View Full Microsoft AI-901 Exam Dumps and Practice Test Dumps

Question 301. Which Azure AI Foundry feature can help developers experiment with different models and prompts?

1. Playground
2. Azure DNS
3. Azure Files
4. Network Security Group

Correct Answer: 1. Playground

Explanation:

The Azure AI Foundry playground provides an interactive environment for experimenting with generative AI models and prompts. Developers can test instructions, provide sample inputs, adjust settings, and observe model responses without immediately building a complete application. This makes it useful during the early stages of solution development and prompt engineering. Teams can compare how different instructions or models behave for a particular scenario and identify potential limitations before implementation. However, successful playground testing does not replace formal evaluation. Applications should still be tested with representative data, safety scenarios, and performance requirements before being deployed to production.

Question 302. What is zero-shot prompting?

1. Providing many labeled examples before the task
2. Asking a model to perform a task without providing task-specific examples
3. Training a new model from scratch
4. Removing the model’s instructions

Correct Answer: 2. Asking a model to perform a task without providing task-specific examples

Explanation:

Zero-shot prompting asks a generative AI model to perform a task without giving it specific examples of the expected input and output. Instead, the prompt describes the task directly. For example, a developer might ask a model to classify a customer message as either positive or negative while providing no previous classification examples. Zero-shot prompting can be useful when the task is relatively straightforward or when creating examples would add unnecessary complexity. The quality of the result still depends on how clearly the task is described and whether the selected model is capable of performing the requested task.

Question 303. What is the main purpose of few-shot prompting?

1. To provide examples that demonstrate the expected task or output
2. To delete training data
3. To disable model responses
4. To create a network firewall

Correct Answer: 1. To provide examples that demonstrate the expected task or output

Explanation:

Few-shot prompting provides a generative AI model with a small number of examples showing how a task should be performed. These examples can demonstrate the expected format, classification categories, tone, or reasoning pattern without requiring the developer to retrain the underlying model. For example, a prompt might provide several customer messages with their corresponding categories and then ask the model to classify a new message. Carefully selected examples can improve consistency and help the model understand an application’s requirements. However, examples should be relevant and representative, and the resulting behavior should still be evaluated using additional test cases.

Question 304. Which situation is an example of data leakage during machine learning development?

1. Using information during training that would not be available when making real-world predictions
2. Splitting data into training and test datasets
3. Monitoring model performance
4. Removing duplicate records

Correct Answer: 1. Using information during training that would not be available when making real-world predictions

Explanation:

Data leakage occurs when information that would not legitimately be available at prediction time is included in the training process. This can cause a model to appear highly accurate during evaluation while performing poorly in real-world use. For example, if a model predicts whether a customer will cancel a service and the training data includes a field created after the cancellation occurred, the model may learn information that would not exist when the prediction is actually made. Preventing leakage requires careful feature selection, correct data splitting, and understanding when each piece of information becomes available in the real-world workflow.

Question 305. Why is it important to keep training and test data separate?

1. To provide an independent evaluation of model performance
2. To increase the number of model parameters
3. To prevent all model errors
4. To eliminate the need for validation

Correct Answer: 1. To provide an independent evaluation of model performance

Explanation:

Keeping training and test data separate allows developers to evaluate how a model performs on data that it did not use during learning. If the same examples are used for both training and final evaluation, performance measurements may be overly optimistic because the model has already seen those examples. A separate test dataset provides evidence about how well the model may generalize to unseen data. The test set should ideally represent the conditions expected in production. Maintaining proper separation also helps teams detect overfitting and compare different approaches more reliably during the development lifecycle.

Question 306. Which machine learning approach is most appropriate when historical examples already contain known output labels?

1. Supervised learning
2. Unsupervised learning
3. Reinforcement learning only
4. Random data generation

Correct Answer: 1. Supervised learning

Explanation:

Supervised learning uses training examples that include known labels or target values. The model learns a relationship between input features and the desired output so it can make predictions for new data. Classification and regression are common examples of supervised learning. For instance, a dataset containing customer information and a known churn outcome can be used to train a supervised classification model. The quality and representativeness of the labeled training data strongly influence model performance. Developers should also evaluate the trained model using appropriate validation and test data to determine whether it generalizes beyond the examples used for learning.

Question 307. Which machine learning approach is useful when the data does not have predefined labels and the goal is to discover patterns?

1. Supervised learning
2. Unsupervised learning
3. Text-to-speech
4. Optical character recognition

Correct Answer: 2. Unsupervised learning

Explanation:

Unsupervised learning is used when training data does not contain predefined target labels. Instead of learning to predict a known outcome, the algorithm attempts to discover patterns, structures, or relationships within the available data. Clustering is a common unsupervised learning technique because it can group similar records based on selected characteristics. Unsupervised learning can be useful for exploratory analysis, customer segmentation, and identifying previously unknown patterns. Because there are no predefined labels to directly measure against, interpreting the resulting patterns requires careful analysis and domain knowledge to determine whether the discovered structure is meaningful for the intended business scenario.

