AWS AI Practitioner > Applications of Foundation Models > Flashcards

Applications of Foundation Models Flashcards

Analyze how AI technologies are applied across different industries to improve decision-making, automation, and user experiences. (39 cards)

Study These Flashcards
1
Q

A company has fine-tuned a pre-existing model from Amazon Bedrock to enhance document summarization for an internal project. They now want to use the custom model through Amazon Bedrock for production purposes.

What should the company do to enable the use of the custom model?

  1. Allocate Reserved Capacity for the custom model.
  2. Host the custom model on an Amazon SageMaker endpoint for real-time predictions.
  3. List the custom model in the Amazon SageMaker Model Registry.
  4. Enable access to the custom model within Amazon Bedrock.
A

4. Enable access to the custom model within Amazon Bedrock.

Enable access to the custom model within Amazon Bedrock is correct because once the model is fine-tuned, it must be made accessible within Amazon Bedrock to be used in production. Enabling access to the custom model ensures the company can use it for inference tasks like document summarization directly through Bedrock’s infrastructure.

  • Allocate Reserved Capacity for the custom model is incorrect because Amazon Bedrock automatically manages the necessary infrastructure for hosting models. There is no need to allocate reserved capacity manually. Bedrock provides the required scaling and infrastructure management.
  • Host the custom model on an Amazon SageMaker endpoint for real-time predictions is incorrect because Amazon Bedrock provides the hosting and inference capabilities needed to use the custom model. There is no need to deploy it on an Amazon SageMaker endpoint separately.
  • List the custom model in the Amazon SageMaker Model Registry is incorrect because while the Amazon SageMaker Model Registry helps manage model versions and deployments, it is not required when using Amazon Bedrock. Bedrock has its own system for managing and deploying models.

Reference:
Amazon Bedrock

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

A company is building a generative AI application with Amazon Bedrock and wants to understand how much data it can include in a single prompt.

What factor should the company consider?

  1. Temperature setting
  2. Context window size
  3. Maximum inference batch
  4. Model architecture
A

2. Context window size

The context window size defines how much input the model can process in a single prompt. It determines the maximum number of tokens (words or characters) that can be included in the input for the model to process.

  • Temperature setting is incorrect because the temperature setting controls the level of randomness in the model’s output, influencing how deterministic or creative the responses are. It doesn’t affect the amount of data that can be included in a prompt.
  • Maximum inference batch is incorrect because batch size refers to the number of inputs processed together during inference, but it does not influence how much information a single prompt can contain. It relates more to processing efficiency rather than input size.
  • Model architecture is incorrect because model architecture describes the structure and components of the neural network, which affect the model’s performance and capabilities, but not how much data can be provided in one prompt.

Reference:
Amazon Bedrock

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

A media company is building a news summarization tool using Amazon Bedrock to generate concise summaries of trending news articles. The company wants to ensure that the generated content avoids misinformation and inappropriate topics.

Which AWS service or feature will help meet this requirement?

  1. Amazon Rekognition
  2. Amazon Bedrock playgrounds
  3. Guardrails for Amazon Bedrock
  4. Agents for Amazon Bedrock
A

3. Guardrails for Amazon Bedrock

Guardrails can be used to ensure the generated content adheres to certain standards, such as avoiding misinformation, inappropriate topics, or harmful content. In this scenario, guardrails would help ensure that the news summaries are both accurate and appropriate for the audience.

  • Amazon Rekognition is incorrect because Amazon Rekognition is used for image and video analysis, such as facial recognition or object detection, and is not related to controlling the appropriateness of text content.
  • Amazon Bedrock playgrounds is incorrect because Amazon Bedrock playgrounds provide an environment for testing and experimenting with foundation models, but they do not offer specific tools to ensure content appropriateness or control over generated outputs.
  • Agents for Amazon Bedrock is incorrect because Agents for Amazon Bedrock help orchestrate complex multi-step tasks using foundation models, but they do not address the need to filter or restrict the type of content being generated.

Reference:
Amazon Bedrock Guardrails

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

Which feature of Amazon OpenSearch Service enables companies to create applications that use vector-based search functionality?

  1. Seamless integration with Amazon S3 for data storage.
  2. Capabilities for handling location-based queries and geospatial data.
  3. Advanced vector indexing and similarity search for high-dimensional data.
  4. Real-time processing of incoming data streams for immediate insights.
A

3. Advanced vector indexing and similarity search for high-dimensional data.

Amazon OpenSearch Service provides vector search capabilities, which are essential for building vector databases. These features allow companies to search for similar items based on vector representations, often used in applications like recommendation engines and AI-driven search.

  • Seamless integration with Amazon S3 for data storage is incorrect because integration with Amazon S3 is primarily used for storing large volumes of data or backups, not for managing or searching vector data.
  • Capabilities for handling location-based queries and geospatial data is incorrect because geospatial queries are used to manage and query geographic data, not vectors. These features help with map-based applications but don’t support vector search.
  • Real-time processing of incoming data streams for immediate insights is incorrect because real-time analysis is related to processing and analyzing streaming data as it arrives. While useful for certain analytics applications, it doesn’t enable vector-based search functionality.

Reference:
Amazon OpenSearch Service

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

A company needs to visualize the total sales of its best-performing products across different retail outlets over the last year. The company wants an automated solution to create these graphs based on its data.

Which AWS service should the company use?

  1. Amazon QuickSight Q for automated insights
  2. Amazon SageMaker for custom model deployment
  3. Amazon EC2 for data processing and reporting
  4. AWS Glue for data visualization
A

1. Amazon QuickSight Q for automated insights

Amazon QuickSight Q is a natural language query tool that allows users to generate graphs and reports automatically by asking questions in plain language. It’s specifically designed to create visualizations like sales reports without manual effort.

