Amazon Practice Questions, Discussions & Exam Topics by our Authors
A company is building an application that needs to generate synthetic data that is based on existing data.Which type of...
The company is building an application that needs to generate synthetic data based on existing data. Let's evaluate each option in the context of this requirement:
Option A: Generative adversarial network (GAN)
- Purpose: A Generative Adversarial Network (GAN) is a type of machine learning model that is specifically designed for generating synthetic data that closely resembles real data. GANs work by training two neural networks—the generator and the discriminator—in a game-like setting, where the generator creates data and the discriminator evaluates it, guiding the generator to improve.
- Reason for selection: GANs are the most suitable choice for generating synthetic data, as their core purpose is to learn the underlying distribution of the existing data and generate new, similar data. They are widely used in applications like image generation, data augmentation, and more. GANs provide the flexibility to generate high-quality synthetic data, which directly meets the company's requirement of creating data based on existing datasets.
- Effort, Time, Cost: GANs require significant training time and computational resources, but they are the most effective method for generating realistic synthetic data. The initial setup might involve substantial effort, but once trained, they can be highly efficient for generating large datasets with minimal manual intervention.
Option B: XGBoost
- Purpose: XGBoost is a popular machine learning algorithm primarily used for supervised learning tasks, such as classification and regression. It works by combining decision trees in an ensemble method to make predictions. XGBoost is not designed for generating synthetic data; instead, it excels at making predictions based on labeled datasets.
- Reason for rejection: XGBoost is no...
Author: Amira · Last updated Jul 19, 2026
A digital devices company wants to predict customer demand for memory hardware. The company does not have coding experience or knowledge of ML algorithms and needs to develop a data-driven predictive model. The company needs t...
The company needs to predict customer demand for memory hardware and doesn't have coding experience or knowledge of machine learning (ML) algorithms. This suggests that the solution should be user-friendly and allow the company to easily analyze both internal and external data without requiring coding expertise.
Let’s evaluate each option:
A) Store the data in Amazon S3. Create ML models and demand forecast predictions by using Amazon SageMaker built-in algorithms that use the data from Amazon S3.
- Rejection Reasoning: While storing the data in Amazon S3 and using SageMaker built-in algorithms is a valid approach, it requires a certain level of expertise in machine learning and coding to prepare the data, choose the right algorithms, and interpret the results. Since the company does not have coding experience, this option may not be the best fit for them as it would likely require more technical effort.
B) Import the data into Amazon SageMaker Data Wrangler. Create ML models and demand forecast predictions by using SageMaker built-in algorithms.
- Rejection Reasoning: Amazon SageMaker Data Wrangler is a powerful tool for data preparation, and it simplifies the process of importing and transforming data. However, building models and making predictions with SageMaker built-in algorithms still involves some complexity in terms of understanding the algorithms and model deployment. Although Data Wrangler helps with data preparation, the overall solution may still require more technical know-how than the company possesses.
C) Import the data into Amazon SageMaker Data Wrangler. Build ML models and demand forecast predictions by using an Amazon Pers...
Author: Sophia Clark · Last updated Jul 19, 2026
A company has installed a security camera. The company uses an ML model to evaluate the security camera footage for potential thefts. The company has discovered that the model disproportionately flags people wh...
To address the type of bias affecting the model’s output in this scenario, let's analyze the options based on the description that the model disproportionately flags people from a specific ethnic group. The goal is to identify the type of bias that directly leads to this situation. Each option will be examined in the context of the problem.
Option A: Measurement Bias
- Explanation: Measurement bias occurs when there is an error in how the data is measured or collected, leading to incorrect or skewed results. This could happen if, for example, the camera's resolution, angle, or environmental conditions distort the features of individuals in a way that disproportionately affects certain groups.
- Relevance: While measurement bias can cause issues in how data is captured, it doesn't seem to directly address the disproportionate flagging of one specific ethnic group. The problem described seems more related to the model's behavior rather than the raw data quality or capture process itself.
- Use case: Measurement bias would apply if the camera setup itself is flawed, leading to errors in the raw footage, but it's less relevant in this case where the issue seems to be with the model’s interpretation of data.
Option B: Sampling Bias
- Explanation: Sampling bias occurs when the data used to train the model is not representative of the real-world population, meaning that certain groups are either overrepresented or underrepresented in the training data. In this case, if the model was trained on data that disproportionately represented one ethnic group (or did not include enough diversity), the model may learn patterns that result in biased flagging of individuals from that group.
- Relevance: This option is highly relevant because it directly explains the issue. If the model was trained on a dataset that is not representative of the real-world distribution of ethnicities or over-represented certain groups, the model would develop biased behavior, flagging certain individuals more frequently based on ethnicity.
- Use case: Sampling bias is a strong contender when the model’s training data is skew...
Author: Joseph · Last updated Jul 19, 2026
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....
To build a customer service chatbot that can improve its responses over time by learning from past interactions and online resources, the company needs a learning strategy that allows for continuous adaptation and refinement. Let’s evaluate the options in detail:
A) Supervised learning with a manually curated dataset of good responses and bad responses
In supervised learning, the model learns from a labeled dataset, which includes pairs of inputs (e.g., customer inquiries) and corresponding outputs (e.g., good and bad responses). This approach can work well for training the chatbot initially with examples of both effective and ineffective responses, but it doesn’t inherently support the idea of continuous self-improvement. The model can only improve based on the fixed dataset it was trained on, so without constant manual updates and curation of new good and bad responses, this approach would not support dynamic learning from ongoing interactions.
Pros: Good for initial training; can improve with high-quality curated datasets.
Cons: Lacks continuous self-improvement and adaptation; requires manual effort to update the dataset.
B) Reinforcement learning with rewards for positive customer feedback
Reinforcement learning (RL) enables the chatbot to learn by interacting with its environment (e.g., customers) and receiving feedback. The model can adjust its behavior based on rewards (positive feedback) or penalties (negative feedback) for specific actions. This strategy is highly suited for a chatbot that wants to learn from its past interactions and improve over time. The feedback mechanism allows for dynamic learning, and the model can continuously adapt its responses as it receives more interactions. This is a true self-improvement capability, as the chatbot learns from both successes and mistakes.
Pros: Enables continuous, dynamic learning based on real interactions; can self-improve over time.
Cons: Requires significant setup and management to ensure that the feedback system is properly implemented and that t...
Author: Ming88 · Last updated Jul 19, 2026
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 me...
To evaluate the performance of a deep learning model that classifies the types of materials in images, we need to choose an appropriate metric based on the task (classification) and the evaluation goals.
A) Confusion matrix
A confusion matrix is a powerful tool for evaluating classification models, especially when there are multiple classes. It provides insight into how well the model distinguishes between different types of materials, by showing the counts of true positive, true negative, false positive, and false negative predictions for each class. This can help identify which material types the model is confusing with others and assess the overall performance in terms of precision, recall, and F1-score.
- Pros:
- Perfect for classification tasks, particularly multi-class classification.
- Helps in understanding where the model is making errors and which classes it struggles with.
- Provides detailed metrics like precision, recall, and F1-score, which are very informative in classification.
- Cons:
- Does not give a single numeric measure of overall performance by itself (but is used as a basis for other metrics).
Best use case: Evaluating classification tasks, especially when the goal is to analyze the specific performance of the model for each class.
B) Correlation matrix
A correlation matrix is used to measure the correlation between different variables. It is typically used for regression tasks or when analyzing relationships between continuous variables. In a classification scenario, this wouldn't be as useful because the target variable (material types) is categorical, not continuous.
- Pros:
- Useful for identifying correlations between continuous variables.
- Cons:
- Not suitable for classification tasks, especially multi-class classification.
- Doesn’t directly evaluate model performance.
Best use case: Understanding the relationships ...
Author: Stella · Last updated Jul 19, 2026
A company has built a chatbot that can respond to natural language questions with images. The company wants to ensure that the chatbot does not return inapprop...
Let's analyze each option based on the requirement: ensuring that the chatbot does not return inappropriate or unwanted images. We'll consider the effectiveness, effort, time, cost, and suitability of each solution.
Option A: Implement moderation APIs
Moderation APIs are designed to automatically filter and flag inappropriate content, including images, based on predefined criteria such as violence, nudity, or hate speech. These APIs can be integrated into the chatbot system, ensuring that any images returned by the chatbot are checked before being delivered to the user. This is a straightforward solution that reduces the effort, time, and cost associated with manually curating content. It also scales easily with minimal changes required in the chatbot's workflow.
Selected reason: Moderation APIs provide an effective and automated way to filter out inappropriate or unwanted images. They are specifically designed for this kind of content moderation task and are easy to implement, making them the ideal solution.
Option B: Retrain the model with a general public dataset
Retraining the model with a more diverse dataset may help to some extent in improving the quality of the model's outputs. However, this would not directly address the issue of ensuring that the model returns appropriate or unwanted images. Retraining can be resource-intensive, time-consuming, and expensive, and may not guarantee that the model will completely avoid generating inappropriate content, as even with a well-balanced dataset, some risks might still remain.
