability or accuracy.
- D. Matching the company's language style improves engagement and familiarity but does not directly address confidence in the accuracy or reliability of answers.
Question 60
A company has terabytes of data in a database that the company can use for business analysis. The company wants to build an AI-based application that can build a SQL query from input text that employees provide. The employees have minimal experience with technology.
Which solution meets these requirements?
Show answer and explanation
Correct answer: A. Generative pre-trained transformers (GPT)
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.
Question 61
A company wants to improve the accuracy of the responses from a generative AI application. The application uses a foundation model (FM) on Amazon Bedrock.
Which solution meets these requirements MOST cost-effectively?
Show answer and explanation
Correct answer: D. Use prompt engineering.
Prompt engineering is the most cost-effective way to improve FM accuracy on Bedrock. It requires no additional training, computational resources, or model modifications, only iterative refinement of input prompts. Fine-tuning requires labeled data and computational overhead, retraining is prohibitively expensive, and training a new FM from scratch is the most costly option.
Why the other options are wrong
- A. Fine-tuning requires labeled data, infrastructure investment, and significant computational costs compared to prompt engineering.
- B. Retraining an FM is extremely expensive and requires substantial computational resources and expertise.
- C. Training a new FM from scratch is the most expensive and impractical option for this use case.
Question 62
A company wants to create a new solution by using AWS Glue. The company has minimal programming experience with AWS Glue.
Which AWS service can help the company use AWS Glue?
Show answer and explanation
Correct answer: A. Amazon Q Developer
Amazon Q Developer is an AWS service that provides AI-powered assistance for developing, debugging, and optimizing code including AWS Glue jobs. It helps developers with minimal AWS Glue experience by generating code snippets, explaining Glue concepts, and providing real-time guidance for creating Glue solutions.
Why the other options are wrong
- B. AWS Config is a compliance and configuration management service, not a development assistance tool.
- C. Amazon Personalize is a recommendation engine service unrelated to AWS Glue development.
- D. Amazon Comprehend is a natural language processing service, not relevant to Glue development assistance.
Question 63
A company is making a chatbot. The chatbot uses Amazon Lex and Amazon OpenSearch Service. The chatbot uses the company's private data to answer questions. The company needs to convert the data into a vector representation before storing the data in a database.
Which type of foundation model (FM) meets these requirements?
Show answer and explanation
Correct answer: C. Text embeddings model
A text embeddings model is specifically designed to convert text data into vector representations (embeddings) that can be stored and searched in vector databases like Amazon OpenSearch Service. This is essential for semantic search and retrieval-augmented generation (RAG) systems, which allow the chatbot to find and use relevant private data to answer questions.
Why the other options are wrong
- A. Text completion models generate text continuations but do not convert text into vector representations.
- B. Instruction following models follow user instructions to generate responses but are not designed for creating vector embeddings.
- D. Image generation models create images from text descriptions, not relevant to converting text data into vectors for search purposes.
Question 64
A company wants to create a chatbot that answers questions about human resources policies. The company is using a large language model (LLM) and has a large digital documentation base.
Which technique should the company use to optimize the generated responses?
Show answer and explanation
Correct answer: A. Use Retrieval Augmented Generation (RAG).
Retrieval Augmented Generation (RAG) is the optimal technique for this scenario. RAG allows the LLM to retrieve relevant information from the company's large documentation base and incorporate it into responses, ensuring answers are grounded in actual HR policies rather than generated from the model's training data alone, thereby providing accurate and policy-compliant answers.
Why the other options are wrong
- B. Few-shot prompting provides examples but does not leverage the company's documentation base for authoritative policy information.
- C. Setting temperature to 1 maximizes randomness, which is counterproductive for a system that should provide consistent, policy-accurate responses.
- D. Decreasing token size limits response length and does not optimize the quality or accuracy of HR policy answers.
Question 65
A news agency publishes articles in English. The agency wants to make articles available in other languages.
Which solution meets these requirements?
Show answer and explanation
Correct answer: B. Use the Amazon Translate real-time translation feature.
Amazon Translate is the correct service for translating text content into multiple languages. It can process articles in real-time to make them available in other languages. Amazon Transcribe converts speech to text, Amazon Personalize provides recommendations, and Amazon Textract extracts text from documents; none of these directly address translating text content into multiple languages.
Why the other options are wrong
- A. Amazon Transcribe converts speech to text, not for translating written articles into other languages.
- C. Amazon Personalize provides personalized recommendations, not translation services.
- D. Amazon Textract extracts text from documents, not for translating content into other languages.
Question 66
A company wants to keep its foundation model (FM) relevant by using the most recent data. The company wants to implement a model training strategy that includes regular updates to the FM.
Which solution meets these requirements?
Show answer and explanation
Correct answer: B. Continuous pre-training
Continuous pre-training is the technique of regularly updating a foundation model with new data to keep it current and relevant. This allows the model to incorporate recent information and trends while maintaining its general capabilities, directly addressing the requirement for regular updates with the most recent data.
Why the other options are wrong
- A. Batch learning processes data in batches but doesn't specifically address the requirement for regular, continuous updates.
- C. Static training implies a one-time training process without updates, which contradicts the requirement for ongoing relevance.
- D. Latent training is not a standard technique for keeping foundation models updated with recent data.
Question 67
A social media company wants to prevent users from posting discriminatory content on the company's application. The company wants to use Amazon Bedrock as part of the solution.
How can the company use Amazon Bedrock to meet these requirements?
Show answer and explanation
Correct answer: B. Block interactions related to predefined topics.
Guardrails for Amazon Bedrock allows blocking interactions related to predefined topics and policies, enabling the company to prevent discriminatory content by filtering conversations before they reach users. This is the built-in content moderation approach in Bedrock.
Why the other options are wrong
- A. User preference interaction does not prevent discriminatory content from being posted.
- C. Restricting to predefined topics is limiting but does not specifically target discriminatory content.
- D. Offering response variety does not prevent users from posting discriminatory content.
Question 68
A company is using Amazon SageMaker Studio notebooks to build and train ML models. The company stores the data in an Amazon S3 bucket.
The company needs to manage the flow of data from Amazon S3 to SageMaker Studio notebooks.
Which solution will meet this requirement?
Show answer and explanation
Correct answer: C. Configure SageMaker to use a VPC with an S3 endpoint.
Configuring SageMaker to use a VPC with an S3 endpoint enables secure, efficient data flow from Amazon S3 to SageMaker Studio notebooks. The S3 endpoint allows private connectivity without traversing the public internet, optimizing both performance and security for data management between these services.
Why the other options are wrong
- A. Amazon Inspector is a vulnerability assessment tool, not a data flow management solution.
- B. Amazon Macie is a data discovery and protection service for sensitive data, not a data flow management tool.
- D. S3 Glacier Deep Archive is a cold storage class for archival, not suitable for active data access in machine learning workflows.
Question 69
A manufacturing company wants to create product descriptions in multiple languages.
Which AWS service will automate this task?
Show answer and explanation
Correct answer: A. Amazon Translate
Amazon Translate is the AWS service designed to automatically translate text into multiple languages. It can convert product descriptions into various languages at scale, automating the task of creating multilingual content. Transcribe converts speech to text, Kendra provides search functionality, and Polly converts text to speech; none of these services automate translation to multiple languages.
Why the other options are wrong
- B. Amazon Transcribe converts audio to text, not text to multiple languages.
- C. Amazon Kendra is an intelligent search service, not a translation tool.
- D. Amazon Polly converts text to speech in various languages but does not translate content.
Question 70
Which strategy evaluates the accuracy of a foundation model (FM) that is used in image classification tasks?
Show answer and explanation
Correct answer: B. Measure the model's accuracy against a predefined benchmark dataset.