  • Amazon SageMaker for custom model deployment is incorrect because Amazon SageMaker is used to build, train, and deploy machine learning models. It’s not intended for generating sales graphs or creating automated visualizations based on business data.
  • Amazon EC2 for data processing and reporting is incorrect because Amazon EC2 provides compute resources for running virtual machines, but it doesn’t offer tools for data visualization or graph generation. The company would still need a separate service like QuickSight to generate graphs.
  • AWS Glue for data visualization is incorrect because AWS Glue is a data integration service designed for ETL (extract, transform, load) processes. It is not a visualization tool, and it cannot directly generate graphs or reports from data.

Reference:
Amazon QuickSight

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

Which method is used to evaluate the accuracy of a foundation model (FM) applied to image classification tasks?

  1. Calculate the total cost of resources consumed by the model.
  2. Measure the model’s accuracy using a benchmark dataset specifically designed for image classification.
  3. Count the total number of neural network layers in the model architecture.
  4. Assess the color accuracy of the images that the model processes.
A

2. Measure the model’s accuracy using a benchmark dataset specifically designed for image classification.

Evaluating the accuracy of a foundation model involves comparing its predictions to known labels from a predefined dataset. Benchmark datasets are specifically curated for tasks like image classification to assess how well a model performs against a standardized set of images.

  • Calculate the total cost of resources consumed by the model is incorrect because resource consumption measures efficiency, not model accuracy. Evaluating costs is important for financial considerations, but it doesn’t reflect how well the model classifies images.
  • Count the total number of neural network layers in the model architecture is incorrect because the number of layers in the neural network doesn’t directly measure the model’s performance or accuracy. While deeper networks can often be more powerful, accuracy is evaluated through performance on actual data.
  • Assess the color accuracy of the images that the model processes is incorrect because color accuracy pertains to the visual quality or fidelity of an image. It has no direct relationship to the model’s ability to classify images correctly.

Reference:
Image Classification - MXNet

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

A company has customized a foundation model (FM) with Amazon Bedrock to handle customer support queries. The company now wants to test the model’s ability to respond accurately to new types of queries. They need to upload and store a new dataset that Amazon Bedrock can access for the validation process.

Which AWS service should they use for storing this dataset?

  1. Amazon S3
  2. Amazon FSx
  3. Amazon RDS
  4. AWS Snowball Edge
A

1. Amazon S3

Amazon S3 is a scalable, cost-effective object storage service that is widely used for storing large datasets. It is commonly integrated with machine learning services, including Amazon Bedrock, for tasks like training and validating models. S3 allows easy access to datasets needed for model validation.

  • Amazon FSx is incorrect because Amazon FSx is a fully managed file system service designed for specific high-performance workloads like Windows file storage or Lustre for HPC (high-performance computing). It is not typically used for general-purpose dataset storage for model validation.
  • Amazon RDS is incorrect because Amazon RDS is a managed relational database service used for structured data and transactional workloads. While RDS is useful for database-driven applications, it is not suitable for storing large datasets required for model validation in machine learning.
  • AWS Snowball Edge is incorrect because AWS Snowball Edge is a physical device used for transferring large volumes of data in or out of AWS when network transfer is impractical. It is used for data migration, not for directly storing and validating datasets in cloud-based machine learning services.

Reference:
Amazon S3

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

A company is developing an application using Amazon Bedrock. With a limited budget, the company seeks a flexible pricing model that does not require long-term commitments.

Which Amazon Bedrock pricing model is most suitable for this requirement?

  1. On-Demand
  2. Reserved Instances
  3. Pay-per-request
  4. Subscription-based
A

1. On-Demand

The On-Demand pricing model allows the company to pay only for the resources it uses, with no long-term commitments. This offers flexibility and is ideal for companies with a limited budget who want to avoid upfront costs.

  • Reserved Instances is incorrect because reserved instances require a commitment to a long-term contract, which contradicts the company’s need for flexibility and a limited budget.
  • Pay-per-request is incorrect because while pay-per-request may exist for some services, it is not the standard pricing model for Amazon Bedrock. On-Demand is the appropriate model for flexible usage.
  • Subscription-based is incorrect because subscription-based pricing often involves recurring charges, which may not provide the cost control and flexibility the company is looking for.

Reference:
Amazon Bedrock Pricing

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

A media company is using a large language model (LLM) on Amazon Bedrock to summarize movie reviews. The company wants the model to generate concise summaries that capture the overall sentiment of the review.

Which prompt engineering strategy should the company use?

  1. Provide detailed instructions on how LLMs generate summaries and sentiment analysis.
  2. Include a few examples of movie reviews with their corresponding summaries before providing the new review to summarize.
  3. Provide the new review without any context or examples and ask the model to summarize it.
  4. Include instructions for other tasks, such as generating a product description or categorizing the review, along with the summarization task.
A

2. Include a few examples of movie reviews with their corresponding summaries before providing the new review to summarize.

Providing a few examples of reviews with their corresponding summaries helps guide the model by showing it how to perform the task. This is known as few-shot learning and helps the LLM generate more accurate summaries based on the patterns in the provided examples.

  • Provide detailed instructions on how LLMs generate summaries and sentiment analysis is incorrect because giving the model detailed technical explanations doesn’t guide it in generating the specific type of summary required.
  • Provide the new review without any context or examples and ask the model to summarize it is incorrect because without context or examples, the model may not generate consistent or accurate summaries.
  • Include instructions for other tasks, such as generating a product description or categorizing the review, along with the summarization task is incorrect because mixing different tasks in a single prompt can confuse the model and lead to poor-quality outputs for the specific task of summarization.

Reference:
What is Amazon Bedrock?

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

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?