Rejection reason: While retraining could potentially reduce some inappropriate outputs, it doesn't specifically address the need for content moderation and may not be as reliable or efficient as using dedicated moderation APIs. It's a more complex and expensive approach.
...
Author: Ryan · Last updated Jul 19, 2026
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 moni...
To monitor model input and output data effectively when using Amazon Bedrock for summarizing session chats from the customer service department, the AI practitioner needs a strategy to store invocation logs and track both model requests and responses. Let's analyze each of the options provided:
A) Configure AWS CloudTrail as the logs destination for the model.
- Rejection Reasoning: AWS CloudTrail is primarily used for logging API activity across AWS services, which helps in auditing and tracking who did what and when. While it’s useful for monitoring AWS API calls and security-related information, it is not designed specifically for logging model inputs and outputs. CloudTrail won’t capture detailed information like the content of the model invocations or responses, making it unsuitable for tracking the specific data of model interactions.
B) Enable invocation logging in Amazon Bedrock.
- Selection Reasoning: Amazon Bedrock offers invocation logging as part of its functionality, which is specifically designed to track model inputs and outputs. By enabling invocation logging directly within Amazon Bedrock, the AI practitioner can ensure that the inputs and outputs of each model invocation are stored for monitoring, analysis, and auditing purposes. This option is optimized for the use case of tracking model interactions in real-time, and it will provide the necessary logs to assess model performance or review output from cust...
Author: Liam123 · Last updated Jul 19, 2026
A company is building an ML model to analyze archived data. The company must perform inference on large datasets that are multiple GBs in size. The company does not need to access the model predicti...
To determine the best solution for performing inference on large datasets that are multiple GBs in size, where immediate access to the model predictions is not required, we need to evaluate the available Amazon SageMaker inference options in the context of the requirements.
Option A: Batch transform
- Description: Batch Transform is an Amazon SageMaker feature that allows you to process large datasets in bulk. It is ideal for inference on large datasets where the model predictions can be processed asynchronously.
- Key factors:
- Large datasets: Batch Transform is specifically designed to handle large datasets efficiently, making it suitable for multi-GB datasets.
- Inference timing: Since the company does not require immediate access to the predictions, Batch Transform fits well because it processes the data in batches and stores the results in an S3 bucket, allowing you to access them when ready.
- Effort & time: The process is handled in a batch manner, and although it may take time depending on the dataset size, it's efficient for large-scale inference.
- Cost: You only pay for the compute resources used during the inference job, and the costs are relatively predictable for batch processing.
- Why it’s suitable: Batch Transform is designed for situations where you need to process large volumes of data asynchronously and do not need immediate access to predictions.
Option B: Real-time inference
- Description: Real-time inference in Amazon SageMaker is used for low-latency predictions, where the model needs to provide results immediately after receiving input.
- Key factors:
- Large datasets: Real-time inference is typically not ideal for multi-GB datasets, as it is designed for smaller, quicker requests. Handling large datasets in real-time would require splitting the data into smaller parts, which adds complexity.
- Inference timing: Since real-time inference requires the model to respond instantly, it’s not suitable for scenarios where the predictions don't need to be accessed immediately.
- Cost: Real-time inference incurs charges based on the number of requests and the compute resources used. If the dataset is large and inference isn't required immediately, this could lead to inefficient usage and higher costs.
- Why it’s not suitable: This is not a good fit for large datasets that don't need to be processed in real-time.
Option C: Serverless inference
- Description: Serverless inference allows you to run models ...
Author: SolarFalcon11 · Last updated Jul 19, 2026
Which term describes the numerical representations of real-world objects and concepts that AI and natural language processing (NLP) m...
Option Analysis:
A) Embeddings:
- Reasoning: Embeddings are numerical representations of real-world objects, words, or concepts in a lower-dimensional space that AI and NLP models use to improve understanding. They capture semantic relationships between words and can represent complex concepts by encoding the meaning and context of terms. Embeddings are essential in AI and NLP because they help models interpret and process textual data more effectively.
- Use case: Embeddings are specifically used to enhance understanding in tasks like text classification, sentiment analysis, machine translation, and more. They are a key part of most modern NLP models like BERT and GPT.
- Conclusion: Selected.
B) Tokens:
- Reasoning: Tokens refer to the individual pieces of text that are processed by an NLP model (e.g., words, characters, or subwords). While tokens are essential for NLP tasks, they represent the raw input rather than the numerical representations that improve understanding. Tokens themselves are not used to encode semantic meaning, which is the core purpose of embeddings.
- Use case: Tokens are used as part of preprocessing in NLP, but they are not the direct answer to the question of numerical representations used for improving understanding.
- Conclusion: Rejected.
C) Models:
- Reasoning: Models are the algorithms or systems used to process and analyze data. While models use embeddings as part of their architecture to improve understanding, models...
Author: Krishna · Last updated Jul 19, 2026
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 b...
In this scenario, the research company is using a foundation model (FM) from Amazon Bedrock to implement a chatbot, but the FM is struggling with complex scientific terms found in the research papers. The company wants to improve the chatbot's performance to better handle these terms. Let’s evaluate each of the options:
Option A: Use few-shot prompting to define how the FM can answer the questions
- Explanation: Few-shot prompting involves providing a small number of examples within the prompt to guide the model on how to answer specific questions. This method can be effective when you need the model to understand how to respond to specific types of questions or tasks, but it may not directly solve the issue of understanding complex scientific terms.
- Relevance: While few-shot prompting could help guide the model’s behavior, it doesn't directly address the core issue, which is the model’s difficulty with complex scientific terminology. Few-shot prompting would help the model structure its responses more effectively but won't necessarily improve its understanding of specialized language.
- Use case: This option is useful if the problem is related to response formatting or specific types of questions, but it won't help with the model's understanding of complex scientific terms in the research papers.
Option B: Use domain adaptation fine-tuning to adapt the FM to complex scientific terms
- Explanation: Domain adaptation fine-tuning involves training the foundation model on a domain-specific dataset to help the model better understand the language and terminology specific to that domain. In this case, fine-tuning the FM on research papers containing complex scientific terms will help the model adapt to those terms and improve its understanding and performance.
- Relevance: This option is highly relevant because it directly addresses the issue of the model’s poor performance with scientific terms. By fine-tuning the FM on research papers that use the terminology the model is struggling with, the model will learn how to better handle such terms and improve its performance.
- Use case: Domain adaptation fine-tuning is ideal in this scenario because it allows the company to adapt the model to the specific language of their research papers, improving the chat...
Author: Evelyn · Last updated Jul 19, 2026
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 adjus...
To make the responses from the large language model (LLM) on Amazon Bedrock more consistent for the same input prompt, the company should decrease the temperature value. Here's the reasoning:
Key factors for selecting Decrease the temperature value:
- Temperature parameter: In LLMs, the temperature setting controls the randomness of the model's output. A higher temperature leads to more diverse, creative, and potentially unpredictable responses, while a lower temperature generates more deterministic and consistent responses.
- Consistency: By decreasing the temperature value, the model becomes more likely to produce the same output for identical input prompts. This is exactly what the company needs for ensuring consistent sentiment analysis results.
- Fit: In sentiment analysis, consistency is important as the company wants reliable, repeatable results. Lowering the temperature ensures that the model doesn't introduce unnecessary randomness, which is critical in scenarios where accuracy and reliability are key.
Why other options are rejected:
- B) Increase the temperature value: Increasing the temperature value makes the model’s responses more varied and creative, which can lead to inconsistent answers for the same input. This would not help with achieving consistency in sentiment analysis, where reliable output i...
Author: Ethan · Last updated Jul 19, 2026
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 acc...
To meet the company's requirements of using Amazon Bedrock to develop a large language model (LLM) application while adhering to the security policy that restricts teams to accessing only their own customer data in Amazon S3, let's evaluate each option based on factors like security, ease of implementation, scalability, and adherence to the policy.
A) Create an Amazon Bedrock custom service role for each team that has access to only the team's customer data
- Benefits: This option involves creating a dedicated service role for each team, restricting their access to only the customer data belonging to their team. It directly aligns with the security policy, ensuring that teams only access the data they are permitted to. This solution provides granular access control, which is ideal for compliance.
- Drawback: This approach may increase the administrative overhead, as creating and managing multiple service roles could become complex as the number of teams grows. However, this is outweighed by the strong security benefits.
- Use Case: This is the most direct solution to meet the policy requirements, especially when fine-grained access control is crucial and the number of teams is manageable.
B) Create a custom service role that has Amazon S3 access. Ask teams to specify the customer name on each Amazon Bedrock request
- Benefits: This option simplifies access control by using a single service role for Amazon S3 access, with teams specifying the customer name during each request.
- Drawback: The security of this solution is weak because it relies on the teams to correctly specify the customer name. This could lead to potential mistakes or misconfigurations, which is a risk for data privacy and compliance. It also doesn't provide automatic enforcement of the policy at the service level, and it adds manual responsibility for ensuring correct data access.