Measuring accuracy against a predefined benchmark dataset is the standard and correct approach to evaluate the performance of a foundation model on image classification tasks. This quantitative metric directly assesses how well the model performs its intended function.
Why the other options are wrong
- A. Resource cost calculation is relevant to operational efficiency, not model accuracy evaluation.
- C. Counting neural network layers is a model architecture characteristic unrelated to accuracy measurement.
- D. Color accuracy of processed images is not a standard metric for evaluating image classification model accuracy.
Question 71
A company wants to use AI to protect its application from threats. The AI solution needs to check if an IP address is from a suspicious source.
Which solution meets these requirements?
Show answer and explanation
Correct answer: C. Develop an anomaly detection system.
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.
Question 72
A company has an ML model. The company wants to know how the model makes predictions.
Which term refers to understanding model predictions?
Show answer and explanation
Correct answer: A. Model interpretability
Model interpretability is the term that specifically refers to understanding how a machine learning model makes predictions, encompassing explainability and transparency of the decision-making process.
Why the other options are wrong
- B. Model training refers to the process of fitting a model to data, not understanding predictions.
- C. Model interoperability concerns compatibility and integration between different systems, not prediction understanding.
- D. Model performance measures accuracy and effectiveness but does not explain how predictions are generated.
Question 73
What is tokenization used for in natural language processing (NLP)?
Show answer and explanation
Correct answer: C. To break text into smaller units for processing
Tokenization is the fundamental NLP process of breaking down text into smaller units called tokens (words, subwords, or characters). This segmentation allows models to process and understand text at a manageable granularity, where each token can be converted to numerical representations for computation.
Why the other options are wrong
- A. Tokenization is not an encryption technique; it is a text parsing operation.
- B. While tokenization may incidentally reduce storage needs, compression is not its primary purpose.
- D. Tokenization is a text processing technique, not a translation mechanism between languages.
Question 74
A company wants to display the total sales for its top-selling products across various retail locations in the past 12 months.
Which AWS solution should the company use to automate the generation of graphs?
Show answer and explanation
Correct answer: C. Amazon Q in Amazon QuickSight
Amazon Q in Amazon QuickSight is the correct solution for automating the generation of graphs and visualizations for business analytics. QuickSight is AWS's native business intelligence and visualization service, and Amazon Q integration within it enables AI-assisted analysis and automated graph generation. Amazon Q in EC2 provides general assistance, Amazon Q Developer is for code development, and Amazon Q in AWS Chatbot is for chat-based notifications; none are designed for automated analytics visualization.
Why the other options are wrong
- A. Amazon Q in EC2 provides general assistance but is not designed for automated graph generation.
- B. Amazon Q Developer is focused on code development rather than business analytics visualization.
- D. Amazon Q in AWS Chatbot is for chat notifications and does not generate analytical graphs.
Question 75
A company is creating an agent for its application by using Amazon Bedrock Agents. The agent is performing well, but the company wants to improve the agent’s accuracy by providing some specific examples.
Which solution meets these requirements?
Show answer and explanation
Correct answer: A. Modify the advanced prompts for the agent to include the examples.
Advanced prompts in Amazon Bedrock Agents allow you to provide specific examples (few-shot prompting) and detailed instructions to improve agent accuracy and behavior. This is the direct mechanism for customizing agent performance without modifying the underlying model or architecture.
Why the other options are wrong
- B. Guardrails are designed to enforce safety constraints and prevent harmful outputs, not to provide examples for improving accuracy.
- C. SageMaker Ground Truth is for labeling large datasets for model training, not for providing quick examples to an existing agent.
- D. Running Lambda scripts to add examples to a training dataset would require retraining the model, which is unnecessary when advanced prompts can achieve the goal directly.
Question 76
A company is building a large language model (LLM) question answering chatbot. The company wants to decrease the number of actions call center employees need to take to respond to customer questions.
Which business objective should the company use to evaluate the effect of the LLM chatbot?
Show answer and explanation
Correct answer: B. Average call duration
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.
Question 77
A company wants to build a generative AI application by using Amazon Bedrock and needs to choose a foundation model (FM). The company wants to know how much information can fit into one prompt.
Which consideration will inform the company's decision?
Show answer and explanation
Correct answer: B. Context window
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.
Question 78
An AI practitioner is evaluating the performance of an Amazon SageMaker model. The AI practitioner must choose a performance metric. The metric must show the ratio of the number of correctly classified items to the total number of correctly and incorrectly classified items.
Which metric meets these requirements?
Show answer and explanation
Correct answer: A. Accuracy
Accuracy is defined as the ratio of correctly classified items to the total number of items classified (both correct and incorrect). This matches exactly the requirement stated in the question. Precision measures correct positive predictions out of predicted positives, Recall measures correct positives out of actual positives, and F1 score combines precision and recall.
Why the other options are wrong
- B. Precision is the ratio of true positives to all predicted positives, not total classifications.
- C. F1 score is the harmonic mean of precision and recall, not the ratio described.
- D. Recall is the ratio of true positives to all actual positives, not total classifications.
Question 79
A company is building a contact center application and wants to gain insights from customer conversations. The company wants to analyze and extract key information from the audio of the customer calls.
Which solution meets these requirements?
Show answer and explanation
Correct answer: B. Transcribe call recordings by using Amazon Transcribe.
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.
Question 80
A company is building a chatbot to improve user experience. The company is using a large language model (LLM) from Amazon Bedrock for intent detection. The company wants to use few-shot learning to improve intent detection accuracy.
Which additional data does the company need to meet these requirements?
Show answer and explanation
Correct answer: C. Pairs of user messages and correct user intents
Few-shot learning for intent detection requires pairs of user messages and their correct intent labels. The model learns to map user inputs to intents by observing a small number of labeled examples. Option C directly provides this mapping. Options A, B, and D involve chatbot responses, which are not needed for intent detection; the task is to classify user messages into intents, not to generate or evaluate responses.
Why the other options are wrong
- A. Chatbot responses are not needed for intent detection; the focus is on classifying user intent.
- B. Chatbot responses are outputs, not inputs needed to train intent detection.
- D. This pairs intents with responses, but intent detection requires mapping messages to intents.
Question 81
A company needs to train an ML model to classify images of different types of animals. The company has a large dataset of labeled images and will not label more data.
Which type of learning should the company use to train the model?
Show answer and explanation
Correct answer: A. Supervised learning
Since the company has a large dataset of labeled images and will not label more data, supervised learning is the appropriate approach. The labeled data can be used directly to train a classifier that learns the relationship between images and animal types. Supervised learning leverages labeled data for classification tasks, which is exactly what this scenario requires.
Why the other options are wrong
- B. Unsupervised learning finds patterns in unlabeled data but would not leverage the existing labeled dataset effectively for classification.
- C. Reinforcement learning requires an agent to learn through interactions and rewards, which is not suitable for image classification with labeled datasets.
- D. Active learning is used when labels are scarce and you need to selectively label additional data, but the company will not label more data.
Question 82
A company wants to make a chatbot to help customers. The chatbot will help solve technical problems without human intervention.
The company chose a foundation model (FM) for the chatbot. The chatbot needs to produce responses that adhere to company tone.
Which solution meets these requirements?
Show answer and explanation
Correct answer: C. Experiment and refine the prompt until the FM produces the desired responses.
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.
Question 83
A financial company is developing a fraud detection system that flags potential fraud cases in credit card transactions. Employees will evaluate the flagged fraud cases. The company wants to minimize the amount of time the employees spend reviewing flagged fraud cases that are not actually fraudulent.
Which evaluation metric meets these requirements?
Show answer and explanation
Correct answer: C. Precision
Precision measures the proportion of flagged fraud cases that are actually fraudulent (true positives divided by all positive predictions). By minimizing false positives, precision directly addresses the company's goal of reducing time employees spend reviewing non-fraudulent cases. A high precision means most flagged cases are genuine fraud, maximizing employee efficiency.