  1. Use few-shot prompting to define how the FM can answer the questions.
  2. Use domain adaptation fine-tuning to adapt the FM to complex scientific terms.
  3. Change the FM inference parameters.
  4. Clean the research paper data to remove complex scientific terms.
A

2. Use domain adaptation fine-tuning to adapt the FM to complex scientific terms.

Lowering the temperature value makes the model’s output more deterministic and consistent. A lower temperature reduces randomness in the generated responses, ensuring the same input yields more similar outputs.

  • Raise the temperature value is incorrect because increasing the temperature introduces more randomness, making the model’s responses less predictable and more varied.
  • Shorten the output token limit is incorrect because reducing the token limit affects the length of the output, not the consistency of the responses. It limits how much text the model can generate but does not control randomness.
  • Extend the maximum sequence length is incorrect because increasing the sequence length allows the model to produce longer outputs but does not directly affect how consistent or varied the responses are.

Reference:
Inference Request Parameters and Response Fields for Foundation Models

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

An education platform is developing a chatbot to help students with homework questions. The company has selected a foundation model (FM) but wants the chatbot’s responses to maintain an encouraging and educational tone.

What should the company do to achieve this?

  1. Limit the token output to control the length of responses.
  2. Refine the prompt to ensure the FM produces responses in the desired tone.
  3. Use batch inference to process multiple student queries at once.
  4. Increase the temperature to make responses more dynamic.
A

2. Refine the prompt to ensure the FM produces responses in the desired tone.

Adjusting and refining the prompt can help guide the model to respond in a consistent, encouraging tone, which aligns with the company’s goals for the chatbot.

  • Limit the token output to control the length of responses is incorrect because limiting tokens affects response length, not the tone or style.
  • Use batch inference to process multiple student queries at once is incorrect because batch inference is used for handling large-scale data and doesn’t control the chatbot’s tone.
  • Increase the temperature to make responses more dynamic is incorrect because increasing temperature makes responses more random and creative, which could make them inconsistent with the desired tone.

Reference:
Prompt Engineering Concepts

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

A fintech company is building a generative AI application using Amazon Bedrock. The company wants to assess the costs related to generating inferences with a large language model (LLM).

Which factor will influence the inference costs?

  1. Number of tokens processed
  2. Model accuracy rate
  3. Size of training dataset
  4. Total memory used for training
A

1. Number of tokens processed

Inference costs for large language models on Amazon Bedrock are driven by the number of tokens processed. Tokens represent pieces of the input or output text, and the more tokens involved in a single inference request, the higher the cost. Monitoring token usage helps control and manage costs effectively.

  • Model accuracy rate is incorrect because the accuracy of the model does not directly influence the cost of inference, which is based on the number of tokens processed during requests.
  • Size of training dataset is incorrect because the cost of inference is unrelated to the size of the training dataset. It’s the inference stage that affects token usage and costs.
  • Total memory used for training is incorrect because memory usage during training impacts training costs, not inference costs. Inference costs are driven by token processing during real-time requests.

Reference:
Amazon Bedrock Pricing

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

A marketing agency needs to select a model from Amazon Bedrock that will be used internally to generate campaign slogans and advertisements. The agency must find a model that produces content in a tone and style that aligns with the agency’s creative standards.

What should the agency do to meet these requirements?

  1. Evaluate the models using built-in datasets for prompt evaluation.
  2. Evaluate the models by testing custom prompts and collecting feedback from the agency’s creative team.
  3. Use popular model benchmarks and rankings to identify the best model.
  4. Analyze the model InvocationLatency runtime metrics in Amazon CloudWatch to assess response times.
A

2. Evaluate the models by testing custom prompts and collecting feedback from the agency’s creative team.

The best way to find a model that fits the agency’s style and tone is to use custom prompts that reflect real-world use cases and collect feedback from employees who are familiar with the company’s preferences. This ensures that the chosen model aligns with internal creative standards.

  • Evaluate the models using built-in datasets for prompt evaluation is incorrect because built-in datasets may not reflect the company’s specific needs or style preferences. Custom prompts are more effective for evaluating the model in a real-world context.
  • Use popular model benchmarks and rankings to identify the best model is incorrect because public leaderboards often rank models based on general performance metrics, which may not reflect how well a model fits the company’s specific style requirements.
  • Analyze the model InvocationLatency runtime metrics in Amazon CloudWatch to assess response times is incorrect because InvocationLatency measures the speed of a model’s responses, not the quality or style of its outputs.

Reference:
Prompt Engineering Concepts

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

A healthcare company is training a foundation model (FM) to analyze medical records. The company wants to improve the model’s accuracy until it reaches a specific threshold for acceptable performance.

Which solution will help the company achieve this?

  1. Decrease the batch size.
  2. Increase the epochs.
  3. Lower the learning rate.
  4. Increase the dropout rate.
A

2. Increase the epochs.

Increasing the epochs allows the model to continue learning by passing through the dataset multiple times, which can improve the model’s accuracy. The company can adjust the number of epochs to achieve the desired performance, ensuring the model trains until it meets the accuracy threshold.

  • Decrease the batch size is incorrect because reducing the batch size affects the training dynamics but does not directly improve accuracy to a specific level. It may instead slow down training.
  • Lower the learning rate is incorrect because decreasing the learning rate helps in fine-tuning but may result in longer training times without necessarily improving accuracy beyond a certain point.
  • Increase the dropout rate is incorrect because increasing the dropout rate prevents overfitting, but an overly high dropout rate can reduce the model’s ability to learn, negatively impacting accuracy.

Reference:
Hyperparameters for Optimizing the Learning Process of your Text Generation Models

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

A company wants to develop a chat interface using large language models (LLMs) with Amazon Bedrock to help users navigate the company’s product manuals, which are stored as PDF files. The company needs a cost-effective solution to provide relevant answers from the manuals.

Which solution meets these requirements most cost-effectively?