- Use Case: This option could be used in low-risk scenarios where teams are trusted to manage data access correctly, but it does not offer strong security enforcement, making it less ideal for this case.
C) Redact personal data in Amazon S3. Update the S3 bucket policy to allow team access to customer data
- Ben...
Author: Ava · Last updated Jul 19, 2026
A medical company deployed a disease detection model on Amazon Bedrock. To comply with privacy policies, the company wants to prevent the model from including personal patient information in its responses. The company also wa...
Let’s analyze the options in relation to the requirements of ensuring privacy and compliance for the disease detection model:
A) Use Amazon Macie to scan the model's output for sensitive data and set up alerts for potential violations:
Amazon Macie is a service that helps identify and protect sensitive data, such as personally identifiable information (PII), in Amazon S3 buckets. While Macie can be useful for detecting sensitive data, it is not directly integrated with Amazon Bedrock, and the task here is related to scanning the model’s outputs, not data stored in S3. Using Macie would require additional infrastructure and may not be as seamless for this use case, especially for real-time detection of model output. This solution might be too indirect and require additional effort in integration.
B) Configure AWS CloudTrail to monitor the model's responses and create alerts for any detected personal information:
AWS CloudTrail logs API calls and provides visibility into user activity, but it doesn’t provide a direct mechanism to scan or analyze model outputs for sensitive information. It is great for tracking activity, but it doesn't help in scanning or filtering content like PII. For this use case, CloudTrail would not be effective because it doesn’t actively scan the data produced by the model in real time. It’s more about audit logs and tracking access, not about ensuring compliance with privacy policies by analyzing model responses.
C) Use Guardrails for Amazon Bedrock to filter content. Set up Amazon CloudW...
Author: Lucas · Last updated Jul 19, 2026
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 t...
The education provider wants to dynamically adjust the style of the model's response based on the age range of the user asking the question. Let's evaluate each option with respect to the least implementation effort, considering factors like ease of setup, flexibility, and maintenance:
A) Fine-tune the model by using additional training data that is representative of the various age ranges that the application will support:
- Effort: Fine-tuning involves significant effort in collecting age-specific data, preparing the training pipeline, retraining the model, and validating the results. This option requires substantial investment in terms of time and resources to create datasets, retrain the model, and manage the ongoing maintenance of these different versions.
- Why rejected: Although it can work, it demands considerable ongoing effort and complexity. The cost and effort associated with this approach are higher than the other solutions.
B) Add a role description to the prompt context that instructs the model of the age range that the response should target:
- Effort: This is a relatively low-effort solution. By simply modifying the prompt to include an instruction about the age range, the model can adjust its response style accordingly (e.g., "explain this to a 10-year-old" or "explain this to a college student"). It leverages the flexibility of prompt engineering to steer the model's responses without needing to retrain or fine-tune.
- Why selected: This approach directly addresses the requirement of modifying the response style with minimal effort and complexity. It is adaptable and doesn’t require heavy resources or significant reworking of the underlying model.
C) Use chain-of-thought reasoning to deduce the correct style and complex...
Author: Aria · Last updated Jul 19, 2026
Which strategy evaluates the accuracy of a foundation model (FM) that is used in image classificatio...
When evaluating the accuracy of a foundation model (FM) used in image classification tasks, it is essential to focus on how well the model performs at classifying images. Here's a breakdown of each option:
A) Calculate the total cost of resources used by the model:
- Effort & Relevance: This option assesses the resource consumption (e.g., computational cost, memory usage, etc.), but it doesn't measure how accurately the model performs in its image classification task. Cost-related factors may be important for deployment or scalability but are not a metric for model accuracy.
- Why rejected: While useful for operational insights, this does not provide any insight into the model's performance or accuracy in classifying images.
B) Measure the model's accuracy against a predefined benchmark dataset:
- Effort & Relevance: This strategy directly measures the model's performance by comparing its predictions against a known dataset where the correct answers (labels) are predefined. This is a standard and effective method for evaluating accuracy in classification tasks, including image classification. By calculating metrics like accuracy, precision, recall, and F1 score, the model's performance is clearly assessed.
- Why selected: This is the most common and accurate approach to evaluate the performance of a model in image classification tasks. Benchmark datasets (e.g., ImageNet, CIFAR-10) are widely used for this purpose.
C) Count the number of layers in the neural network:
- Effort & Relevance: Whil...
Author: Isabella · Last updated Jul 19, 2026
An accounting firm wants to implement a large language model (LLM) to automate document processing. The firm must proceed responsibly to avoid potential harms.What ...
In this scenario, the accounting firm needs to implement a large language model (LLM) for document processing while ensuring the model behaves responsibly and avoids potential harms. The firm must consider fairness, bias mitigation, model robustness, and ethical deployment.
Option Analysis:
A) Include fairness metrics for model evaluation.
- Pros: Evaluating the model with fairness metrics ensures that it performs equitably across various demographic groups or use cases. This is crucial in document processing, where biased outputs could have legal or financial consequences. Fairness metrics help identify and mitigate potential harms like discrimination or unequal treatment based on sensitive factors.
- Cons: Implementing fairness metrics can add some complexity to the evaluation process and may require additional tools or frameworks to measure fairness. This may increase effort and time, but the benefits in reducing harm are worth it.
- When to use: This should be used as a part of responsible AI practices in any application, particularly when dealing with legal, financial, or sensitive data, as in the case of the accounting firm.
B) Adjust the temperature parameter of the model.
- Pros: Adjusting the temperature in a model controls the randomness of the model’s responses. A lower temperature makes the model’s output more deterministic, while a higher temperature makes it more creative or varied.
- Cons: While adjusting the temperature can affect the model's creativity, it does not directly address responsible deployment, fairness, or ethical concerns. It is more about controlling the output style, not mitigating potential harms such as bias or unethical decision-making in document processing.
- When to use: Adjusting the temperature could be useful for fine-tuning the model's output for specific tasks (e.g., balancing creativity and precision), but it is not central to ensuring the model is deployed responsibly.
C) Modify the training data to mitigate bias.
- Pros: Bias in the training data can lead to biased outputs, which is a significant concern in many applications, especially in finance and legal documents. Modifying the training data to reduce bias ensures that the model outputs are more fair and representative, reducing potential harm in decision-making.
- Cons: This step may involve significant effort, as it requires carefully curating the data to avoid introducing biases or un...
Author: Ming · Last updated Jul 19, 2026
A company is building an ML model. The company collected new data and analyzed the data by creating a correlation matrix, calculating statistics, and visualizing...
The company is currently in the Exploratory Data Analysis (EDA) stage. Here's the reasoning:
Key factors for selecting EDA:
- Services and effort: EDA involves analyzing and understanding the data before any model-building begins. This can include visualizations, calculating basic statistics, and examining relationships between variables (like using a correlation matrix). It's more about gaining insights into the data's structure.
- Time and cost: This phase typically requires significant time for investigation and understanding, but it is a crucial part of the model pipeline as it guides subsequent decisions.
- Focus on Data: The company is focusing on data exploration and understanding the relationships within the data (as indicated by their use of a correlation matrix and visualizations).
Why other options are rejected:
- A) Data pre-processing: Data pre-processing usually comes before EDA, where tasks like cleaning data (removing missing values, handling outliers), normalization, or transformation take place. EDA is more about exploration and gaining insights, while pre-processing focuses on preparing the da...
Author: Emma Brown · Last updated Jul 19, 2026
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 ...
In this scenario, the company needs to suggest potential words to fill in missing text in documents. This suggests the need for a model that can understand the context of the surrounding text and generate or predict missing words accordingly.
Option Analysis:
A) Topic Modeling:
- Reasoning: Topic modeling techniques (such as Latent Dirichlet Allocation - LDA) are used to identify the underlying themes or topics within a collection of documents. While it can give a high-level understanding of document themes, it does not focus on predicting or filling in missing words in the context of the text. This is not ideal for this specific task.
- Use case: Topic modeling would be useful for categorizing documents based on themes, but not for filling in missing text.
- Conclusion: Rejected.
B) Clustering Models:
- Reasoning: Clustering models group similar documents or text together based on features. However, these models do not directly address the problem of predicting missing words. They are used for grouping data, but not for text completion.
- Use case: Clustering models might be helpful for organizing documents or detecting similar content, but they would not be able to suggest words for missing text.
- Conclusion: Rejected.
C) Prescriptive ML Models:
- Reasoning: Prescriptive ML models provide recommendations for actions based on data and insights, but they are often more concerned with making decisions or suggest...
Author: StarlightBear · Last updated Jul 19, 2026
A company wants to display the total sales for its top-selling products across various retail locations in the past 12 months.Which AWS solutio...
The company wants to automate the generation of graphs showing the total sales for its top-selling products across various retail locations over the past 12 months. The goal here is to visualize this sales data efficiently with minimal manual effort.