Why the other options are wrong
- A. Recall measures the proportion of actual fraud cases that are caught, which doesn't address minimizing time spent on false positives.
- B. Accuracy measures overall correctness but doesn't specifically minimize false positives in this fraud detection context.
- D. Lift charts are used for model comparison and ranking, not for evaluating the quality of individual predictions.
Question 84
In which stage of the generative AI model lifecycle are tests performed to examine the model's accuracy?
Show answer and explanation
Correct answer: D. Evaluation
The Evaluation stage is where model accuracy and performance are tested against validation datasets. This is the phase where metrics like precision, recall, F1 score, and accuracy are calculated to determine if the model meets quality requirements. Deployment comes after evaluation, Data selection precedes model training, and Fine-tuning is part of the training optimization process.
Why the other options are wrong
- A. Deployment is when the model is put into production, not when testing occurs.
- B. Data selection is the early stage of preparing data, not where accuracy testing happens.
- C. Fine-tuning is the optimization of model parameters, not the primary testing phase.
Question 85
A company wants to create an application by using Amazon Bedrock. The company has a limited budget and prefers flexibility without long-term commitment.
Which Amazon Bedrock pricing model meets these requirements?
Show answer and explanation
Correct answer: A. On-Demand
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.
Question 86
A company deployed a model to production. After 4 months, the model inference quality degraded. The company wants to receive a notification if the model inference quality degrades. The company also wants to ensure that the problem does not happen again.
Which solution will meet these requirements?
Show answer and explanation
Correct answer: B. Retrain the model. Monitor model drift by using Amazon SageMaker Model Monitor.
Amazon SageMaker Model Monitor is the AWS service specifically designed to monitor model performance and data drift in production, providing notifications when metrics degrade beyond thresholds. Re-training the model addresses the immediate degradation, and Model Monitor together ensures ongoing detection. Clarify is for bias and fairness analysis, Feature Store manages feature data, and JumpStart is for quick model deployment, none of which are designed for ongoing drift monitoring and alerting.
Why the other options are wrong
- A. SageMaker Clarify is for analyzing model bias and fairness, not for monitoring model drift and inference quality degradation.
- C. Feature Store manages feature data but is not designed for monitoring model drift, and building a new model doesn't ensure drift detection going forward.
- D. JumpStart is for quickly deploying pre-trained models, not for monitoring model drift and receiving notifications about degradation.
Question 87
A company built an AI-powered resume screening system. The company used a large dataset to train the model. The dataset contained resumes that were not representative of all demographics.
Which core dimension of responsible AI does this scenario present?
Show answer and explanation
Correct answer: A. Fairness
Fairness in responsible AI concerns ensuring that models do not discriminate against or disadvantage specific demographic groups. A training dataset that is not representative of all demographics introduces bias and unfair outcomes for underrepresented groups, directly violating fairness principles.
Why the other options are wrong
- B. Explainability refers to how well the model's decisions can be understood, not whether the training data represents all demographics.
- C. Privacy and security involve protecting personal data and securing systems, not the representativeness of training data demographics.
- D. Transparency refers to disclosing model capabilities and limitations, not the demographic composition of training data.
Question 88
A company’s large language model (LLM) is experiencing hallucinations.
How can the company decrease hallucinations?
Show answer and explanation
Correct answer: C. Decrease the temperature inference parameter for the model.
Decreasing the temperature inference parameter reduces randomness and variability in model outputs, making the model more deterministic and less likely to generate hallucinations. Lower temperature values cause the model to favor more probable tokens, resulting in more grounded and factually consistent responses.
Why the other options are wrong
- A. Agents for Amazon Bedrock help orchestrate tasks but don't directly reduce hallucinations during inference.
- B. Data pre-processing cannot remove hallucinations since they are generated during inference, not caused by training data.
- D. While better-trained models exist, this is not a practical approach; temperature adjustment is a direct and immediate control mechanism.
Question 89
A design company is using a foundation model (FM) on Amazon Bedrock to generate images for various projects. The company wants to have control over how detailed or abstract each generated image appears
Which model parameter should the company modify?
Show answer and explanation
Correct answer: C. Generation step
The generation step parameter controls the number of iterations or steps the model takes to generate an image, which directly affects the level of detail and quality. More generation steps typically produce more detailed images, while fewer steps may result in more abstract or less refined outputs. This parameter allows control over the trade-off between detail and abstraction as described.
Why the other options are wrong
- A. Model checkpoint refers to saved model weights at different training points, not a parameter controlling output detail or abstraction.
- B. Batch size controls how many samples are processed together, not the visual characteristics of generated images.
- D. Token length applies to text generation, not image generation parameters for controlling detail versus abstraction.
Question 90
A company wants to build a lead prioritization application for its employees to contact potential customers. The application must give employees the ability to view and adjust the weights assigned to different variables in the model based on domain knowledge and expertise.
Which ML model type meets these requirements?
Show answer and explanation
Correct answer: A. Logistic regression model
Logistic regression models provide interpretable weights for each variable that employees can view and adjust based on domain expertise. The model coefficients directly represent the contribution of each feature, making it transparent and easy to understand how changes to weights affect predictions. This transparency is essential for a lead prioritization system where domain experts need to fine-tune the model based on business knowledge.
Why the other options are wrong
- B. Deep learning models built on principal components obscure the original variable relationships, making it difficult for employees to understand and adjust weights based on domain knowledge.
- C. K-NN models don't have explicit weights that can be easily viewed and adjusted; they rely on distance calculations that are not transparent to users.
- D. Neural networks are black-box models where individual weights are difficult for non-experts to interpret and adjust meaningfully based on domain expertise.
Question 91
Which prompting attack directly exposes the configured behavior of a large language model (LLM)?
Show answer and explanation
Correct answer: D. Extracting the prompt template
Extracting the prompt template is a direct attack that exposes the configured behavior, system instructions, and hidden directives that govern how an LLM responds. By revealing the underlying prompt structure, an attacker gains insight into the model's configured constraints and operational logic.
Why the other options are wrong
- A. Prompted persona switches attempt to change the model's behavior but don't directly expose its configuration.
- B. Exploiting friendliness and trust manipulates the model's tendency to be helpful but doesn't expose internal configuration.
- C. Ignoring the prompt template is a result of attacks but isn't the attack itself that directly exposes configuration.
Question 92
A company deployed a Retrieval Augmented Generation (RAG) application on Amazon Bedrock that gathers financial news to distribute in daily newsletters. Users have recently reported politically influenced ideas in the newsletters.
Which Amazon Bedrock guardrail can identify and filter this content?
Show answer and explanation
Correct answer: B. Denied topics
Denied topics is the Amazon Bedrock guardrail that can identify and filter politically influenced or bias-driven content. This guardrail allows specification of topics that should be avoided or filtered from responses, enabling the company to prevent politically influenced ideas from appearing in generated newsletters. It is specifically designed to exclude content related to specified topics.
Why the other options are wrong
- A. Word filters focus on blocking specific words or phrases, not on detecting politically influenced ideas or bias.
- C. Sensitive information filters protect personal data like credentials or PII, not ideological bias or political content.
- D. Content filters are generic safety measures, not specifically targeted at identifying and blocking politically influenced content.
Question 93
A financial institution is building an AI solution to make loan approval decisions by using a foundation model (FM). For security and audit purposes, the company needs the AI solution's decisions to be explainable.
Which factor relates to the explainability of the AI solution's decisions?