  1. Use prompt engineering to add one relevant PDF file as context to the user prompt when submitted to Amazon Bedrock.
  2. Use prompt engineering to add all PDF files as context to every user prompt submitted to Amazon Bedrock.
  3. Fine-tune a model with Amazon Bedrock using all the PDF files and process user prompts with the fine-tuned model.
  4. Upload PDF documents to an Amazon Bedrock knowledge base and use the knowledge base to provide context when users submit prompts.
A

4. Upload PDF documents to an Amazon Bedrock knowledge base and use the knowledge base to provide context when users submit prompts.

Uploading PDF documents to an Amazon Bedrock knowledge base allows the company to store the manuals and provide relevant context dynamically based on user prompts. This solution is cost-effective because the model does not need to be fine-tuned or retrained, and it leverages an existing knowledge base to serve the necessary information without repeatedly submitting large amounts of data.

  • Use prompt engineering to add one relevant PDF file as context to the user prompt when submitted to Amazon Bedrock is incorrect because adding a single document as context may not provide sufficient information for accurate responses when the user’s query spans multiple manuals.
  • Use prompt engineering to add all PDF files as context to every user prompt submitted to Amazon Bedrock is incorrect because adding all the PDF files as context to each prompt would be expensive in terms of token usage and processing, making it a less cost-effective solution.
  • Fine-tune a model with Amazon Bedrock using all the PDF files and process user prompts with the fine-tuned model is incorrect because fine-tuning a model with a large dataset like all the product manuals would be costly and unnecessary when a knowledge base can serve the same purpose more efficiently.

Reference:
Prompt Engineering Concepts

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

A healthcare company’s AI development team wants to quickly deploy and consume a foundation model (FM) within their VPC to process medical data securely.

Which AWS service or feature will help them achieve this?

  1. Amazon Comprehend Medical
  2. AWS Fargate
  3. Amazon SageMaker endpoints
  4. AWS Lambda
Study These Flashcards
A

3. Amazon SageMaker endpoints

Amazon SageMaker endpoints allow the team to securely deploy machine learning models, including foundation models, within their VPC. This ensures the data stays within the VPC while enabling real-time access to the deployed models for secure processing of medical data.

  • Amazon Comprehend Medical is incorrect because Amazon Comprehend Medical is a service designed to extract and analyze medical information from text, not to deploy and host foundation models within a VPC.
  • AWS Fargate is incorrect because AWS Fargate is a serverless compute engine for containerized applications and is not specifically designed for deploying or hosting machine learning models.
  • AWS Lambda is incorrect because AWS Lambda is a serverless compute service for running functions in response to events. It’s not designed for hosting and deploying foundation models within a VPC.

Reference:
Real-time Inference

17
Q

A global transportation company receives thousands of requests daily from customers seeking updates on package deliveries. To manage the volume, the company wants to deploy Agents for Amazon Bedrock to streamline responses and automate workflows.

What are the key benefits of using Amazon Bedrock agents that could assist the transportation company?

  1. Simplification of data visualization for customer tracking insights
  2. Automation of routine inquiries and coordination of multi-step processes
  3. Generation of marketing analytics to predict future package deliveries
  4. Enhancement of email campaigns for customer engagement based on delivery data
Study These Flashcards
A

2. Automation of routine inquiries and coordination of multi-step processes

Amazon Bedrock agents automate repetitive tasks like responding to common customer inquiries, such as package status, and orchestrate complex workflows, including escalating unresolved issues or initiating specific processes across different systems. This improves efficiency in handling large volumes of customer service requests.

  • Simplification of data visualization for customer tracking insights is incorrect because Bedrock agents focus on automating tasks and workflows, not on creating visualizations or dashboards for tracking insights.
  • Generation of marketing analytics to predict future package deliveries is incorrect because Bedrock agents are not used for marketing analytics or predictive delivery models. They are designed for workflow automation and task handling.
  • Enhancement of email campaigns for customer engagement based on delivery data is incorrect because Bedrock agents are not designed for email campaign management or customer engagement tools.

Reference:
Amazon Bedrock Agents

18
Q

A financial services company is using few-shot prompting on a base model hosted on Amazon Bedrock to generate daily reports. The model currently uses 10 examples in each prompt and performs well. However, the company wants to reduce monthly operational costs.

Which solution will meet these requirements?

  1. Customize the model through fine-tuning.
  2. Decrease the number of tokens in the prompt.
  3. Increase the number of examples used in the prompt.
  4. Deploy the model on a dedicated server instance.
Study These Flashcards
A

2. Decrease the number of tokens in the prompt.

Decreasing the number of tokens in the prompt will directly reduce the amount of input the model processes during each invocation, leading to lower inference costs. Fewer tokens in each prompt mean the company can maintain performance while cutting costs.

  • Customize the model through fine-tuning is incorrect because fine-tuning adds complexity and cost. It does not directly help in reducing inference costs based on token usage.
  • Increase the number of examples used in the prompt is incorrect because increasing the number of examples would result in more tokens being processed, which would raise costs rather than lowering them.
  • Deploy the model on a dedicated server instance is incorrect because running the model on a dedicated server may not necessarily reduce costs. It can, in fact, increase operational expenses depending on usage patterns.

Reference:
Design a Prompt

19
Q

A museum is developing an AI-powered virtual tour guide to explain historical artifacts to visitors. The AI needs to adjust its language and tone depending on the visitor’s background, such as children, history enthusiasts, or academic researchers. The visitor’s background will be provided to the model when they ask questions.

Which solution meets these requirements with the least implementation effort?