Let's evaluate each option:
Option A: Amazon Q in Amazon EC2
- Purpose: Amazon EC2 is an IaaS (Infrastructure as a Service) that provides virtual machines (instances) for running applications. While EC2 can be used to run custom applications (including data analytics tools or custom dashboards), it requires manual setup, management, and scaling of infrastructure. Using EC2 would not be an optimal solution for automating graph generation since the company would need to configure and manage the application stack.
- Reason for rejection: EC2 provides infrastructure but lacks automation and is not tailored for the direct purpose of automating the generation of visual data insights such as graphs. It would require more effort and resources compared to other options.
Option B: Amazon Q Developer
- Purpose: Amazon Q Developer is a tool designed to help developers build generative AI applications. While it can help create applications and automate some tasks, it isn't specifically focused on data visualization or graph generation for sales data.
- Reason for rejection: While generative AI can be useful for some automated tasks, this option is not aligned with the specific need for visualizing sales data in the form of graphs. It would be an over-engineered solution for this use case, where a straightforward data visualization tool is better suited.
Option C: Amazon Q in Amazon QuickSight
- Purpose: Amazon QuickSight is a scalable business intelligence (BI) service that allows users to create visualizations, reports, and dashboards from their data. Amazon Q is a feature within QuickSight that allows users to query data using natural language (NLQ, or natural language queries) an...
Author: Andrew · Last updated Jul 19, 2026
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...
In this scenario, the company is using few-shot learning to improve the accuracy of intent detection for their chatbot, and they are utilizing a large language model (LLM) from Amazon Bedrock. The goal is to improve intent detection accuracy, which means the system needs to learn how to map user messages to the correct intents effectively. Let's evaluate the options based on this context.
A) Pairs of chatbot responses and correct user intents
While pairs of chatbot responses and correct user intents might seem useful for evaluating the chatbot’s responses, this pairing focuses on the final output of the chatbot, rather than the critical step of mapping the user's input (message) to the correct intent. The model needs to learn how to detect the user's intent from their message, not from the chatbot’s response.
- Pros:
- Could be useful in assessing chatbot performance, but not for intent detection directly.
- Cons:
- This option doesn’t directly address the core requirement, which is to improve intent detection based on user inputs.
Best use case: This could be used for evaluating or fine-tuning chatbot responses, but not ideal for improving intent detection accuracy directly.
B) Pairs of user messages and correct chatbot responses
While this pairing helps with ensuring the chatbot provides appropriate responses, it doesn’t directly help in training the model for intent detection. The goal is to understand the user's intent based on their input message, and responses may not directly provide insight into this mapping.
- Pros:
- Useful for training the chatbot to improve response quality.
- Cons:
- Does not help in directly improving intent detection, which is the main focus.
Best use case: This would be relevant if the company wanted to improve the chatbot's response generation rather than intent detection.
C) Pairs of user messages and correct user intents
This option provides the user messages...
Author: NebulaEagle11 · Last updated Jul 19, 2026
A company is using few-shot prompting on a base model that is hosted on Amazon Bedrock. The model currently uses 10 examples in the prompt. The model is invoked once daily and is performing well. Th...
The company is using few-shot prompting on a base model hosted on Amazon Bedrock with 10 examples in the prompt. The goal is to lower the monthly cost while maintaining the model's performance, which is performing well with the current setup (invoked once daily).
Let's evaluate each option:
A) Customize the model by using fine-tuning
Fine-tuning a model allows you to adapt it to a specific task using domain-specific data. However, fine-tuning typically requires a significant amount of data and computational resources to retrain the model, which increases costs in the short term. Once fine-tuned, the model may need fewer examples or lower token usage during inference, but the upfront costs for fine-tuning would not immediately reduce the monthly cost, making it a more expensive option initially.
- Pros:
- Can improve model performance in the long term, possibly reducing the number of examples needed for few-shot learning.
- Cons:
- High upfront costs for fine-tuning.
- Not the most cost-effective solution for a company looking to lower monthly costs immediately.
Best use case: Useful for long-term performance optimization but not suitable for reducing short-term monthly costs.
B) Decrease the number of tokens in the prompt
Reducing the number of tokens in the prompt directly impacts the cost of each API call. Since the cost of API usage on Amazon Bedrock is often linked to the number of tokens processed, decreasing the number of tokens would help reduce monthly costs. This can be achieved by:
- Reducing the number of examples in the prompt (from 10 to fewer examples).
- Trimming down unnecessary content in the prompt without sacrificing performance.
- Pros:
- Directly lowers cost per invocation.
- Easy to implement by optimizing the prompt without significant effort or reworking the model.
- Cons:
- Reducing the number of tokens too much might impact model per...
Author: Kai99 · Last updated Jul 19, 2026
An AI practitioner is using a large language model (LLM) to create content for marketing campaigns. The generated content sounds plausible a...
In this scenario, the LLM is generating content that sounds plausible and factual but is actually incorrect. This points to a specific issue with how the model generates its responses, so let's evaluate the options to understand which one best fits the problem:
A) Data leakage
Data leakage refers to a situation where information from outside the training dataset unintentionally influences the model’s predictions. This happens when the model accidentally "leaks" data during training, leading to overly optimistic performance. Data leakage typically results in an artificially high model performance on training data or validation sets, but it is not directly related to generating content that sounds plausible but is incorrect.
Reason for rejection: Data leakage affects the training process or model performance but is not the cause of the model generating plausible yet incorrect content.
B) Hallucination
Hallucination in the context of large language models refers to the model generating text that seems plausible but is factually incorrect or fabricated. This is a known issue in LLMs, where the model might create text based on patterns learned during training but without access to or understanding of the underlying factual accuracy. The content may sound confident and factual but can be completely incorrect or invented.
Reason for selection: Hallucination perfectly describes the problem where an LLM generates content that is inaccurate, even though it appears to be truthful. This issue is common in language models, especially when they are not fine-tuned with strong factual grounding or external validation mechanisms.
C) Overfitting
Overfitting occurs when a model learns the training data too well, including noise and outliers, which negatively affects its ability to generalize to unseen data. Overfitting might lead to a model perfor...
Author: Noah Williams · Last updated Jul 19, 2026
An AI practitioner trained a custom model on Amazon Bedrock by using a training dataset that contains confidential data. The AI practitioner wants to ensure that the custom model does not generate inference responses based o...
To ensure that a custom model trained with confidential data on Amazon Bedrock does not generate responses based on that confidential data, the AI practitioner must take steps to prevent any leakage of sensitive information during inference. We need to consider security, effort, cost, and practicality in choosing the best solution.
Option Analysis:
A) Delete the custom model. Remove the confidential data from the training dataset. Retrain the custom model.
- Pros: This option directly addresses the issue by removing confidential data from the training process. By retraining the model without this data, the model should no longer generate responses based on confidential information. It ensures that no confidential data is included in the model’s training or behavior.
- Cons: This approach involves significant time and effort. The practitioner would need to reprocess the dataset, retrain the model, and ensure that the new version of the model does not contain any of the sensitive information. This could also be costly, depending on the scale of the dataset and model.
- When to use: This is the most thorough and effective solution to ensure that confidential data is not present in the model. However, it is also the most time-consuming and costly, making it more suitable when the risk of leakage is high and when the practitioner can afford the retraining process.
B) Mask the confidential data in the inference responses by using dynamic data masking.
- Pros: Dynamic data masking is a technique that can prevent certain sensitive data from appearing in inference responses. This is an efficient way to prevent confidential information from being exposed during inference without modifying the model itself.
- Cons: While dynamic data masking can prevent the disclosure of confidential information during inference, it does not prevent the model from having learned from the confidential data in the first place. This means the model could still generate responses influenced by confidential information, even if that data is masked in the output. It is a reactive solution rather than a preventative one.
- When to use: This is useful in situations where data leakage must be immediately addresse...
Author: Mia · Last updated Jul 19, 2026
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 ex...
To evaluate the accuracy of a solution that uses large language models (LLMs) to translate training manuals from English into other languages, the company needs to assess how well the generated translations match the expected outcomes.
Let’s examine the provided evaluation strategies:
A) Bilingual Evaluation Understudy (BLEU)
- Selection Reasoning: BLEU is a widely used metric specifically designed for machine translation tasks. It compares the n-grams of the machine-generated translations with reference translations, and higher scores indicate better translation quality. BLEU focuses on precision and can provide a clear measure of translation accuracy, making it the most appropriate metric for evaluating the output of the generative AI solution for translating manuals into different languages.
- Scenario: BLEU is ideal in this scenario because it’s tailored for machine translation tasks, measuring how well the generated translations align with reference translations.
B) Root mean squared error (RMSE)
- Rejection Reasoning: RMSE is typically used for regression tasks, where the goal is to predict continuous numerical values. It is not suitable for evaluating translation tasks, which involve text and linguistic accuracy. ...
Author: Noah · Last updated Jul 19, 2026
A large retailer receives thousands of customer support inquiries about products every day. The customer support inquiries need to be processed and responded to quickly. The company wants to implement Agents for Amazon ...