Show answer and explanation
Correct answer: A. Model complexity
Model complexity directly relates to explainability: simpler models are easier to interpret and explain, while highly complex models (like deep neural networks) are often black boxes. For a loan approval system requiring audit trails and explainable decisions, model complexity is the primary factor determining whether decisions can be understood and justified. Training time, hyperparameters, and deployment time are operational factors unrelated to explainability.
Why the other options are wrong
- B. Training time reflects computational resources but has no direct relationship to decision explainability.
- C. Number of hyperparameters is a tuning concern, not a factor in whether decisions are explainable.
- D. Deployment time is an operational metric unrelated to the interpretability of the model's decisions.
Question 94
A company needs to build its own large language model (LLM) based on only the company's private data. The company is concerned about the environmental effect of the training process.
Which Amazon EC2 instance type has the LEAST environmental effect when training LLMs?
Show answer and explanation
Correct answer: D. Amazon EC2 Trn series
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.
Question 95
An education company waftion. The application will give users the ability to enter text or provide a picture of a question. The application will respond with a written answer and an explanation of the written answer.
Which model type meets these requirements?
Show answer and explanation
Correct answer: B. Large multi-modal language model
A large multi-modal language model can process both text and image inputs and generate text outputs with explanations. This capability to handle multiple input modalities (pictures and text) and produce text responses matches the application requirements exactly.
Why the other options are wrong
- A. Computer vision models process images but do not generate natural language explanations effectively.
- C. Diffusion models generate images, not text responses to questions.
- D. Text-to-speech models convert text to audio, not answer questions or process image inputs.
Question 96
A company is testing the security of a foundation model (FM). During testing, the company wants to get around the safety features and make harmful content.
Which security technique is this an example of?
Show answer and explanation
Correct answer: D. Jailbreak
A jailbreak is a technique used to bypass safety features and guardrails in AI systems to generate harmful content. This directly matches the scenario where the company is attempting to circumvent safety mechanisms to produce harmful outputs. Jailbreaking is a well-known adversarial technique specific to large language models and foundation models.
Why the other options are wrong
- A. Fuzzing training data finds vulnerabilities in the training process itself, not a method to bypass safety features of a deployed model.
- B. Denial of service attacks aim to make systems unavailable, not to generate harmful content while bypassing safety features.
- C. Authorized penetration testing is a legitimate security practice, but the scenario describes attempting to 'get around' safety features without explicit authorization language, indicating unauthorized testing.
Question 97
A company wants to find groups for its customers based on the customers’ demographics and buying patterns.
Which algorithm should the company use to meet this requirement?
Show answer and explanation
Correct answer: B. K-means
K-means is the standard unsupervised clustering algorithm designed specifically for grouping data points into k clusters based on similarities. It effectively partitions customers into groups based on demographic and behavioral features, making it ideal for customer segmentation tasks.
Why the other options are wrong
- A. K-nearest neighbors is a classification and regression algorithm, not a primary clustering method for grouping.
- C. Decision trees are primarily used for supervised classification tasks, not unsupervised clustering.
- D. Support vector machines are classifiers for supervised learning, not designed for finding natural groupings in unlabeled data.
Question 98
A company wants to create a chatbot to answer employee questions about company policies. Company policies are updated frequently. The chatbot must reflect the changes in near real time. The company wants to choose a large language model (LLM).
Which solution meets these requirements?
Show answer and explanation
Correct answer: C. Create a Retrieval Augmented Generation (RAG) workflow by using Amazon Bedrock Knowledge Bases.
Retrieval Augmented Generation (RAG) using Amazon Bedrock Knowledge Bases is the optimal solution for frequently updated company policies. RAG retrieves current policy information from knowledge bases in real-time and augments LLM responses without requiring model retraining. Fine-tuning requires retraining when policies change, basic foundation models lack access to current data, and Amazon Q is designed for different use cases.
Why the other options are wrong
- A. Fine-tuning requires retraining the model each time policies update, which is inefficient for frequently changing content.
- B. A basic foundation model without retrieval cannot access or reflect near-real-time policy updates.
- D. Amazon Q Business is designed for different enterprise search use cases, not specifically for policy-based chatbots needing frequent updates.
Question 99
A company is introducing a mobile app that helps users learn foreign languages. The app makes text more coherent by calling a large language model (LLM). The company collected a diverse dataset of text and supplemented the dataset with examples of more readable versions. The company wants the LLM output to resemble the provided examples.
Which metric should the company use to assess whether the LLM meets these requirements?
Show answer and explanation
Correct answer: C. Recall-Oriented Understudy for Gisting Evaluation (ROUGE) score
ROUGE (Recall-Oriented Understudy for Gisting Evaluation) score is specifically designed to measure how well generated text matches reference examples by comparing overlapping n-grams, making it ideal for assessing whether LLM output resembles provided readable examples. The loss function measures training performance rather than output quality relative to examples, semantic robustness measures consistency under perturbations, and latency measures speed rather than output quality.
Why the other options are wrong
- A. Loss function value reflects training convergence, not how well output matches provided examples.
- B. Semantic robustness measures consistency under input variations, not alignment with reference examples.
- D. Latency measures generation speed, which is unrelated to whether output matches the provided readable examples.
Question 100
A company has created a custom model by fine-tuning an existing large language model (LLM) from Amazon Bedrock. The company wants to deploy the model to production and use the model to handle a steady rate of requests each minute.
Which solution meets these requirements MOST cost-effectively?
Show answer and explanation
Correct answer: D. Purchase Provisioned Throughput for the model on Amazon Bedrock.
Provisioned Throughput on Amazon Bedrock provides reserved capacity at a predictable rate, making it cost-effective for steady, consistent workload patterns. This approach is more economical than on-demand pricing when request rates are predictable and sustained, as the company can commit to a baseline throughput level at a lower per-unit cost.
Why the other options are wrong
- A. Using EC2 instances requires managing infrastructure, licensing, and scaling; it's more complex and not necessarily more cost-effective than Bedrock's provisioned throughput.
- B. On-demand throughput has higher per-unit costs for steady, predictable traffic compared to reserved provisioned throughput.
- C. Storing the model in S3 and hosting via Lambda introduces unnecessary complexity, cold start delays, and higher costs compared to using Bedrock's managed provisioned throughput.
Question 101
A company receives a large amount of unstructured user feedback in text format. The company wants to analyze the sentiment of the user feedback.
Which solution will meet these requirements?
Show answer and explanation
Correct answer: A. Use a large language model (LLM) to perform natural language processing (NLP) for sentiment analysis.
Large Language Models excel at Natural Language Processing tasks including sentiment analysis. LLMs can understand context, nuance, and meaning in unstructured text feedback to accurately determine sentiment without requiring predefined categories or historical training data.
Why the other options are wrong
- B. Regression algorithms predict continuous values, not sentiment classification; they require predefined categories which defeats the analysis purpose.
- C. Recommendation engines suggest items based on user preferences and behavior, not designed for sentiment analysis.
- D. Time series algorithms predict future values based on temporal patterns, not for analyzing unstructured text sentiment.
Question 102
A company is developing an ML application. The application must automatically group similar customers and products based on their characteristics.
Which ML strategy should the company use to meet these requirements?
Show answer and explanation
Correct answer: A. Unsupervised learning
Unsupervised learning is the correct approach for automatically grouping similar items without labeled training data. Clustering algorithms like k-means discover natural groupings in customer and product characteristics. Supervised learning requires labeled data with known outcomes, reinforcement learning is for decision-making through trial and error, and semi-supervised learning combines both labeled and unlabeled data but is unnecessary when only automatic grouping is needed.
Why the other options are wrong
- B. Supervised learning requires labeled training data indicating which groups items belong to, but the requirement is automatic discovery.
- C. Reinforcement learning is for learning optimal actions through rewards and penalties, not for grouping similar items.
- D. Semi-supervised learning combines labeled and unlabeled data but is unnecessary for the automatic grouping requirement.