  1. Fine-tune the model by adding specialized datasets for each visitor group, such as children and researchers.
  2. Include a description of the visitor’s background in the prompt to instruct the model on how to adjust its response.
  3. Use a separate machine learning model to analyze and transform the model’s output for different user types.
  4. Implement a multi-step dialogue process where the model asks follow-up questions to adjust its response style based on the visitor’s feedback.
Study These Flashcards
A

2. Include a description of the visitor’s background in the prompt to instruct the model on how to adjust its response.

Including a description of the visitor’s background in the prompt is the most efficient solution, requiring minimal implementation effort. By simply adjusting the prompt, the model can tailor its responses to match the visitor’s background, whether it’s a child or a researcher, without the need for additional training or complex workflows.

  • Fine-tune the model by adding specialized datasets for each visitor group, such as children and researchers is incorrect because fine-tuning the model with additional datasets requires significant effort and is more resource-intensive than prompt engineering.
  • Use a separate machine learning model to analyze and transform the model’s output for different user types is incorrect because using an additional machine learning model introduces unnecessary complexity when prompt adjustments can achieve the desired result more simply.
  • Implement a multi-step dialogue process where the model asks follow-up questions to adjust its response style based on the visitor’s feedback is incorrect because a multi-step dialogue process increases complexity and implementation effort, which is not necessary for simply adjusting the response style based on the provided background.

Reference:
Prompt Engineering Concepts

20
Q

An ecommerce company wants to develop a solution that analyzes customer reviews of products to determine customer sentiments based on the text.

Which AWS services meet these requirements?

(Select TWO.)

  1. Amazon Lex
  2. Amazon Comprehend
  3. Amazon Polly
  4. Amazon Transcribe
  5. Amazon Bedrock
Study These Flashcards
A

2. Amazon Comprehend
5. Amazon Bedrock

Amazon Comprehend is a natural language processing (NLP) service that can analyze customer reviews to detect sentiment, including positive, negative, neutral, or mixed. It’s ideal for automatically analyzing written text to determine customer opinions.

Amazon Bedrock offers foundation models for various AI tasks, including text analysis and sentiment detection. It provides pre-trained models that can be fine-tuned for tasks like sentiment analysis based on customer reviews.

  • Amazon Lex is incorrect because Amazon Lex is designed for building conversational interfaces like chatbots, not for analyzing sentiment in written reviews.
  • Amazon Polly is incorrect because Amazon Polly converts text to speech, which is unrelated to sentiment analysis from written customer reviews.
  • Amazon Transcribe is incorrect because Amazon Transcribe converts speech to text, which is not relevant for analyzing already written customer reviews.

References:

21
Q

Which metric measures the runtime efficiency of operating AI models?

  1. Model accuracy
  2. Training time per epoch
  3. Average response time
  4. Disk space used for model storage
Study These Flashcards
A

3. Average response time

Average response time measures how quickly an AI model provides predictions or outputs in real-time, making it an important metric for assessing runtime efficiency.

  • Model accuracy is incorrect because accuracy measures how well a model performs, not how fast it operates.
  • Training time per epoch is incorrect because training time per epoch measures the duration of training cycles, not how efficiently the model runs during inference.
  • Disk space used for model storage is incorrect because disk space refers to the model’s size, not its runtime performance.

22
Q

A company is building a large language model (LLM) chatbot to help answer customer questions more efficiently. The company wants to reduce the number of actions that call center employees need to take to assist customers.

Which business objective should the company use to evaluate the effect of the LLM chatbot?

  1. Website engagement rate
  2. Average call duration
  3. Employee satisfaction score
  4. Product innovation rate
Study These Flashcards
A

2. Average call duration

Average call duration measures how long call center employees spend on each customer interaction. If the LLM chatbot helps answer questions efficiently, the time employees spend on calls should decrease, making this the most relevant metric to track.

  • Website engagement rate is incorrect because website engagement measures user interaction with the site, not the efficiency of call center operations.
  • Employee satisfaction score is incorrect because while employee satisfaction is important, it does not directly measure the chatbot’s impact on reducing actions in customer service calls.
  • Product innovation rate is incorrect because product innovation tracks new product development and is not related to call center performance or chatbot effectiveness.

Reference:
Unleash AI to Transform Every Customer Interaction

23
Q

A manufacturing company is building a solution to generate designs for protective eyewear. The company needs the solution to have high accuracy and minimize the risk of errors in the design annotations.

Which solution will meet these requirements?

  1. Data augmentation by using an Amazon Bedrock knowledge base
  2. Human-in-the-loop validation by using Amazon SageMaker Ground Truth Plus
  3. Image recognition by using Amazon Rekognition
  4. Data analytics by using AWS Glue
Study These Flashcards
A

2. Human-in-the-loop validation by using Amazon SageMaker Ground Truth Plus

Human-in-the-loop validation through Amazon SageMaker Ground Truth Plus allows human reviewers to validate and correct any annotations made by the model. This ensures high accuracy and minimizes errors in generating designs for protective eyewear.

  • Data augmentation by using an Amazon Bedrock knowledge base is incorrect because data augmentation involves creating variations of data to enhance training, but it does not focus on validating and correcting errors in annotations.
  • Image recognition by using Amazon Rekognition is incorrect because Amazon Rekognition is primarily used for identifying objects in images, not for generating designs or verifying their accuracy.
  • Data analytics by using AWS Glue is incorrect because AWS Glue is used for data processing and ETL tasks, not for image design or validation.

Reference:
Amazon SageMaker Ground Truth

24
Q

An AI engineer is developing a search tool that needs to process user queries containing both text and images.

Which type of foundation model (FM) should the engineer use to meet this requirement?

  1. Multi-modal embedding model
  2. Text classification model
  3. Image recognition model
  4. Language translation model
Study These Flashcards
A

1. Multi-modal embedding model

A multi-modal embedding model can handle and create unified representations of both text and image data. This makes it ideal for search applications that need to process queries containing both formats and return relevant results.

  • Text classification model is incorrect because text classification models are designed to categorize text data and cannot process images.
  • Image recognition model is incorrect because image recognition models focus on identifying objects within images but cannot process or link text inputs with images.
  • Language translation model is incorrect because language translation models are designed to convert text from one language to another, not to process or embed both text and image data.