When analyzing the key benefits of using Amazon Bedrock agents for handling customer support inquiries, we need to assess how each option impacts various factors such as services, effort, time, and cost. Let's explore each option and how it applies to the retailer's needs:
A) Generation of custom foundation models (FMs) to predict customer needs
- Benefits: This option could allow for the creation of custom models tailored specifically to predict customer behavior or identify common issues. It offers more accurate responses based on past interactions and patterns, which can improve service quality.
- Drawback: While beneficial in prediction, the creation of custom models requires significant effort in training and tuning the models. Additionally, it may take a considerable amount of time and resources to generate custom foundation models, which could delay the ability to respond to customer inquiries quickly.
- Use Case: This option is more suited for businesses looking to personalize customer interactions deeply or forecast future needs, but it’s not ideal for this scenario where quick response time is critical.
B) Automation of repetitive tasks and orchestration of complex workflows
- Benefits: Automating repetitive tasks such as order tracking, basic inquiries, and FAQs can significantly reduce the workload for human agents, allowing them to focus on more complex issues. Orchestrating workflows ensures smooth handling of multiple tasks simultaneously, improving efficiency and reducing response times.
- Drawback: While it is a valuable option for improving operational efficiency, it may not directly address the intricacies of customer inquiries that require varied solutions across different products. Some level of customization may still be needed, depending on the inquiry.
- Use Case: This option is ideal for retailers who have many routine inquiries and need to reduce manual effort. It improves response time and can scale easily, making it a great choice for high-volume inquiries.
C) Automatically calling multiple foundation models (FMs) and consolidating the results
- Benefits: By automatically using multiple foundation models, the system can leverage di...
Author: Ishaan · Last updated Jul 19, 2026
Which option is a benefit of ongoing pre-training when fine-tuning a foundation model (FM)?
In the context of ongoing pre-training when fine-tuning a foundation model (FM), let's analyze each option in relation to the benefits it provides:
A) Helps decrease the model's complexity:
- Ongoing pre-training typically doesn't decrease the complexity of the model; instead, it might increase the complexity by enhancing the model’s ability to learn more from additional data. Pre-training generally makes a model more sophisticated to handle various tasks, especially if fine-tuning continues to improve its capabilities. Thus, this option doesn’t align with the main benefit of ongoing pre-training.
B) Improves model performance over time:
- Ongoing pre-training can indeed improve model performance over time. As the model is exposed to more data and continues to learn, its accuracy and ability to generalize can improve, especially when fine-tuning is applied to the foundation model to tailor it for specific tasks. This aligns with the benefits of continuous learning and adapting to new information, which is a key advantage of ongoing pre-training.
C) Decreases the training time requirement:
- Ongoing...
Author: Aarav · Last updated Jul 19, 2026
What are tokens in the context of generative AI models?
In the context of generative AI models, understanding the role of tokens is crucial as they influence both the functioning and efficiency of the model. Let's analyze each option in detail based on the definition and application of tokens in generative AI models.
A) Tokens are the basic units of input and output that a generative AI model operates on, representing words, subwords, or other linguistic units.
This is the most accurate definition of tokens in generative AI. Tokens are the fundamental units that models like GPT-3 or BERT process. These units can represent words, parts of words, or even characters in some cases, depending on the tokenizer used by the model. For example, the word "chatbot" might be split into two tokens: "chat" and "bot." Tokens serve as the building blocks for both the input and output in the model's operation. This option aligns perfectly with the way tokens function in AI models.
B) Tokens are the mathematical representations of words or concepts used in generative AI models.
While tokens may be mathematically represented in vectors for processing (using embeddings, for example), the token itself is not specifically the mathematical representation of a word or concept. The definition in option A is more accurate because it highlights the role of tokens as linguistic units that the model uses to generate responses. Option B is partially true, but it focuses too much on the mathematical aspect and doesn't capture the fundamental nature of tokens as input/output units.
...
Author: Kai · Last updated Jul 19, 2026
A company wants to assess the costs that are associated with using a large language model (LLM) to generate inferences. The company wants to use Amazon Bedrock to build...
In the context of Amazon Bedrock and assessing the costs associated with using a large language model (LLM) for inference, the primary factor driving inference costs would be the number of tokens consumed. Here's the reasoning behind each option:
A) Number of tokens consumed:
- Inference costs in services like Amazon Bedrock are generally based on the number of tokens processed by the model. When you generate inferences or responses from an LLM, each request involves a certain number of tokens (input and output tokens). Since pricing is often tied to how many tokens are processed by the LLM during inference (the prediction stage), this is the key cost driver for inference, which aligns directly with the usage.
B) Temperature value:
- The temperature value is a parameter used to control the randomness or creativity of the generated text. While it does influence the nature of the output, ...
Author: Ethan · Last updated Jul 19, 2026
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 A...
In this case, the company needs to manage the flow of data from Amazon S3 to SageMaker Studio notebooks, which means ensuring efficient and secure access to data stored in Amazon S3 while using SageMaker Studio. Let’s break down each option to see which best meets this requirement.
Option A: Use Amazon Inspector to monitor SageMaker Studio
- Explanation: Amazon Inspector is a security assessment service designed to identify vulnerabilities or deviations from best practices in the AWS environment. It helps monitor the security posture of applications, but it’s not focused on data management or facilitating data flow between S3 and SageMaker Studio.
- Relevance: While Amazon Inspector is useful for security audits and vulnerability assessments, it doesn't address the requirement of managing the flow of data between S3 and SageMaker Studio notebooks. It focuses on security rather than data transfer.
- Use case: This option is useful for monitoring security, but not for managing data flow.
Option B: Use Amazon Macie to monitor SageMaker Studio
- Explanation: Amazon Macie is a data security and privacy service that uses machine learning to discover, classify, and protect sensitive data. It helps identify personally identifiable information (PII) within S3 buckets but does not directly manage or facilitate data flow between S3 and SageMaker Studio notebooks.
- Relevance: Although Amazon Macie is helpful in monitoring data security and privacy, it does not provide functionality for managing the flow of data between S3 and SageMaker Studio. It is primarily a data classification tool rather than a data transfer or management tool.
- Use case: This option is relevant for securing sensitive data in S3 but not for managing the flow of data to SageMaker Studio notebooks.
Option C: Configure SageMaker to use a VPC with an S3 endpoint
- Explanation: Configuring a VPC (Virtual Private Cloud) with an S3 endpoi...
Author: James · Last updated Jul 19, 2026
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 ...
Let's analyze each option based on the requirement: uploading a new dataset to Amazon Bedrock for validating a customized foundation model's (FM) responses.
Option A: Amazon S3
Amazon S3 (Simple Storage Service) is a scalable object storage service designed to store large amounts of data, including datasets. It is commonly used for storing and managing data used in machine learning processes, such as model validation and training. Amazon Bedrock integrates with S3 for tasks like data storage and retrieval. The company can upload the validation dataset to an S3 bucket, and Amazon Bedrock can access the data for validation purposes. S3 is cost-effective, easy to use, and highly scalable.
Selected reason: S3 is the most appropriate service for uploading datasets, especially for validation tasks. It integrates seamlessly with Amazon Bedrock, and offers flexibility in terms of cost, scalability, and ease of use. S3 is commonly used for storing large datasets and is directly compatible with Bedrock workflows.
Option B: Amazon Elastic Block Store (Amazon EBS)
Amazon EBS is designed for block-level storage and is typically used for attaching storage volumes to EC2 instances. While it provides high-performance storage for EC2, it is not an optimal choice for uploading and managing large datasets for model validation. EBS is more suited for use cases that require low-latency access to data on EC2 instances rather than serving as a storage solution for datasets in the context of machine learning validation.
Rejection reason: EBS is primarily designed for block storage associated with EC2 instances and does not offer the same scalability, flexibility, or ease of integration with services like Amazon Bedrock as S3 does. It’s not the best fit for dataset storage in this case.
Option C: Amazon Elastic File System (Amazon ...
Author: Krishna · Last updated Jul 19, 2026
Which prompting attack directly exposes the configured behavior of a large language model (LLM)?
To address the question about which prompting attack directly exposes the configured behavior of a large language model (LLM), let's evaluate each of the given options:
A) Prompted persona switches
- Purpose: This attack involves manipulating the LLM to switch between different personas or personalities during a conversation. The idea is to alter the behavior of the model by prompting it to take on a new role or persona.
- Reason for rejection: While this attack can change the behavior of the model, it doesn’t directly expose the configured behavior or the underlying prompt template. It's more about manipulating the model’s responses or personality, rather than revealing how the model was set up or how it operates.
- Scenario where it can be used: This could be used in situations where a model is designed to take on different personas, and the goal is to trick the model into adopting one that could behave inappropriately or unexpectedly.
B) Exploiting friendliness and trust
- Purpose: This attack involves exploiting the model's tendencies to be polite, friendly, or accommodating in its responses. Attackers can use this to manipulate the LLM into giving undesirable outputs.