Question 103
A company wants to add generative AI functionality to its application by integrating a large language model (LLM). The responses from the LLM must be as deterministic and as stable as possible.
Which solution meets these requirements?
Show answer and explanation
Correct answer: A. Configure the application to automatically set the temperature parameter to 0 when submitting the prompt to the LLM.
The temperature parameter controls randomness in model outputs; setting it to 0 makes the model fully deterministic by always selecting the highest probability token, eliminating randomness. This directly addresses the requirement for stable, deterministic responses. Temperature of 1 increases variability. Adding text prompts cannot control the underlying probabilistic behavior of the model.
Why the other options are wrong
- B. Adding instructions in the prompt cannot override the probabilistic sampling mechanism.
- C. Prompt placement does not fundamentally change the model's probabilistic behavior.
- D. Temperature set to 1 increases randomness rather than reducing it.
Question 104
A social media company wants to use a large language model (LLM) to summarize messages. The company has chosen a few LLMs that are available on Amazon SageMaker JumpStart. The company wants to compare the generated output toxicity of these models.
Which strategy gives the company the ability to evaluate the LLMs with the LEAST operational overhead?
Show answer and explanation
Correct answer: B. Automatic model evaluation
Automatic model evaluation provides the least operational overhead by using metrics and algorithms to assess model outputs without human involvement. For evaluating toxicity, automated metrics can quickly measure output quality across multiple models without requiring crowd sourcing, hiring, or feedback collection infrastructure.
Why the other options are wrong
- A. Crowd-sourced evaluation requires managing external contributors and scaling evaluation efforts, creating significant overhead.
- C. Human worker evaluation involves recruitment, management, and quality assurance processes that add substantial operational burden.
- D. RLHF is a training technique to improve models, not an evaluation strategy for comparing existing models.
Question 105
A manufacturing company has an application that ingests consumer complaints from publicly available sources. The application uses complex hard-coded logic to process the complaints. The company wants to scale this logic across markets and product lines.
Which advantage do generative AI models offer for this scenario?
Show answer and explanation
Correct answer: B. Adaptability
Generative AI models offer adaptability as a key advantage because they can generalize across different contexts, markets, and product lines without requiring hard-coded logic changes. Rather than maintaining complex conditional rules for each market or product, a generative AI model trained on complaint data can learn patterns and adapt its responses to new variations, making it easy to scale across diverse scenarios without extensive manual reconfiguration.
Why the other options are wrong
- A. Generative AI models are known for variability in outputs, not predictability, which is actually a limitation in some use cases.
- C. Generative AI models are sensitive to input variations and can produce different outputs based on subtle changes in prompts or data.
- D. Generative AI models are often considered black boxes with limited explainability compared to traditional rule-based systems.
Question 106
An ecommerce company is deploying a chatbot. The chatbot will give users the ability to ask questions about the company’s products and receive details on users’ orders. The company must implement safeguards for the chatbot to filter harmful content from the input prompts and chatbot responses.
Which AWS feature or resource meets these requirements?
Show answer and explanation
Correct answer: A. Amazon Bedrock Guardrails
Amazon Bedrock Guardrails is specifically designed to filter harmful content from both input prompts and model responses, implementing safeguards for chatbot deployments. It allows companies to define and enforce safety policies to prevent inappropriate or harmful outputs.
Why the other options are wrong
- B. Bedrock Agents enable building AI assistants with tool integration but do not provide content filtering capabilities.
- C. Inference APIs provide model access but lack built-in content filtering and safeguard mechanisms.
- D. Custom models focus on training domain-specific models rather than filtering harmful content.
Question 107
An ecommerce company receives multiple gigabytes of customer data daily. The company uses the data to train an ML model to forecast future product demand. The company needs a solution to perform inferences once each day.
Which inference type meets these requirements?
Show answer and explanation
Correct answer: A. Batch inference
Batch inference is the correct choice for this scenario because the company needs to perform inferences once each day on multiple gigabytes of data. Batch inference processes large volumes of data at scheduled intervals, making it ideal for daily forecasting tasks where real-time responses are not required. This approach is cost-effective for non-urgent inference workloads.
Why the other options are wrong
- B. Asynchronous inference is for long-running predictions with unpredictable latency, not specifically designed for scheduled daily batch processing.
- C. Real-time inference is for immediate predictions on individual or small requests, not suitable for processing gigabytes of data once daily.
- D. Serverless inference refers to a deployment option without managing servers, but doesn't address the core requirement of daily scheduled batch processing.
Question 108
Which ML technique ensures data compliance and privacy when training AI models on AWS?
Show answer and explanation
Correct answer: C. Federated learning
Federated learning is a distributed machine learning technique that trains models on decentralized data while keeping data at its source, ensuring privacy and compliance. Unlike centralized approaches where data is collected and moved to a central location, federated learning allows AWS and other platforms to train models without directly accessing or moving sensitive data, making it the ideal technique for privacy and compliance requirements.
Why the other options are wrong
- A. Reinforcement learning is a training paradigm focused on learning through rewards, not a privacy-preserving technique.
- B. Transfer learning reuses knowledge from pre-trained models but does not inherently address data privacy or compliance.
- D. Unsupervised learning is an algorithm category that discovers patterns in data without labels, not a privacy-preserving approach.
Question 109
A company wants to implement a large language model (LLM) based chatbot to provide customer service agents with real-time contextual responses to customers' inquiries. The company will use the company's policies as the knowledge base.
Which solution will meet these requirements MOST cost-effectively?
Show answer and explanation
Correct answer: C. Implement Retrieval Augmented Generation (RAG) for in-context responses.
Retrieval Augmented Generation (RAG) is the most cost-effective solution for incorporating external knowledge bases into LLM responses. It retrieves relevant policy documents at inference time without requiring model retraining or fine-tuning, avoiding the significant computational costs and time associated with model customization while maintaining real-time contextual accuracy.
Why the other options are wrong
- A. Retraining an LLM from scratch is extremely expensive and time-consuming, making it cost-prohibitive.
- B. Fine-tuning still requires substantial computational resources and is more expensive than RAG for knowledge base integration.
- D. Pre-training and data augmentation are even more resource-intensive than fine-tuning and unnecessary for this use case.
Question 110
A company wants to use large language models (LLMs) with Amazon Bedrock to develop a chat interface for the company's product manuals. The manuals are stored as PDF files.
Which solution meets these requirements MOST cost-effectively?
Show answer and explanation
Correct answer: D. Upload PDF documents to an Amazon Bedrock knowledge base. Use the knowledge base to provide context when users submit prompts to Amazon Bedrock.
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.
Question 111
An AI practitioner is using an Amazon Bedrock base model to summarize session chats from the customer service department. The AI practitioner wants to store invocation logs to monitor model input and output data.
Which strategy should the AI practitioner use?
Show answer and explanation
Correct answer: B. Enable invocation logging in Amazon Bedrock.
Amazon Bedrock has a native invocation logging feature that captures model inputs and outputs for monitoring and compliance. This is the purpose-built solution for storing logs from Bedrock model invocations.
Why the other options are wrong
- A. AWS CloudTrail logs API calls and AWS service actions, not model invocation inputs and outputs.
- C. AWS Audit Manager is for compliance auditing and assessment, not for capturing model invocation logs.
- D. Amazon EventBridge is an event routing service; it doesn't specifically store model invocation logs as a primary function.
Question 112
A company that uses multiple ML models wants to identify changes in original model quality so that the company can resolve any issues.
Which AWS service or feature meets these requirements?
Show answer and explanation
Correct answer: D. Amazon SageMaker Model Monitor
Amazon SageMaker Model Monitor is the AWS service designed to detect changes and degradation in model quality over time. It continuously monitors model predictions and performance metrics, identifies data drift, prediction drift, and other quality issues, and alerts users when model performance deviates from baseline, enabling proactive issue resolution.