Reference:
Cost-effective Document Classification Using the Amazon Titan Multimodal Embeddings Model

25
A company is developing a chatbot to enhance user experience by using a large language model (LLM) from Amazon Bedrock for intent detection. The company wants to use few-shot learning to improve the accuracy of intent detection. **Which additional data does the company need to meet these requirements?** 1. Pairs of chatbot responses and correct user intents 2. Pairs of user messages and correct chatbot responses 3. Pairs of user messages and correct user intents 4. Pairs of user intents and correct chatbot responses
**3.** Pairs of user messages and correct user intents ## Footnote To improve intent detection with few-shot learning, the company needs pairs of user messages and correct user intents. These pairs allow the LLM to learn from a small number of examples and map user messages to their corresponding intents more accurately. * Pairs of chatbot responses and correct user intents is incorrect because the chatbot’s responses are not relevant for intent detection accuracy. The focus should be on understanding the user’s message and linking it to the correct intent. * Pairs of user messages and correct chatbot responses is incorrect because chatbot responses do not help with detecting user intent. The model needs to understand user messages and map them to the right intents. * Pairs of user intents and correct chatbot responses is incorrect because the pairing of user intents and chatbot responses is more relevant for generating responses, not for improving intent detection accuracy. **Reference:** [Amazon Bedrock](https://aws.amazon.com/bedrock/)
26
Which option is a benefit of ongoing pre-training when fine-tuning a foundation model (FM)? 1. Helps decrease the model's complexity 2. Improves model performance over time 3. Reduces hardware resource requirements 4. Simplifies the model’s architecture
**2.** Improves model performance over time ## Footnote Ongoing pre-training helps the model learn from additional data, which allows it to continually improve its performance when fine-tuning for specific tasks. This approach enables the model to adapt to new patterns and trends, enhancing accuracy and effectiveness. * Helps decrease the model's complexity is incorrect because ongoing pre-training typically increases the model’s ability to capture more complex patterns, not decrease its complexity. * Reduces hardware resource requirements is incorrect because ongoing pre-training usually requires more computational resources, not less. * Simplifies the model’s architecture is incorrect because pre-training doesn’t necessarily simplify the architecture but improves the model’s ability to handle more diverse tasks. **Reference:** [What are Foundation Models?](https://aws.amazon.com/what-is/foundation-models/)
27
A company needs to build a system that can process and analyze large amounts of real-time log data to generate immediate insights. **Which feature of Amazon OpenSearch Service will help the company meet these requirements?** 1. Seamless integration with Amazon S3 for long-term storage of log data. 2. Capabilities for handling location-based queries and geospatial data. 3. Real-time ingestion and analysis of streaming data through integrations with services like Amazon Kinesis. 4. Advanced vector indexing for similarity searches across high-dimensional data.
**3.** Real-time ingestion and analysis of streaming data through integrations with services like Amazon Kinesis. ## Footnote Amazon OpenSearch Service integrates with services like Amazon Kinesis to enable real-time ingestion, processing, and analysis of streaming data. This allows companies to derive immediate insights from log data and other real-time events as they occur. * Seamless integration with Amazon S3 for long-term storage of log data is incorrect because S3 integration is more for storage, not real-time log analysis. * Capabilities for handling location-based queries and geospatial data is incorrect because location-based queries focus on geospatial data, not real-time log analysis. * Advanced vector indexing for similarity searches across high-dimensional data is incorrect because vector indexing is used for similarity searches, not real-time data ingestion and analysis. **Reference:** [Amazon OpenSearch Service](https://aws.amazon.com/opensearch-service/)
28
A retail company is using an Amazon Bedrock foundation model (FM) to generate product descriptions. The company wants to ensure that the model's outputs reflect its specific product catalog and internal data sources. **What solution should the company implement to meet this requirement?** 1. Fine-tune the foundation model with the company's product data. 2. Lower the temperature to make the responses more consistent. 3. Use a different foundation model that specializes in retail. 4. Increase the token limit to capture more detailed product descriptions.
**1.** Fine-tune the foundation model with the company's product data. ## Footnote By fine-tuning the foundation model with specific internal product data, the company can improve the relevance of the generated product descriptions, ensuring that the outputs align closely with its unique catalog and offerings. * Lower the temperature to make the responses more consistent is incorrect because lowering the temperature reduces randomness, but it does not incorporate the company’s internal data into the model. * Use a different foundation model that specializes in retail is incorrect because switching to another model won’t necessarily improve alignment with the company’s specific product data without fine-tuning. * Increase the token limit to capture more detailed product descriptions is incorrect because increasing the token limit allows for longer outputs, but it won’t integrate the company’s product data into the model’s responses. **Reference:** [Customize your Model to Improve its Performance for your Use Case](https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html)
29
A logistics company has collected sensor data from thousands of vehicles, with most of the data unlabeled, but a small portion of it has been labeled indicating vehicle failures. The company wants to predict vehicle maintenance needs by leveraging both the labeled and unlabeled data to make better predictions. **Which machine learning approach should the company use to achieve this?** 1. Semi-supervised learning 2. Unsupervised learning 3. Supervised learning 4. Transfer learning
**1.** Semi-supervised learning ## Footnote Semi-supervised learning is ideal when a dataset contains a small amount of labeled data and a large amount of unlabeled data. The company can use the labeled data on vehicle failures to train the model while utilizing the vast amounts of unlabeled sensor data to improve the model’s accuracy in predicting vehicle maintenance needs. * Unsupervised learning is incorrect because unsupervised learning works entirely with unlabeled data and would not leverage the labeled vehicle failure data the company possesses. * Supervised learning is incorrect because supervised learning requires fully labeled data, and the company only has a small portion of labeled data, making semi-supervised learning a better fit. * Transfer learning is incorrect because transfer learning involves applying a pre-trained model from one domain to another, which is not the focus here. **Reference:** [What’s the Difference Between Supervised and Unsupervised Learning?](https://aws.amazon.com/compare/the-difference-between-machine-learning-supervised-and-unsupervised/)
30