- Reason for rejection: While exploiting friendliness and trust can manipulate the outputs or behaviors of the model, it doesn’t necessarily expose the internal configured behavior or prompt template. It’s more about exploiting the behavioral traits of the model rather than revealing its configuration or structure.
- Scenario where it can be used: This attack could be useful in making the model give inappropriate responses by relying on its politeness, but it doesn't directly expose how the model works under the hood.
C) Ignoring the prompt template
- Purpose: This attack involves c...
Author: StarlightBear · Last updated Jul 19, 2026
A company wants to use Amazon Bedrock. The company needs to review which security aspects the company is responsible for when using Amazon Bedr...
When evaluating the responsibilities a company has while using Amazon Bedrock, it’s important to understand the shared responsibility model between AWS and the customer. Here's a detailed analysis of each option:
A) Patching and updating the versions of Amazon Bedrock
- Reasoning: Amazon Bedrock is a fully managed service, meaning AWS handles the patching and updates for the underlying infrastructure and software components. The company is not responsible for patching or updating Amazon Bedrock itself.
- Rejected because: AWS takes care of patching and version updates for managed services like Bedrock.
B) Protecting the infrastructure that hosts Amazon Bedrock
- Reasoning: Amazon Bedrock runs on AWS’s infrastructure, which is already secured and maintained by AWS. The responsibility for physical and network-level security of the infrastructure (such as hardware, network, and hypervisor) lies with AWS, not the customer.
- Rejected because: The security of AWS's infrastructure, including physical servers and networking, is managed by AWS, so the company doesn’t need to worry about this.
C) Securing the company's data in transit and at rest
- Reasoning: This is a key responsibility for t...
Author: Isabella · Last updated Jul 19, 2026
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 mode...
To evaluate the toxicity of the LLM outputs with the least operational overhead, let's analyze each of the options:
A) Crowd-sourced evaluation
Crowd-sourcing involves recruiting a large number of individuals, usually via platforms like Amazon Mechanical Turk, to manually assess the toxicity of the model outputs. While this can provide diverse perspectives, it introduces significant operational overhead. The company would need to set up tasks for workers, manage large-scale data collection, handle quality control of the responses, and ensure that the evaluations are accurate. There are also time and cost considerations for compensating workers.
Reason for rejection: While this can work, it’s resource-intensive and requires a lot of manual management, making it less efficient than the other options.
B) Automatic model evaluation
Automatic evaluation involves using pre-built toxicity detection models or tools to assess the outputs of the LLMs. These tools, such as AWS services like Amazon Comprehend, can automatically analyze the generated content for toxicity based on predefined criteria. This strategy is low-effort and can be set up quickly, as no manual work or human intervention is required. Additionally, it can be scaled easily, making it a highly efficient and cost-effective solution.
Reason for selection: This approach offers the least operational overhead. It leverages existing services to automate the evaluation process, which saves time, cost, and effort. It also ensures scalability and consistency.
C) Model evaluation with human workers
This option is similar to crowd-sourcing but generally involves fewer workers or a more structured evaluation process. Human workers are engaged to manually assess the outputs, but this still leads to significant operational overhead (scheduling, quality control, ...
Author: Vikram · Last updated Jul 19, 2026
A company is testing the security of a foundation model (FM). During testing, the company wants to get around the safety features and make ha...
In this scenario, the company is testing the security of a foundation model (FM) and aims to bypass its safety features to generate harmful content. This suggests that the company is trying to intentionally find weaknesses in the model by manipulating its behavior.
Option Analysis:
A) Fuzzing training data to find vulnerabilities:
- Description: Fuzzing involves providing random or unexpected inputs to a system in order to find vulnerabilities, often related to handling edge cases or unexpected inputs.
- Reasoning: While fuzzing could be used in testing a foundation model, it typically focuses on finding vulnerabilities in input handling or data processing. In the context of testing the FM’s security, fuzzing the training data might help identify weaknesses in how the model processes certain inputs, but the specific goal here is to bypass the safety features to generate harmful content, which is a more deliberate attempt to subvert the model's behavior.
- Scenario Usage: Fuzzing could be used to identify general input vulnerabilities, but it's not the best fit for deliberately bypassing safety features to produce harmful content.
B) Denial of service (DoS):
- Description: A DoS attack aims to make a service or system unavailable by overwhelming it with excessive requests or exploiting vulnerabilities that cause the system to crash or malfunction.
- Reasoning: A DoS attack focuses on making a service unavailable, not on bypassing security features to produce harmful content. The company’s goal in this scenario is not to disrupt the model’s operation, but rather to test the model’s ability to generate harmful content despite safety features.
- Scenario Usage: DoS is not relevant to the scenario described, as it does not involve testing for harmful content generation or subverting safety features.
C) Penetration testing with authorization:
- Description: Penetration testing in...
Author: Lucas · Last updated Jul 19, 2026
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 enviro...
To meet the company's requirements of running Amazon SageMaker jobs in an isolated environment without internet access while ensuring regulatory compliance, let's evaluate each option in detail based on the criteria of isolation, effort, cost, and time.
A) Run SageMaker training and inference by using SageMaker Experiments
SageMaker Experiments is used to track and organize machine learning experiments, models, and data to compare different versions of models or experiments. However, SageMaker Experiments does not specifically address the need for an isolated environment or controlling internet access during SageMaker jobs. While useful for managing experiments, it does not fulfill the regulatory requirement to isolate the environment from the internet.
Rejection Reason: SageMaker Experiments is useful for model tracking and management but does not provide the necessary isolation from the internet for regulatory compliance.
B) Run SageMaker training and Inference by using network Isolation
Amazon SageMaker offers a feature called "Network Isolation" where you can configure your training and inference jobs to run within a Virtual Private Cloud (VPC) that does not have any internet access. This ensures that the SageMaker jobs are isolated from the public internet, which aligns directly with the company's regulatory requirement for running jobs in an isolated environment. You can configure this isolation by specifying VPC subnets with no route to the internet and controlling access via security groups and network ACLs.
Reason for Selection: This option directly meets the requirement to isolate SageMaker jobs from the internet. ...
Author: Ravi Patel · Last updated Jul 19, 2026
An ML research team develops custom ML models. The model artifacts are shared with other teams for integration into products and services. The ML team retains the model training code and data. The ML team wants to build a mechanism that the ML...
To address the ML team's requirement to audit models and ensure traceability and transparency of the models, we need to evaluate the available options considering factors like ease of implementation, transparency, security, and long-term sustainability.
Option Analysis:
A) Create documents with the relevant information. Store the documents in Amazon S3.
- Pros: This option allows the team to store documents with relevant information about the models. It can be a simple solution with relatively low upfront cost and time.
- Cons: This solution doesn't provide a structured and automated way to ensure full model transparency or integration with model management tools. It also requires manual effort to maintain consistency in documenting models and their details. It doesn't offer built-in features for auditing, version control, or specific integrations for model performance tracking.
- When to use: This could be used for basic documentation or for small-scale, less complex projects, but it lacks robustness and scalability for ML model auditing and tracking over time.
B) Use AWS AI Service Cards for transparency and understanding models.
- Pros: AWS AI Service Cards provide insights into the models and their components, offering transparency. These cards are designed to document the AI service, model types, and use cases, helping teams understand the scope of the model's capabilities.
- Cons: AWS AI Service Cards are primarily useful for AWS prebuilt AI services and may not be directly applicable for custom ML models developed by a research team. Additionally, this approach may not offer the level of control and auditing needed for custom models.
- When to use: This would be ideal if the team was working with AWS prebuilt models or services, but it is not an ideal choice for custom models developed in-house.
C) Create Amazon SageMaker Model Cards with intended uses and training and inference details.
- Pros: Amazon SageMaker Model Cards are specifically designed for documenting ML models, providing a structured format for model details, training data, intended use cases, performance metrics, and o...
Author: Joseph · Last updated Jul 19, 2026
A software company builds tools for customers. The company wants to use AI to increase software development product...
To determine which solution best increases software development productivity, we need to assess each option based on its ability to streamline or enhance the development process while considering factors like services, effort, time, cost, and their direct impact on productivity.
1. Use a binary classification model to generate code reviews:
- Relevance: A binary classification model typically categorizes inputs into two groups (e.g., "good" vs. "bad" code). While code reviews are important for maintaining code quality, generating them automatically through a binary classification model might not be the most productive solution, as reviews often require nuanced understanding and context.
- Reason for rejection: This solution might help identify certain issues, but it may not fully address productivity in terms of speeding up development or improving the coding process directly.
- Scenario: This could be useful in specific cases for automating simple code quality checks, but it’s not a comprehensive solution for boosting overall development productivity.
2. Install code recommendation software in the company's developer tools:
- Relevance: Code recommendation software can suggest snippets, functions, or even entire lines of code based on the developer’s input, making it easier and faster for developers to write code. This can significantly speed up development by reducing the time spent on writing boilerplate code or looking up documentation.
- Reason for selection: This solution has the potential to improve productivity by providing quick suggestions that reduce development time and help avoid errors. It directly addresses the goal of increasing software development productivity.
