17 free Microsoft AI-300 practice questions from the CertStash pack of 170, each with the correct answer and a full explanation. The rule is the same for every exam: ten free on the product page, 17 free here, all 170 in the pack, and each step includes the one before it.
Get all 170 questions (US$39) · Download the first 10 as a PDF
Question 1
An organization maintains separate Azure Machine Learning workspaces for development and production. Both environments must use the same validated assets without duplicating them. Assets must be shared across workspaces while maintaining centralized governance and version control. You need to enable reuse of assets across workspaces without copying them. What should you do?
Show answer and explanation
Correct answer: D. Publish the asset to an Azure Machine Learning registry.
An Azure Machine Learning registry enables organizations to publish and share assets (models, components, environments, datasets) across multiple workspaces without duplication. A registry provides centralized governance, version control, and asset reusability across development and production environments while maintaining a single source of truth. This is the designed mechanism for cross-workspace asset sharing in Azure ML.
Why the other options are wrong
- A. Git integration syncs code repositories but does not provide asset-level sharing or governance across workspaces.
- B. Publishing as a pipeline component only makes the asset available within the current workspace's pipeline ecosystem, not across separate workspaces.
- C. A shared environment only enables runtime dependency sharing, not the broader asset reuse and governance required for validated assets across workspaces.
Question 2
An Azure Machine Learning workspace processes sensitive training data. The workspace must NOT be accessible from the public internet. You need to restrict network access. Which configuration should you implement?
Show answer and explanation
Correct answer: B. Private endpoints
Private endpoints create private network connections to Azure Machine Learning workspaces, ensuring they are not exposed to the public internet. By implementing private endpoints, network traffic flows through private IP addresses within a virtual network rather than over public routes, providing the complete isolation required for sensitive data processing.
Why the other options are wrong
- A. Azure Firewall rules control traffic between networks but do not prevent public internet exposure; the workspace would still be internet-accessible.
- C. Network security groups control traffic within virtual networks but do not isolate a workspace from the public internet.
- D. Service endpoints provide controlled access through Microsoft's backbone network but still allow public internet accessibility to the workspace.
Question 3
A team is experimenting with traditional models for a classification workflow in Azure Machine Learning. The team requires a consistent way to manage assets that are created during experimentation. You need to ensure that artifacts can be reused and governed across projects. Which asset should you register?
Show answer and explanation
Correct answer: A. Model
Registering a model in Azure Machine Learning creates a versioned, governed asset that can be tracked, reused, and deployed across projects. Model registration provides consistent management of trained classification models, enables version control, supports lineage tracking, and facilitates deployment governance, directly addressing the need for artifact reuse and governance across experimental projects.
Why the other options are wrong
- B. Components are reusable building blocks for pipelines but are better suited for encapsulating steps or logic rather than managing trained model artifacts.
- C. Environments manage runtime dependencies and packages but do not serve as the primary mechanism for registering and governing trained model artifacts.
- D. Pipelines are workflows that orchestrate steps but are not the asset type for registering and governing individual trained models from experimentation.
Question 4
Note: This question is part of a series of questions that present the same scenario. Each question in the series contains a unique solution that might meet the stated goals. Some question sets might have more than one correct solution, while others might not have a correct solution. After you answer a question in this section, you will NOT be able to return to it. As a result, these questions will not appear on the review screen. You manage an Azure Machine Learning workspace. The Python script named script.py reads an argument named training_data. The training_data argument specifies the path to the training data in a file named dataset1.csv. You plan to run the script.py Python script as a command job that trains a machine learning model. You need to provide the command to pass the path for the dataset as a parameter value when you submit the script as a training job. Solution: python train.py --training_data training_data Does the solution meet the goal?
Show answer and explanation
Correct answer: B. No
The solution does not meet the goal. The command provided contains an error: it references 'train.py' instead of 'script.py', and it passes 'training_data' as a literal string value rather than providing the actual path to dataset1.csv. The command should reference the correct script file and properly pass the dataset path as an argument value.