A company wants to increase developer productivity and streamline the software development process by leveraging generative AI. The company plans to use Amazon Q Developer to enhance its development workflows. **Which feature of Amazon Q Developer will help meet these goals?** 1. Provide real-time suggestions for code refactoring to improve efficiency. 2. Offer real-time serverless deployment to reduce infrastructure management overhead. 3. Enable voice-activated coding and natural language search for easier code navigation. 4. Analyze software performance and automatically suggest improvements to boost application speed.
**3.** Enable voice-activated coding and natural language search for easier code navigation. ## Footnote Amazon Q Developer allows developers to leverage generative AI through features like voice-activated coding and natural language search. These capabilities help developers navigate codebases more efficiently and reduce the time spent manually searching or writing code, improving productivity. * Provide real-time suggestions for code refactoring to improve efficiency is incorrect because Amazon Q Developer does not focus on automated code refactoring; its primary function is assisting with code generation and search. * Offer real-time serverless deployment to reduce infrastructure management overhead is incorrect because serverless deployment is not the primary functionality of Amazon Q Developer, but rather a feature of services like AWS Lambda. * Analyze software performance and automatically suggest improvements to boost application speed is incorrect because Amazon Q Developer does not analyze performance; it focuses on improving the coding process with generative AI features like voice commands and code search. **Reference:** [What is Amazon Q Developer?](https://docs.aws.amazon.com/amazonq/latest/qdeveloper-ug/what-is.html)
31
A financial services company is training a foundation model (FM) to predict credit risk. The company wants to prevent overfitting and improve the model's generalization to unseen data. **Which solution will help the company achieve this?** 1. Increase the learning rate. 2. Apply regularization techniques. 3. Decrease the dropout rate. 4. Reduce the number of training examples.
**2.** Apply regularization techniques. ## Footnote Regularization techniques (such as L2 regularization) help reduce overfitting by penalizing overly complex models. This encourages the model to generalize better to unseen data, improving its performance on real-world predictions, such as credit risk assessment. * Increase the learning rate is incorrect because increasing the learning rate might lead to faster convergence but can cause the model to miss key patterns, leading to suboptimal performance. * Decrease the dropout rate is incorrect because reducing the dropout rate could worsen overfitting, as dropout helps prevent overfitting by randomly dropping units during training. * Reduce the number of training examples is incorrect because reducing the training dataset size would likely lead to worse generalization, not better performance. **Reference:** [What is Overfitting?](https://aws.amazon.com/what-is/overfitting/)
32
A gaming company is developing a machine learning model to predict player behavior and improve in-game recommendations. The model is performing well but has not yet reached its desired accuracy. The company wants to fine-tune the model to optimize performance without changing the underlying dataset. **Which process should the company use to achieve this?** 1. Feature engineering 2. Data augmentation 3. Hyperparameter tuning 4. Model deployment
**3.** Hyperparameter tuning ## Footnote Hyperparameter tuning is the process of optimizing the model’s hyperparameters, such as learning rate, batch size, and number of layers, to improve the model's accuracy and performance. It helps the company fine-tune the model without altering the dataset or its structure. * Feature engineering is incorrect because feature engineering involves modifying or creating new input variables from the dataset, not adjusting the model's hyperparameters. * Data augmentation is incorrect because data augmentation refers to generating additional data by modifying the existing dataset, which doesn't directly optimize model performance. * Model deployment is incorrect because model deployment refers to making the trained model available for use, not improving its accuracy. **Reference:** [Understand the Hyperparameter Tuning Strategies Available in Amazon SageMaker AI ](https://docs.aws.amazon.com/sagemaker/latest/dg/automatic-model-tuning-how-it-works.html)
33
A company wants to experiment with various foundation models (FMs) to understand how they respond to different prompts before integrating a model into their production environment. The company needs a tool that allows them to quickly test and refine prompts without managing any infrastructure. **Which AWS service should the company use?** 1. Amazon SageMaker Studio 2. Amazon Bedrock Playground 3. Amazon Comprehend 4. AWS Glue
**2.** Amazon Bedrock Playground ## Footnote Amazon Bedrock Playground provides a platform for experimenting with different foundation models and refining prompts without needing to manage infrastructure. It allows companies to test model behavior before deploying them in production environments. * Amazon SageMaker Studio is incorrect because SageMaker Studio is a comprehensive machine learning environment but is more focused on building and deploying custom ML models, not specifically experimenting with foundation models through simple prompts. * Amazon Comprehend is incorrect because Amazon Comprehend is used for natural language processing tasks like sentiment analysis and entity recognition, not for experimenting with foundation models. * AWS Glue is incorrect because AWS Glue is an ETL service used for data preparation and transformation, not for testing and refining AI models. **Reference:** [Generate Responses in the Console Using Playgrounds](https://docs.aws.amazon.com/bedrock/latest/userguide/playgrounds.html)
34
A company wants to improve its customer support chatbot by providing accurate, up-to-date information from a large internal knowledge base. The company is considering a solution that combines document retrieval with generative AI to create responses based on both the knowledge base and the chatbot's model. **Which AI technique should the company implement to achieve this?** 1. Reinforcement learning 2. Retrieval Augmented Generation (RAG) 3. Transfer learning 4. Data augmentation