Why the other options are wrong
- A. SageMaker JumpStart provides pre-built models and solutions, not monitoring capabilities for detecting model quality changes.
- B. SageMaker HyperPod provides distributed training infrastructure, not model quality monitoring functionality.
- C. Data Wrangler is for data preparation and feature engineering, not for monitoring deployed model quality.
Question 113
An AI practitioner wants to evaluate ML models. The AI practitioner wants to provide explanations of model predictions to customers and stakeholders.
Which AWS service or feature will meet these requirements?
Show answer and explanation
Correct answer: D. Amazon SageMaker Clarify
Amazon SageMaker Clarify provides tools to explain model predictions through feature importance analysis, SHAP values, and other explainability techniques. It enables practitioners to generate clear explanations of how models make predictions, essential for communicating results to customers and stakeholders.
Why the other options are wrong
- A. Amazon QuickSight is a business analytics and visualization tool, not designed for explaining individual model predictions.
- B. Amazon Comprehend performs NLP tasks like sentiment analysis, not model prediction explanation.
- C. AWS Trusted Advisor provides recommendations for AWS account optimization, not ML model explainability.
Question 114
A financial company uses AWS to host its generative AI models. The company must generate reports to show adherence to international regulations for handling sensitive customer data.
Which AWS service meets these requirements?
Show answer and explanation
Correct answer: B. AWS Artifact
AWS Artifact provides access to AWS compliance documentation, certifications, and reports (such as SOC 2, ISO 27001, and other regulatory frameworks) that demonstrate how AWS meets international regulations for data handling. This is specifically designed for companies needing to generate compliance reports for regulatory adherence.
Why the other options are wrong
- A. Amazon Macie discovers and protects sensitive data but does not generate compliance reports.
- C. AWS Secrets Manager manages credentials and secrets, not compliance documentation.
- D. AWS Config tracks resource compliance and configuration changes, not regulatory compliance reporting.
Question 115
A company needs to select a generative AI model to build an application. The application must provide responses to users in real time.
Which model characteristic should the company consider to meet these requirements?
Show answer and explanation
Correct answer: C. Inference speed
Inference speed directly determines how quickly the model can process inputs and return outputs to users. For real-time applications requiring immediate responses, inference speed is the critical characteristic. Model complexity and training time affect development, while innovation speed relates to model capabilities rather than performance latency.
Why the other options are wrong
- A. Model complexity affects capability and inference speed but is not the primary metric for real-time responsiveness.
- B. Innovation speed relates to model advancement, not to real-time response requirements.
- D. Training time is a one-time consideration during development, not relevant to real-time user-facing performance.
Question 116
What is the purpose of vector embeddings in a large language model (LLM)?
Show answer and explanation
Correct answer: C. Providing the ability to mathematically compare texts
Vector embeddings convert text and other data into numerical representations (vectors) in high-dimensional space where semantic similarity is preserved through mathematical distance. This enables LLMs to mathematically compare and measure the similarity or relationships between different pieces of text, enabling operations like semantic search, clustering, and similarity matching. This mathematical comparison capability is fundamental to how modern LLMs understand and work with language.
Why the other options are wrong
- A. Splitting text into manageable pieces is tokenization, not the purpose of vector embeddings.
- B. Grouping characters into units is also part of tokenization, not vector embeddings.
- D. Counting word occurrences is a bag-of-words approach, not the purpose or capability of vector embeddings.
Question 117
A company needs to use Amazon SageMaker for model training and inference. The company must comply with regulatory requirements to run SageMaker jobs in an isolated environment without internet access.
Which solution will meet these requirements?
Show answer and explanation
Correct answer: B. Run SageMaker training and Inference by using network Isolation.
Network isolation is the SageMaker feature specifically designed to run training and inference jobs in an isolated environment without internet access. This directly addresses the regulatory requirement for isolated, internet-free execution while maintaining compliance standards. Network isolation prevents the SageMaker job from accessing external resources and the internet.
Why the other options are wrong
- A. SageMaker Experiments is a tool for tracking and comparing experiments, not for network isolation or restricting internet access.
- C. Geospatial encryption capabilities are specific to geospatial data processing and do not provide network isolation requirements.
- D. IAM roles control access permissions and authentication but do not provide network isolation or prevent internet access.
Question 118
Which AWS service makes foundation models (FMs) available to help users build and scale generative AI applications?
Show answer and explanation
Correct answer: B. Amazon Bedrock
Amazon Bedrock is AWS's managed service that provides access to foundation models from various providers, allowing users to build and scale generative AI applications without managing infrastructure. It abstracts the complexity of working with multiple FMs and provides a unified API for accessing these models.
Why the other options are wrong
- A. Amazon Q Developer is an AI assistant for developers but doesn't provide foundation models as a core service.
- C. Amazon Kendra is a search service for retrieving information, not a foundation model platform.
- D. Amazon Comprehend is focused on natural language processing tasks like sentiment analysis and entity recognition, not providing foundation models.
Question 119
A company has a foundation model (FM) that was customized by using Amazon Bedrock to answer customer queries about products. The company wants to validate the model's responses to new types of queries. The company needs to upload a new dataset that Amazon Bedrock can use for validation.
Which AWS service meets these requirements?
Show answer and explanation
Correct answer: A. Amazon S3
Amazon S3 is the standard storage service used with Amazon Bedrock for uploading and managing datasets for fine-tuning, validation, and evaluation. Bedrock integrates natively with S3, making it the appropriate choice for storing datasets that will be used for model validation.
Why the other options are wrong
- B. Amazon EBS is block storage for EC2 instances and is not integrated with Bedrock's dataset management.
- C. Amazon EFS is a file system for EC2 and on-premises servers, not designed for Bedrock integration.
- D. AWS Snowcone is for data transfer in disconnected or remote environments, not the primary interface for Bedrock dataset uploads.
Question 120
A company deployed an AI/ML solution to help customer service agents respond to frequently asked questions. The questions can change over time. The company wants to give customer service agents the ability to ask questions and receive automatically generated answers to common customer questions.
Which strategy will meet these requirements MOST cost-effectively?
Show answer and explanation
Correct answer: D. Use Retrieval Augmented Generation (RAG) with prompt engineering techniques.
Retrieval Augmented Generation (RAG) with prompt engineering is the most cost-effective approach because it allows the model to dynamically retrieve relevant context and answer questions without expensive fine-tuning or retraining. It adapts to changing questions through prompt engineering rather than model modification.
Why the other options are wrong
- A. Fine-tuning regularly is computationally expensive and requires retraining, making it less cost-effective than RAG.
- B. Training the model with context data requires significant computational resources and model updates.
- C. Pre-training and benchmarking are also expensive and require substantial upfront investment compared to RAG approaches.
Question 121
An online media streaming company wants to give its customers the ability to perform natural language-based image search and filtering. The company needs a vector database that can help with similarity searches and nearest neighbor queries.
Which AWS service meets these requirements?
Show answer and explanation
Correct answer: D. Amazon OpenSearch Service
Amazon OpenSearch Service supports vector search capabilities and can function as a vector database for similarity searches and nearest neighbor queries, making it ideal for natural language-based image search and filtering applications. Comprehend is for natural language processing and sentiment analysis, Personalize is for recommendation systems, and Polly is for text-to-speech conversion.
Why the other options are wrong
- A. Comprehend performs NLP tasks like sentiment analysis and entity recognition but is not a vector database.
- B. Personalize is designed for building recommendation systems, not for vector similarity searches.
- C. Polly is a text-to-speech service and has no capability for vector databases or similarity searches.
Question 122
A hospital wants to use a generative AI solution with speech-to-text functionality to help improve employee skills in dictating clinical notes.