Question 308. What is reinforcement learning based on?

1. Learning through interactions and rewards or penalties
2. Only labeled historical records
3. Manual database indexing
4. Document translation

Correct Answer: 1. Learning through interactions and rewards or penalties

Explanation:

Reinforcement learning involves an agent interacting with an environment and learning which actions are useful based on feedback such as rewards or penalties. Over time, the agent attempts to develop a strategy that maximizes the accumulated reward. This differs from supervised learning, where examples contain known target outputs. Reinforcement learning can be applied to scenarios involving sequential decisions, where an action can influence future states and outcomes. The design of the reward function is particularly important because the system will optimize according to the feedback it receives. Poorly designed rewards can encourage behavior that does not match the intended objective.

Question 309. Which Azure service is designed to store large amounts of unstructured data such as images and documents?

1. Azure Blob Storage
2. Azure DNS
3. Azure Firewall
4. Azure Key Vault only

Correct Answer: 1. Azure Blob Storage

Explanation:

Azure Blob Storage is designed to store large amounts of unstructured data, including images, videos, documents, backups, and other files. AI applications frequently use Blob Storage as a location for training datasets, source documents, images, or other information required by machine learning and AI workloads. The storage service can integrate with other Azure services, allowing applications to retrieve and process stored content. Developers should still consider appropriate access controls, encryption, data retention, and organizational policies when storing information. Using a suitable storage service provides a scalable foundation for managing large collections of files used by AI solutions.

Question 310. What is the purpose of a data asset in Azure Machine Learning?

1. To provide a managed reference to data used by machine learning workflows
2. To replace the model itself
3. To create a firewall rule
4. To generate speech audio

Correct Answer: 1. To provide a managed reference to data used by machine learning workflows

Explanation:

A data asset in Azure Machine Learning provides a managed way to reference data used by machine learning workflows. Instead of treating every dataset as an unmanaged collection of files, teams can organize and reference data resources within the machine learning environment. This can improve consistency and make it easier for different jobs or experiments to use known datasets. Data assets can be useful for tracking which data is associated with particular workflows and supporting repeatable development processes. Teams should still maintain appropriate data governance, permissions, quality controls, and versioning practices to ensure that machine learning workloads use suitable information.

Question 311. Which metric measures the average absolute difference between predicted and actual values in a regression model?

1. Precision
2. Recall
3. Mean Absolute Error
4. Accuracy

Correct Answer: 3. Mean Absolute Error

Explanation:

Mean Absolute Error, or MAE, measures the average absolute difference between predicted values and actual values in a regression problem. For each prediction, the absolute error is calculated, and these errors are then averaged across the evaluated examples. Because the differences are converted to absolute values, positive and negative errors do not cancel each other out. MAE is expressed in the same units as the target variable, which can make it easier to interpret. A lower MAE generally indicates that predictions are closer to actual values, although the appropriate evaluation metric should always depend on the specific business problem.

Question 312. Which metric is commonly used to measure the average squared error in regression?

1. Mean Squared Error
2. Recall
3. Precision
4. F1 score

Correct Answer: 1. Mean Squared Error

Explanation:

Mean Squared Error, or MSE, calculates the average of the squared differences between predicted and actual values. Squaring the errors means that larger errors receive greater weight, so MSE can be particularly sensitive to significant prediction mistakes. It is commonly used as a regression evaluation metric and can be useful when larger errors are especially undesirable. However, because the errors are squared, the resulting measurement is expressed in squared units of the target variable, which may make it less intuitive to interpret than MAE. Developers should select evaluation metrics according to the consequences and requirements of the specific prediction problem.

Question 313. What does precision measure in a classification model?

1. The proportion of predicted positive cases that are actually positive
2. The total number of training records
3. The amount of storage used
4. The percentage of all inputs that are unique

Correct Answer: 1. The proportion of predicted positive cases that are actually positive

Explanation:

Precision measures how many of the cases predicted as positive are actually positive. It is calculated using true positives divided by the sum of true positives and false positives. Precision becomes especially important when false positive predictions are costly or disruptive. For example, if an application flags transactions as potentially fraudulent, a low precision value could result in many legitimate transactions being incorrectly flagged. Precision should generally be considered alongside recall because improving one can sometimes affect the other. The correct balance depends on the business scenario and the relative consequences of false positives and false negatives.

Question 314. What does recall measure in a classification model?

1. The proportion of actual positive cases correctly identified
2. The number of database records stored
3. The size of the model file
4. The number of features in a dataset

Correct Answer: 1. The proportion of actual positive cases correctly identified

Explanation:

Recall measures the ability of a classification model to identify actual positive cases. It is calculated as true positives divided by the total number of actual positive cases, which includes both true positives and false negatives. Recall is particularly important when failing to identify a positive case has significant consequences. For example, a monitoring system may prioritize identifying as many genuine incidents as possible even if that results in additional false alarms. Recall should not be considered in isolation because a model could achieve high recall while producing many false positives. Precision and other metrics should also be evaluated.