- Scenario: This is widely used in developer tools (e.g., GitHub Copilot) to help developers write code more efficiently by suggesting solutions based on context, thus speeding up the overall development process.
3. ...
Author: Vikram · Last updated Jul 19, 2026
A retail store wants to predict the demand for a specific product for the next few weeks by using the Amazon SageMaker DeepAR forecasti...
To predict the demand for a specific product for the next few weeks using the Amazon SageMaker DeepAR forecasting algorithm, it’s crucial to understand the nature of the algorithm and the type of data it requires. The DeepAR algorithm is designed for time series forecasting, which involves making predictions about future values based on historical data points over time.
Let’s analyze each option in detail:
A) Text data
- Reasoning: Text data involves unstructured data, typically used for natural language processing (NLP) tasks like sentiment analysis, translation, or text generation. It’s not suitable for forecasting demand or any time-dependent numerical predictions.
- Rejected because: DeepAR is specifically designed for time series data, not text data. Text data would not work well for demand forecasting.
B) Image data
- Reasoning: Image data is generally used in computer vision tasks, such as object detection or image classification. It is not directly applicable to forecasting numerical trends over time.
- Rejected because: Image data is not used in time series forecasting tasks like predicting deman...
Author: Mia · Last updated Jul 19, 2026
A large retail bank wants to develop an ML system to help the risk management team decide on loan allocations for different demograph...
To develop an unbiased ML model for the bank's loan allocation system, the best approach is Option D: Measure class imbalance on the training dataset and adapt the training process accordingly.
Analysis of each option:
1. A) Reduce the size of the training dataset.
- Reducing the size of the training dataset could lead to underfitting or loss of important information, which is especially problematic for an unbiased model. A smaller dataset may miss out on critical data that reflects the diversity of the different demographics, potentially increasing bias.
- Effort: Medium (if you reduce data, you risk losing important patterns).
- Time: Medium (reduction could save some time in training but would likely decrease model performance).
- Cost: Medium (resources are wasted on incomplete data, and training may be less effective).
2. B) Ensure that the ML model predictions are consistent with historical results.
- While historical results can provide context and benchmarks, consistent predictions with historical data do not ensure the model is unbiased. If historical data is biased, replicating the same patterns will simply perpetuate that bias. The goal should be to mitigate bias, not maintain it.
- Effort: Low, but would not effectively address the problem of bias.
- Time: Low, but ineffective for bias correction.
- Cost: Low, but also inefficient and unhelpful in eliminating bias.
3. C) Create a different ML model for each demographic group.
- Creating separate models for each demographic would introduce additional complexity and may result in inconsistent decision-making across groups. It could also lead to segmentation ...
Author: Isabella1 · Last updated Jul 19, 2026
Which prompting technique can protect against prompt injection attacks?
Prompt injection attacks involve manipulating the prompt given to an AI model in order to influence its response in unintended ways. To protect against these attacks, we need a strategy that can either verify or guide the model’s reasoning process to ensure the input prompt remains under control.
Let’s analyze each option based on its relevance to protecting against prompt injection attacks:
A) Adversarial prompting
- Reasoning: Adversarial prompting typically involves creating inputs that are specifically designed to trick or manipulate the AI model into making errors or misbehaving. While it’s a term used to describe a form of attack, it’s not a defensive technique. In fact, adversarial prompting is more about attacking the model, not protecting it.
- Rejected because: It’s an attack method, not a defense.
B) Zero-shot prompting
- Reasoning: Zero-shot prompting refers to prompting the model to perform tasks without any task-specific training. It allows the model to generate responses to unseen tasks, but this technique doesn’t inherently protect against prompt injection. The model may still be vulnerable to manipulation if the prompt is injected with malicious instructions.
- Rejected because: Zero-shot prompting doesn't specifically address security concerns or protection against prompt injection attacks.
C) Least-to-most prompting
- Reasoning: Least-to-most prompting is a technique that starts with simpler, less specific prompts ...
Author: Carlos Garcia · Last updated Jul 19, 2026
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 policie...
To determine the most cost-effective solution, we need to evaluate each option based on the factors of services, effort, time, cost, and their suitability for providing real-time contextual responses for a customer service chatbot.
A) Retrain the LLM on the company policy data
Retraining an LLM from scratch would be a highly resource-intensive and costly option. It involves significant computing resources, expertise, and time to train the model from the ground up. This method would likely take the longest time and be the most expensive, as the company would need to start from scratch in terms of training.
Rejection Reason: This option is impractical for cost-efficiency due to the heavy resource consumption and long development time.
B) Fine-tune the LLM on the company policy data
Fine-tuning an already trained LLM on the company’s specific policy data is a more efficient method than retraining. Fine-tuning involves updating the model's parameters using a smaller dataset (in this case, the company policies) to adapt it to the company’s specific domain. This method is less costly and faster than retraining, but still requires time and resources, particularly for high-quality data preparation and model adjustments.
Rejection Reason: Although fine-tuning is more cost-effective than retraining, it still requires substantial effort and resources for model adjustments, and it does not guarantee the real-time contextual responses needed for this case.
C) Implement Retrieval Augmented Generation (RAG) for in-context responses
RAG combines a pre-trained language model with a retrieval system that dynami...
Author: Scarlett · Last updated Jul 19, 2026
A company wants to create a new solution by using AWS Glue. The company has minimal programming experience with AWS Glue....
To help the company use AWS Glue with minimal programming experience, the best option would be Amazon Q Developer. Here's the reasoning:
Key factors for selecting Amazon Q Developer:
- Services and effort: Amazon Q Developer provides a low-code environment for developing data-driven applications, and it simplifies the process of working with AWS Glue. It is specifically designed to allow users with minimal programming experience to leverage AWS Glue and other AWS services for building data pipelines and applications.
- Time and cost: Using a low-code solution like Amazon Q Developer significantly reduces development time because it provides an intuitive, easy-to-use interface, with pre-built templates and features to integrate with AWS Glue, minimizing the need for complex coding. This can also reduce costs in terms of training and development.
- Fit: Since the company has minimal programming experience, a low-code solution that simplifies AWS Glue's use is the most fitting option. Amazon Q Developer helps abstract the complexity of coding and makes it easier to interact with AWS Glue in a more accessible way.
Why other options are rejected:
- B) AWS Config: AWS Config is a service that helps manage and track configurations of AWS resources. It is not designed for simplifying data pipeline creation or reducing the programming effort for using AWS Glue. It is more about governance ...
Author: Olivia Johnson · Last updated Jul 19, 2026
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 geograp...
In the given scenario, the company is developing a mobile ML app that diagnoses and treats insect bites. The key focus here is ensuring the image classification model is trained on a diverse dataset that includes insect bite photos from different genders, ethnicities, and geographic locations. Let's break down how this aligns with the principles of responsible AI.
Option A: Fairness
- Efficiency: Fairness is about ensuring that an AI model performs equitably across all demographic groups. By using a diverse dataset that includes varied representations of genders, ethnicities, and geographic locations, the company is ensuring that the model does not develop biases towards any particular group. This ensures fair performance across different population segments, which is especially crucial for healthcare-related applications like diagnosing insect bites.
- Cost: Although gathering a diverse dataset can be resource-intensive, it ensures that the model generalizes well to all users, which reduces the likelihood of biased predictions and costly errors in real-world applications.
- Time: The time spent on collecting and curating a diverse dataset is an investment in avoiding potential disparities in model performance down the line.
- Key Factor: The company’s emphasis on diversity directly aligns with the principle of fairness because it ensures that the model is not biased towards certain groups and treats all users equitably.
Option B: Explainability
- Efficiency: Explainability refers to how well the model's decisions can be understood by humans. While important in many applications, explainability is not the focus of this particular scenario. The company’s goal is more about creating a model that performs well across diverse populations rather than ensuring the model's reasoning is transparent.
- Cost: Explainability models can be complex and computationally expensive, but that’s not the primary focus here.
- Time: The time invested in making a model interpretable is not aligned with the core objective of ensuring fairness across diverse groups in this case.
- Key Factor: This scenario is more about avoiding bias rather than explaining the reasoning behind the m...
Author: Sara · Last updated Jul 19, 2026
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 expla...
To select the most appropriate solution for detecting bias and explaining predictions in an ML model for loan approvals, we must evaluate the services based on their capabilities to handle model fairness (bias detection) and interpretability (explainability).
A) Amazon SageMaker Clarify
Amazon SageMaker Clarify is specifically designed to address both bias detection and model explainability. It offers capabilities to detect bias in machine learning models by assessing if predictions are influenced by certain sensitive features (like gender, age, or ethnicity). Additionally, it provides model explainability tools, such as feature importance and SHAP (Shapley Additive Explanations) values, which help explain individual predictions by showing how features contribute to the outcome. These features make it ideal for the company’s requirements in detecting bias and explaining predictions.
Reason for Selection: SageMaker Clarify is the most comprehensive and purpose-built solution for bias detection and explainability in machine learning models. It directly meets the company’s needs for both fairness and interpretability.