Which AWS service meets these requirements?
Show answer and explanation
Correct answer: D. AWS HealthScribe
AWS HealthScribe is specifically designed for the healthcare industry and provides speech-to-text functionality that automatically converts clinical dictation into structured clinical notes. It is purpose-built to help healthcare professionals improve their dictation practices and streamline documentation, making it the ideal AWS service for this hospital's requirement to improve employee skills in dictating clinical notes.
Why the other options are wrong
- A. Amazon Q Developer is a generative AI assistant for software development tasks, not for clinical documentation.
- B. Amazon Polly converts text to speech, which is the opposite of the required speech-to-text functionality.
- C. Amazon Rekognition performs image and video analysis, not speech-to-text or clinical documentation.
Question 123
A food service company wants to develop an ML model to help decrease daily food waste and increase sales revenue. The company needs to continuously improve the model's accuracy.
Which solution meets these requirements?
Show answer and explanation
Correct answer: A. Use Amazon SageMaker and iterate with newer data.
Amazon SageMaker is the appropriate solution for developing and continuously improving ML models. It provides tools for training, evaluating, and iterating models with new data over time. The food service company can use SageMaker to build a model for predicting food waste and sales optimization, then continuously retrain it with newer data to improve accuracy and meet business objectives.
Why the other options are wrong
- B. Amazon Personalize is designed for recommendation systems, not for predicting food waste and optimizing sales through iterative model improvement.
- C. Amazon CloudWatch is a monitoring and logging service, not an ML platform for building and iterating models.
- D. Amazon Rekognition is designed for image and video analysis, not for food waste prediction and revenue optimization modeling.
Question 124
A company is implementing the Amazon Titan foundation model (FM) by using Amazon Bedrock. The company needs to supplement the model by using relevant data from the company's private data sources.
Which solution will meet this requirement?
Show answer and explanation
Correct answer: C. Create an Amazon Bedrock knowledge base.
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.
Question 125
An education provider is building a question and answer application that uses a generative AI model to explain complex concepts. The education provider wants to automatically change the style of the model response depending on who is asking the question. The education provider will give the model the age range of the user who has asked the question.
Which solution meets these requirements with the LEAST implementation effort?
Show answer and explanation
Correct answer: B. Add a role description to the prompt context that instructs the model of the age range that the response should target.
Adding a role description to the prompt context that specifies the target age range requires minimal implementation effort and no model retraining. The LLM can adjust its response style based on contextual instructions in the prompt, making this the least effortful solution.
Why the other options are wrong
- A. Fine-tuning requires collecting large representative datasets for each age range and retraining, which is significantly more effort.
- C. Chain-of-thought reasoning adds complexity to the prompt and doesn't directly address the age-based style adaptation requirement.
- D. Summarizing post-response is a post-processing step that doesn't truly adapt the explanation style and complexity as the model generates it.
Question 126
A company manually reviews all submitted resumes in PDF format. As the company grows, the company expects the volume of resumes to exceed the company's review capacity. The company needs an automated system to convert the PDF resumes into plain text format for additional processing.
Which AWS service meets this requirement?
Show answer and explanation
Correct answer: A. Amazon Textract
Amazon Textract is AWS's document analysis service that automatically extracts text and data from PDFs and images, converting them into machine-readable formats including plain text. It is purpose-built for this document processing task.
Why the other options are wrong
- B. Amazon Personalize is for building recommendation systems, not document conversion.
- C. Amazon Lex is for conversational interfaces and chatbots, not document text extraction.
- D. Amazon Transcribe converts audio to text, not PDF documents to plain text.
Question 127
A company has created multiple ML models. The company needs a solution for storing, managing, and versioning the models.
Which AWS service or feature meets these requirements?
Show answer and explanation
Correct answer: C. Amazon SageMaker Model Registry
Amazon SageMaker Model Registry is specifically designed for storing, managing, versioning, and organizing ML models. It provides a centralized repository where data scientists and engineers can register models, track versions, manage model metadata, and control model lifecycle. AWS Audit Manager is for compliance auditing, Model Monitor tracks model performance in production, and Canvas is a no-code tool for building models.
Why the other options are wrong
- A. AWS Audit Manager is for compliance auditing and governance, not for model storage and versioning.
- B. Amazon SageMaker Model Monitor tracks model performance and data quality in production, not for storing and versioning models.
- D. Amazon SageMaker Canvas is a no-code tool for building ML models, not for managing or versioning them.
Question 128
An AI practitioner wants to use a foundation model (FM) to design a search application. The search application must handle queries that have text and images.
Which type of FM should the AI practitioner use to power the search application?
Show answer and explanation
Correct answer: A. Multi-modal embedding model
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.
Question 129
Which statement correctly describes embeddings in generative AI?
Show answer and explanation
Correct answer: A. Embeddings represent data as high-dimensional vectors that capture semantic relationships.
Embeddings are mathematical representations of data (text, images, etc.) as high-dimensional vectors that encode semantic meaning and relationships. They allow models to understand similarity and context between different pieces of information. Option B describes Retrieval-Augmented Generation (RAG), Option C describes quantization, and Option D describes vector databases.
Why the other options are wrong
- B. This describes retrieval or search functionality, not embeddings themselves.
- C. This describes quantization, a technique for reducing model size, not embeddings.
- D. This describes vector storage systems, which store embeddings but are not embeddings.
Question 130
A company needs to monitor the performance of its ML systems by using a highly scalable AWS service.
Which AWS service meets these requirements?
Show answer and explanation
Correct answer: A. Amazon CloudWatch
Amazon CloudWatch is a highly scalable monitoring service that provides metrics, logs, and alarms for ML systems and other AWS resources. It can monitor model performance, training jobs, and inference endpoints at scale. CloudTrail tracks API calls, Trusted Advisor provides resource optimization recommendations, and Config tracks configuration changes; none of these are designed specifically for monitoring ML system performance metrics.
Why the other options are wrong
- B. AWS CloudTrail logs API calls and account activity, not ML system performance metrics.
- C. AWS Trusted Advisor provides resource optimization and security recommendations, not ML performance monitoring.
- D. AWS Config tracks configuration compliance and changes, not ML system performance metrics.
Question 131
Which option is a characteristic of transformer-based language models?
Show answer and explanation
Correct answer: C. Transformer-based language models use self-attention mechanisms to capture contextual relationships.
Self-attention mechanisms are the core innovation of transformer-based models. They allow each token to attend to and learn relationships with all other tokens in the input sequence simultaneously, capturing long-range dependencies and contextual relationships effectively without the sequential processing limitations of recurrent models.
Why the other options are wrong
- A. Transformers do not primarily use convolutional layers; they use attention mechanisms instead, which operate globally rather than locally.
- B. Modern transformer models can process multimodal data including images, audio, and video, not just text.
- D. Transformers process entire sequences in parallel using attention, not one element at a time in cyclic iterations.
Question 132
A manufacturing company uses AI to inspect products and find any damages or defects.
Which type of AI application is the company using?
Show answer and explanation
Correct answer: C. Computer vision
Computer vision is the AI application type that enables machines to interpret and analyze visual information from images and video. In this manufacturing scenario, computer vision is used to inspect products by analyzing images to detect damages, defects, and quality issues, which is a core computer vision use case.
Why the other options are wrong
- A. Recommendation systems suggest products or content to users based on preferences and behavior, not for defect detection.
- B. Natural language processing deals with text and language understanding, not visual inspection of products.
- D. Image processing is a lower-level technique for manipulating pixel data, whereas computer vision is the broader field that uses image processing for understanding and analyzing visual content.
Question 133
A company has fine-tuned a large language model (LLM) to answer questions for a help desk. The company wants to determine if the fine-tuning has enhanced the model's accuracy.