**2.** Retrieval Augmented Generation (RAG) ## Footnote Retrieval Augmented Generation (RAG) enhances generative AI models by retrieving relevant documents or knowledge base entries and using that information to generate accurate and contextually relevant responses. This makes it ideal for customer support scenarios where up-to-date, factual information is critical. * Reinforcement learning is incorrect because reinforcement learning focuses on decision-making based on rewards, not combining document retrieval with generative AI. * Transfer learning is incorrect because transfer learning involves applying knowledge from one task to another, not retrieving documents to assist in generating responses. * Data augmentation is incorrect because data augmentation increases the size of the training dataset by modifying the existing data, not by retrieving external knowledge sources to improve generative responses. **Reference:** [Retrieve Data and Generate AI Responses with Amazon Bedrock Knowledge Bases](https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html)
35
A company is developing a chatbot and wants it to answer complex customer queries by breaking down the problem into intermediate reasoning steps. **Which prompt engineering technique should the company use to achieve this?** 1. Zero-shot learning 2. Few-shot learning 3. Single-shot prompting 4. Chain-of-thought prompting
**4.** Chain-of-thought prompting ## Footnote Chain-of-thought prompting helps the model break down complex questions into intermediate steps, improving its reasoning and accuracy in answering multi-step problems. This technique guides the model through a series of thought processes to arrive at the final answer. * Zero-shot learning is incorrect because zero-shot learning involves making predictions without any prior examples, but it doesn’t help break down complex queries into reasoning steps. * Few-shot learning is incorrect because few-shot learning refers to learning from a few examples, but it doesn't involve breaking down complex reasoning into steps. * Single-shot prompting is incorrect because single-shot prompting involves providing only one example for the model, which doesn't focus on solving multi-step problems. **Reference:** [Prompt Engineering Concepts](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-engineering-guidelines.html)
36
A company is developing a product recommendation system using a large language model (LLM). The company wants the model to generate accurate responses without providing any examples in the prompt. **Which prompt engineering technique should the company use?** 1. Chain-of-thought prompting 2. Few-shot learning 3. Zero-shot learning 4. Prompt templates
**3.** Zero-shot learning ## Footnote Zero-shot learning allows the model to generate responses without any prior examples in the prompt. This technique is useful when the model needs to make predictions or generate outputs based on its general knowledge without being provided specific examples. * Chain-of-thought prompting is incorrect because chain-of-thought prompting involves breaking down reasoning into steps, which is not the goal in this case. * Few-shot learning is incorrect because few-shot learning involves giving the model a small number of examples to guide its predictions, while zero-shot learning does not use any examples. * Prompt templates is incorrect because prompt templates provide a structured format for generating outputs, but they don't specifically focus on zero-shot scenarios. **Reference:** [Prompt Engineering Concepts](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-engineering-guidelines.html)
37
A startup is developing a language translation model but wants to save time and resources by leveraging an existing pre-trained model rather than training one from scratch. The company plans to fine-tune the pre-trained model for its specific use case. **Which machine learning technique should the company use?** 1. Transfer learning 2. Unsupervised learning 3. Reinforcement learning 4. Zero-shot learning
**1.** Transfer learning ## Footnote Transfer learning involves using a pre-trained model and fine-tuning it for a specific task. This approach saves time and computational resources, allowing the company to adapt an existing model for language translation without starting from scratch. * Unsupervised learning is incorrect because unsupervised learning finds patterns in data without labeled examples, and it doesn't involve leveraging pre-trained models. * Reinforcement learning is incorrect because reinforcement learning focuses on decision-making through rewards and penalties, not reusing pre-trained models. * Zero-shot learning is incorrect because zero-shot learning refers to making predictions without having seen any examples of a particular task, not fine-tuning a pre-trained model. **Reference:** [What is Transfer Learning?](https://aws.amazon.com/what-is/transfer-learning/)
38
A company is using a foundation model to develop a custom AI application. They want to adapt the model to their specific use case by training it with additional data that reflects their business needs. **Which approach should the company use to adjust the foundation model for their use case?** 1. Pre-training 2. Zero-shot learning 3. Fine-tuning 4. Data augmentation
**3.** Fine-tuning ## Footnote Fine-tuning is the process of adapting a pre-trained foundation model to a specific use case by training it on additional data that is relevant to the task. This allows the company to customize the foundation model without training from scratch. * Pre-training is incorrect because pre-training refers to the initial phase where the model is trained on large datasets, not the customization of an already pre-trained model. * Zero-shot learning is incorrect because zero-shot learning involves making predictions without training on specific examples, not fine-tuning a model. * Data augmentation is incorrect because data augmentation involves expanding the training dataset, not adjusting the model to a specific use case. **Reference:** [Fine-Tune a Model](https://docs.aws.amazon.com/sagemaker/latest/dg/jumpstart-fine-tune.html)
39
A company is using a foundation model to classify text into categories. The company wants the model to classify new categories without needing to train the model on specific labeled examples for each new category. **Which approach should the company use?** 1. Pre-training 2. Fine-tuning 3. Zero-shot learning 4. Semi-supervised learning
**3.** Zero-shot learning ## Footnote Zero-shot learning allows the model to classify or perform tasks for new categories or tasks without having been explicitly trained on examples for those categories. The model relies on general knowledge to make predictions in cases where labeled examples are unavailable. * Pre-training is incorrect because pre-training refers to the initial large-scale training phase of the model, not predicting new categories without labeled examples. * Fine-tuning is incorrect because fine-tuning requires additional labeled data to adjust the model for specific tasks, whereas zero-shot learning does not require this. * Semi-supervised learning is incorrect because semi-supervised learning involves using a small amount of labeled data and a large amount of unlabeled data, which is not the same as predicting categories without any specific examples. **Reference:** [Amazon SageMaker Model Training](https://aws.amazon.com/sagemaker/train/)
AWS AI Practitioner
flashcards
Decks in class (5)
# Cards
Brainscape helps you reach your goals faster, through stronger study habits.
© 2026 Bold Learning Solutions. Terms and Conditions