B) Amazon SageMaker Data Wrangler
Amazon SageMaker Data Wrangler is a tool for preparing, analyzing, and transforming data for machine learning. It allows data preprocessing and feature engineering, but it is not specifically designed for bias detection or model explainability. While it can help with data handling and preparing datasets, it does not offer tools to detect bias in model predictions or explain those predictions after model deployment.
Rejection Reason: While SageMaker Data Wrangler is useful for data preparatio...
Author: Ravi Patel · Last updated Jul 19, 2026
A company has developed a generative text summarization model by using Amazon Bedrock. The company will use Amazon Bedrock automatic model evaluation capabilities.Wh...
When evaluating the accuracy of a generative text summarization model using Amazon Bedrock’s automatic model evaluation capabilities, the most appropriate metric is Option C: BERTScore.
Analysis of each option:
1. A) Area Under the ROC Curve (AUC) score
- The AUC score is commonly used for evaluating binary classification models, specifically in terms of distinguishing between two classes (positive vs. negative). Since text summarization is not a binary classification task but rather a generation task, AUC is not an appropriate metric for assessing the accuracy of a generative text model.
- Effort: Low, but irrelevant for this task.
- Time: Low, but it won't provide meaningful insights into summarization performance.
- Cost: Low, but ineffective for summarization evaluation.
2. B) F1 score
- The F1 score is a harmonic mean of precision and recall, typically used for classification tasks where you need to measure how well a model balances precision and recall. While it could be used in some text summarization contexts (e.g., for extractive summarization), it’s not the best metric for generative summarization, where fluency, coherence, and semantic quality are more important than exact matches of words or phrases.
- Effort: Medium, but not ideal for generative text summarization.
- Time: Medium, and does not fully reflect the quality of generative summaries.
- Cost: Medium, but F1 score doesn’t capture all aspects of generative model quality.
3. C) BERTScore
- BERTScore is specifically designed for evaluating text generation tasks like summarization. It le...
Author: Suresh · Last updated Jul 19, 2026
An AI practitioner wants to predict the classification of flowers based on petal length, petal width, sepal length, and s...
The goal of predicting the classification of flowers based on attributes like petal length, petal width, sepal length, and sepal width suggests that we are dealing with a classification problem. Let's evaluate each option to determine which algorithm is best suited:
A) K-nearest neighbors (k-NN)
- Reasoning: K-nearest neighbors (k-NN) is a classification algorithm that works by identifying the "k" closest data points to a given query point and assigning the class based on the majority of those neighbors. Since the task involves classifying flowers based on measurements (a classification problem with discrete labels), k-NN is a natural fit for this scenario.
- Accepted because: k-NN is specifically designed for classification tasks, making it an ideal choice for predicting flower classifications based on the given features.
B) K-mean
- Reasoning: K-means is a clustering algorithm, not a classification algorithm. It groups data into "k" clusters based on similarity, but it doesn't provide a class label for each data point. Since we are looking to classify flowers into categories (not group them based on similarity), K-means is not the right choice.
- Rejected because: K-means is for clustering, not classifi...
Author: Leah Davis · Last updated Jul 19, 2026
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 t...
To determine the best solution for encrypting model artifacts using a company-managed encryption key in Amazon Bedrock, let's evaluate each of the available options based on their suitability for encryption, effort, cost, and the specific requirements mentioned.
A) AWS Key Management Service (AWS KMS)
AWS KMS is designed for managing encryption keys for various AWS services and resources. It allows customers to create and manage their own encryption keys (customer managed keys, CMKs) and use them to encrypt data, including model artifacts. KMS integrates seamlessly with Amazon Bedrock, making it an ideal solution for managing encryption keys and encrypting data during model customization jobs. KMS offers fine-grained access control, key rotation, and logging, which would be critical for security compliance.
Reason for Selection: AWS KMS is the best option for managing encryption keys and ensuring that model artifacts are securely encrypted using a company-managed key. It is cost-effective and integrates directly with AWS services, including Amazon Bedrock.
B) Amazon Inspector
Amazon Inspector is a service for assessing the security posture of AWS resources, focusing on vulnerabilities and compliance. While it helps identify security issues within your infrastructure, it does not directly handle encryption or key management. It would be valuable for ensuring the security of the overall environment, but it is not relevant to the specific task of encrypting model artifacts.
Rejection Reason: Amazon Inspector is n...
Author: Siddharth · Last updated Jul 19, 2026
A company wants to use large language models (LLMs) to produce code from natural language code comment...
To determine the most suitable LLM feature for generating code from natural language comments, we need to focus on the task described: translating natural language code comments into code. Let’s analyze each feature based on its capabilities, effort, time, cost, and alignment with the requirements.
1. Text summarization:
- Relevance: Text summarization is used to condense a large body of text into a shorter version, retaining key information. This feature is generally used for tasks like summarizing articles, reports, or documents, not for generating code from comments.
- Reason for rejection: This feature doesn’t align with the task of converting natural language comments into executable code.
- Scenario: Summarization is ideal for applications like summarizing news articles or long documents but doesn't fit the coding task.
2. Text generation:
- Relevance: Text generation refers to the ability of an LLM to produce new text based on a prompt. In the context of the task, it can generate code from natural language input (like comments). This feature can take the natural language comment and produce corresponding code, which matches the requirement of the company.
- Reason for selection: Text generation is a direct fit for transforming natural language into code. It allows the model to create new content (code) based on the input comment, making it highly suited for tasks that involve code synthesis from natural language.
- Scenario: This feature is commonly used in applications like GitHub Copilot, where natural language comments or descriptions are converted into code.
3. Text c...
Author: John · Last updated Jul 19, 2026
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 ...
In this scenario, the company wants to ensure that the output generated by the large language model (LLM) aligns with the provided examples of more readable, coherent text. The goal is to assess whether the LLM is improving the readability of the text, not necessarily focusing on speed or raw prediction performance.
A) Value of the loss function
The loss function measures the difference between the model's predictions and the actual target values during training. While the loss function helps guide the model during training, it doesn't directly assess the quality of the LLM's output in terms of readability or coherence. It’s more of an internal measure for model optimization, not an external metric that reflects how the generated text compares to the provided examples.
Reason for rejection: The loss function is essential during training but is not a suitable metric for evaluating how readable or coherent the output is in relation to the examples provided by the company.
B) Semantic robustness
Semantic robustness refers to how well a model maintains meaning and correctness when subjected to various changes, such as paraphrasing, noise, or distortions. While this is an interesting concept in NLP, the focus in this case is on coherence and readability, not how well the model retains meaning when altered. Semantic robustness typically evaluates how consistent and semantically accurate a model is across different transformations, rather than how readable and coherent the text is compared to the examples.
Reason for rejection: Semantic robustness is more about meaning preservation under transformations, and doesn't directly measure the readability or coherence of generated text in comparison to example texts.
C) Recall-Oriented Understudy for Gisting Evaluation (ROUGE) score
The ROUGE score is a set of metrics used to evaluate the quality of summaries or generated text by comparing it to reference text. It focuses on recall-based metrics, measuring the overlap of n-grams, word sequences, and word pairs between the generated output an...
Author: Liam · Last updated Jul 19, 2026
A company notices that its foundation model (FM) generates images that are unrelated to the prompts. The company wants to modify the prompt techniques to ...
In the given scenario, the company is facing an issue where its foundation model (FM) generates images unrelated to the prompts. To address this, the goal is to adjust the prompt techniques to reduce the occurrence of irrelevant images. Let's break down the options:
Option A: Use zero-shot prompts
- Efficiency: Zero-shot prompts are prompts where the model is asked to perform a task without being explicitly trained for it. While zero-shot prompts can be used in many contexts, they are not tailored to the specific task or to guide the model's behavior in generating relevant images. The model may struggle to understand the specific intent behind the prompt, leading to irrelevant outputs.
- Cost: Zero-shot prompting may require more computational effort for the model to infer the task, as it has no prior context or fine-tuning.
- Time: Zero-shot prompts could result in less accurate results, as the model might need more time to "figure out" the task without sufficient guidance.
- Key Factor: Zero-shot prompts are more suited for tasks where no specific prior knowledge or refinement is necessary, but they don’t directly help in preventing unrelated images when the model is already generating irrelevant results.
Option B: Use negative prompts
- Efficiency: Negative prompts specify what should not be included in the generated image. This is a highly effective approach to guide the model away from generating irrelevant images. By explicitly telling the model what to avoid, negative prompts help in focusing the generation on the relevant aspects of the prompt.
- Cost: The use of negative prompts adds some complexity but is still computationally efficient and can help in significantly improving the quality of generated content by restricting certain outputs.
- Time: Negative prompts reduce the likelihood of generating irrelevant or unrelated images, thus saving time by minimizing the need for post-generation edits or multiple attempts.
- Key Factor: Negative prompts are very effective when the issue is generating unrelated images. They guide the model to avoid certain features or elements, resulting i...