Which metric should the company use for the evaluation?
Show answer and explanation
Correct answer: C. F1 score
The F1 score is the appropriate metric for evaluating a fine-tuned LLM's accuracy on a help desk question-answering task. F1 score balances precision and recall, providing a comprehensive measure of overall accuracy that accounts for both false positives and false negatives, making it ideal for assessing whether fine-tuning has enhanced model performance.
Why the other options are wrong
- A. Precision alone measures only true positives among predicted positives and ignores recall performance.
- B. Time to first token measures latency, not accuracy or the effectiveness of fine-tuning.
- D. Word error rate is primarily used for speech recognition tasks, not for evaluating question-answering accuracy.
Question 134
A company has documents that are missing some words because of a database error. The company wants to build an ML model that can suggest potential words to fill in the missing text.
Which type of model meets this requirement?
Show answer and explanation
Correct answer: D. BERT-based models
BERT-based models are specifically designed for natural language understanding and masked language modeling tasks, which directly address the requirement of predicting missing words in text. BERT can be fine-tuned to fill in missing words by learning contextual representations. Topic modeling identifies document themes, clustering models group similar data points, and prescriptive ML models recommend actions; none of these directly solve word prediction for missing text.
Why the other options are wrong
- A. Topic modeling identifies themes in documents rather than predicting specific missing words.
- B. Clustering models group similar items but do not predict missing words in text.
- C. Prescriptive ML recommends actions but is not designed for text completion tasks.
Question 135
A company wants to enhance response quality for a large language model (LLM) for complex problem-solving tasks. The tasks require detailed reasoning and a step-by-step explanation process.
Which prompt engineering technique meets these requirements?
Show answer and explanation
Correct answer: D. Chain-of-thought prompting
Chain-of-thought prompting is specifically designed for complex problem-solving by encouraging the model to explain its reasoning step-by-step. This technique dramatically improves performance on tasks requiring detailed reasoning and logical progression, making it ideal for enhancing response quality in complex scenarios.
Why the other options are wrong
- A. Few-shot prompting provides examples but doesn't inherently encourage step-by-step reasoning for complex problems.
- B. Zero-shot prompting attempts tasks without examples, making it less effective for complex reasoning requiring detailed explanations.
- C. Directional stimulus prompting is not a standard prompt engineering technique for reasoning tasks.
Question 136
A company has developed a generative text summarization model by using Amazon Bedrock. The company will use Amazon Bedrock automatic model evaluation capabilities.
Which metric should the company use to evaluate the accuracy of the model?
Show answer and explanation
Correct answer: C. BERTScore
BERTScore is a semantic evaluation metric specifically designed for natural language generation tasks like summarization. It evaluates the similarity between generated and reference texts using contextual embeddings, making it ideal for assessing summarization accuracy in generative models.
Why the other options are wrong
- A. AUC is used for binary classification tasks, not for evaluating text summarization quality.
- B. F1 score is a classification metric, not appropriate for evaluating generative summarization accuracy.
- D. RWK score is not a standard metric in Amazon Bedrock or established ML evaluation frameworks.
Question 137
Which ML technique uses training data that is labeled with the correct output values?
Show answer and explanation
Correct answer: A. Supervised learning
Supervised learning is the ML technique that uses labeled training data where each input has a corresponding correct output value. The model learns to map inputs to outputs by studying these labeled examples, making it the fundamental approach for tasks like regression and classification.
Why the other options are wrong
- B. Unsupervised learning works with unlabeled data and finds patterns without predefined correct outputs.
- C. Reinforcement learning uses rewards and penalties rather than labeled correct outputs to guide model learning.
- D. Transfer learning reuses knowledge from one task to another but depends on the underlying learning paradigm (supervised, unsupervised, or reinforcement).
Question 138
Which strategy will determine if a foundation model (FM) effectively meets business objectives?
Show answer and explanation
Correct answer: C. Assess the model's alignment with specific use cases.
Determining whether a foundation model effectively meets business objectives requires assessing its alignment with specific use cases. While benchmark performance is useful, it doesn't guarantee the model solves the actual business problem. The critical evaluation is whether the model's capabilities and outputs address the specific business needs and use cases the company intends to deploy it for.
Why the other options are wrong
- A. Evaluating performance on benchmark datasets tells you how the model performs on standard tests but doesn't directly measure whether it meets your specific business objectives.
- B. Analyzing architecture and hyperparameters is a technical assessment that doesn't determine whether the model achieves business goals.
- D. Measuring computational resources addresses deployment feasibility but not whether the model actually meets business objectives.
Question 139
A company wants to develop a large language model (LLM) application by using Amazon Bedrock and customer data that is uploaded to Amazon S3. The company's security policy states that each team can access data for only the team's own customers.
Which solution will meet these requirements?
Show answer and explanation
Correct answer: A. Create an Amazon Bedrock custom service role for each team that has access to only the team's customer data.
Creating team-specific Amazon Bedrock service roles with fine-grained permissions to access only each team's customer data in S3 directly enforces the security policy at the IAM level. This ensures each role can only retrieve data for its own customers, with no risk of teams accessing other teams' data.
Why the other options are wrong
- B. Relying on teams to specify customer names is not a technical control and is vulnerable to human error or intentional misuse.
- C. Redacting personal data removes necessary information and doesn't enforce team-based access restrictions; S3 bucket policies don't provide team-level granularity.
- D. Creating one Bedrock role with full S3 access violates the security policy by allowing any team to access all customer data, then relying on IAM team roles doesn't protect data accessed through Bedrock.
Question 140
A customer service team is developing an application to analyze customer feedback and automatically classify the feedback into different categories. The categories include product quality, customer service, and delivery experience.
Which A1 concept does this scenario present?
Show answer and explanation
Correct answer: B. Natural language processing (NLP)
Natural Language Processing (NLP) is the correct AI concept for this scenario. The application analyzes customer feedback text and classifies it into categories (product quality, customer service, delivery experience), which is a core NLP task involving text understanding and classification.
Why the other options are wrong
- A. Computer vision processes images and visual data, not text-based customer feedback.
- C. Recommendation systems suggest items to users; they do not classify feedback into predefined categories.
- D. Fraud detection identifies suspicious transactions or activities, not categorizes customer feedback.
Question 141
A company is developing a new model to predict the prices of specific items. The model performed well on the training dataset.
When the company deployed the model to production, the model's performance decreased significantly.
What should the company do to mitigate this problem?
Show answer and explanation
Correct answer: C. Increase the volume of data that is used in training.
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.
Question 142
An AI practitioner has a database of animal photos. The AI practitioner wants to automatically identify and categorize the animals in the photos without manual human effort.
Which strategy meets these requirements?
Show answer and explanation
Correct answer: A. Object detection
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.
Question 143
An AI practitioner has built a deep learning model to classify the types of materials in images. The AI practitioner now wants to measure the model performance.
Which metric will help the AI practitioner evaluate the performance of the model?
Show answer and explanation
Correct answer: A. Confusion matrix
A confusion matrix is the appropriate metric for evaluating multi-class classification models. It shows true positives, true negatives, false positives, and false negatives across all classes, allowing calculation of precision, recall, F1-score, and overall accuracy for material classification tasks.
Why the other options are wrong
- B. Correlation matrix measures relationships between variables, not classification performance.
- C. R² score is used for regression tasks to measure variance explained, not classification accuracy.
- D. Mean squared error is a regression metric; it's inappropriate for classification tasks.
That was 143 of 571.
The full AWS AIF-C01 pack has all 571 questions, each with the answer, the explanation and why the other options are wrong, plus a questions-only copy for timed runs. US$39, paid once, with free monthly updates and a pass-or-your-money-back guarantee.
Not ready yet? Get the first 285 free as one PDF by email. Get the Half Pack →