Question 315. What is the purpose of a validation dataset during model development?

1. To help evaluate and tune a model before final testing
2. To permanently store production predictions
3. To replace all training data
4. To encrypt model parameters

Correct Answer: 1. To help evaluate and tune a model before final testing

Explanation:

A validation dataset provides data that can be used during model development to evaluate different model configurations, features, or hyperparameters. It is separate from the training data and is useful for making development decisions before the final model is evaluated on a test dataset. Keeping a separate test set helps prevent repeated tuning against the final evaluation data. Validation is therefore an important part of the development lifecycle, especially when comparing approaches or identifying overfitting. The validation data should also be representative of the intended use case so that development decisions are based on realistic expectations.

Question 316. What is a hyperparameter in machine learning?

1. A configuration value selected before or during model training
2. A final prediction generated by the model
3. A user’s password
4. A document stored in Blob Storage

Correct Answer: 1. A configuration value selected before or during model training

Explanation:

Hyperparameters are configuration values that influence how a machine learning algorithm learns or operates but are generally not learned directly from the training examples in the same way as model parameters. Examples can include learning rate, tree depth, number of clusters, or other algorithm-specific settings. Developers can adjust hyperparameters to improve model performance, but those adjustments should be evaluated using appropriate validation data. Automated machine learning can help explore combinations of settings in supported scenarios. Hyperparameter tuning should be controlled carefully because repeatedly optimizing against the same validation data can also lead to overly optimistic development results.

Question 317. Which action can help address underfitting in a machine learning model?

1. Using a model that is better able to capture the relevant patterns
2. Removing all useful features
3. Reducing the training data to one example
4. Ignoring validation results

Correct Answer: 1. Using a model that is better able to capture the relevant patterns

Explanation:

Underfitting occurs when a model is too simple or insufficiently trained to capture important patterns in the data. It can result in poor performance on both training and unseen datasets. Depending on the situation, possible approaches include using a more suitable model, adding useful features, improving the training process, adjusting relevant hyperparameters, or providing better-quality training data. The appropriate solution depends on why the model is underfitting. Developers should compare training and validation performance to understand the problem before changing the model. Simply making a model more complex without evidence may introduce unnecessary complexity or create other issues.

Question 318. Which responsible AI principle focuses on ensuring an AI system performs appropriately across relevant groups and scenarios?

1. Fairness
2. Storage optimization
3. Network availability
4. Compression

Correct Answer: 1. Fairness

Explanation:

Fairness in AI involves considering whether a system provides appropriate and equitable performance across relevant groups and scenarios. A model may perform well overall while producing significantly different error rates for different populations. Evaluating performance across relevant groups can help identify such differences. Developers should use representative data, appropriate metrics, and domain knowledge when assessing fairness. Fairness is not simply a matter of achieving identical outcomes in every situation because different applications have different requirements and contexts. Instead, teams should identify relevant risks, evaluate them systematically, and apply suitable safeguards based on the intended use of the AI system.

Question 319. What is one benefit of using alerts in model monitoring?

1. They can notify teams when defined performance or operational conditions occur
2. They guarantee perfect model accuracy
3. They eliminate the need for testing
4. They automatically rewrite the training data

Correct Answer: 1. They can notify teams when defined performance or operational conditions occur

Explanation:

Monitoring alerts can notify responsible teams when predefined conditions or thresholds are reached. For an AI application, alerts might relate to service failures, unusual latency, data drift, changes in prediction behavior, or other operational and model-related indicators. Alerts allow teams to investigate potential problems sooner rather than relying entirely on manual observation. However, an alert itself does not determine the correct response or guarantee that a model is inaccurate. Teams need appropriate monitoring criteria, escalation procedures, and investigation processes. Alerting is therefore one component of a broader lifecycle for maintaining reliable AI applications after deployment.

Question 320. Which practice is most appropriate when an AI model will be used for a high-impact decision?

1. Deploying it without testing
2. Relying entirely on the model’s output
3. Applying appropriate evaluation, safeguards, and human oversight
4. Removing all access controls

Correct Answer: 3. Applying appropriate evaluation, safeguards, and human oversight

Explanation:

High-impact AI applications require careful consideration because incorrect or biased outputs can have significant consequences for people or organizations. Appropriate evaluation should test the model against representative scenarios, relevant performance measures, and known risks. Safeguards can include access controls, monitoring, content filtering, documentation, and defined escalation procedures. Human oversight may also be appropriate when decisions have significant consequences or when the model encounters uncertain or exceptional cases. The specific controls depend on the application and its risks, but the overall goal is to ensure that AI outputs are not treated as automatically correct and that responsible people can review important decisions.