Question 5
Note: This question is part of a series of questions that present the same scenario. Each question in the series contains a unique solution that might meet the stated goals. Some question sets might have more than one correct solution, while others might not have a correct solution. After you answer a question in this section, you will NOT be able to return to it. As a result, these questions will not appear on the review screen. You manage an Azure Machine Learning workspace. The Python script named script.py reads an argument named training_data. The training_data argument specifies the path to the training data in a file named dataset1.csv. You plan to run the script.py Python script as a command job that trains a machine learning model. You need to provide the command to pass the path for the dataset as a parameter value when you submit the script as a training job. Solution: python script.py --trainingdata ${{inputs.training_data}} Does the solution meet the goal?
Show answer and explanation
Correct answer: B. No
The solution does not meet the goal. While the command correctly references 'script.py' and uses proper parameter passing syntax, the argument name is incorrect: the script expects an argument named 'training_data' but the command passes '--trainingdata' (without the underscore). Additionally, the syntax '${{inputs.training_data}}' requires that training_data be properly defined as an input parameter in the job configuration, not just referenced in the command. The command alone does not establish this connection without proper job definition.
Question 6
You manage an Azure Machine learning workspace. You develop a machine learning model. You must deploy the model to use a low-priority VM with a pricing discount. You need to deploy the model. Which compute target should you use?
Show answer and explanation
Correct answer: B. Azure Machine Learning compute clusters
Azure Machine Learning compute clusters support low-priority VMs with pricing discounts through the use of spot instances. This is the only compute target option that provides the required combination of low-priority VM capabilities with the pricing discount benefit. Compute clusters can be configured to use low-priority VMs at a reduced cost while maintaining scalability for model deployment.
Why the other options are wrong
- A. Azure Container Instances do not support low-priority VMs or provide pricing discounts.
- C. Local deployment is for development environments only and does not use Azure compute resources.
- D. Azure Kubernetes Service is typically used for production workloads but is not the primary mechanism for utilizing low-priority VMs with pricing discounts.
Question 7
A team manages an Azure Machine Learning workspace where they deploy models to online endpoints. The team needs to introduce a new version of a model to production without disrupting existing users. The team must validate the new version before full rollout. You need to reduce risk during deployment. What should you do?
Show answer and explanation
Correct answer: B. Split traffic between deployments.
Splitting traffic between deployments implements a canary deployment pattern that reduces risk by allowing the new model version to serve a portion of production traffic while the existing version continues handling the majority. This approach enables validation of the new version's performance in production without disrupting all users, and allows for rollback if issues arise. The team can gradually increase traffic to the new version as confidence builds.
Why the other options are wrong
- A. Batch endpoints are for asynchronous, offline inference and do not provide online serving for user traffic.
- C. Replacing the existing endpoint immediately exposes all users to the new version without validation, creating maximum risk.
- D. Routing all traffic immediately to the new deployment without validation bypasses the validation requirement and maximizes disruption risk.
Question 8
You have a deployment of an Azure OpenAI Service base model. You plan to fine-tune the model. You need to prepare a file that contains training data. Which file format should you use?
Show answer and explanation
Correct answer: C. JSONL
Azure OpenAI Service fine-tuning requires training data in JSONL (JSON Lines) format, where each line contains a single JSON object representing a training example. This format is the standard for OpenAI fine-tuning jobs and is the only format accepted by the Azure OpenAI Service fine-tuning API.
Why the other options are wrong
- A. CSV format is not supported for Azure OpenAI Service fine-tuning operations.
- B. TSV format is not supported for Azure OpenAI Service fine-tuning operations.
- D. JSON format (with separate objects or arrays) is not the correct format; JSONL (JSON Lines) with one object per line is required.
Question 9
You have a deployment of an Azure OpenAI Service base model. You plan to fine-tune the model. You need to prepare a file that contains training data for multi-turn chat. Which file encoding method should you use?
Show answer and explanation
Correct answer: C. UTF-8
UTF-8 is the standard encoding method for fine-tuning data files in Azure OpenAI Service. UTF-8 is Unicode-compatible, supports all character sets including multi-language support for global datasets, and is the industry standard for text encoding in machine learning and AI applications. All training data files must be encoded in UTF-8 format.
Why the other options are wrong
- A. ISO-8859-1 is a legacy encoding with limited character support and is not the standard for modern AI training data.
- B. UTF-16 is inefficient for text data and is not the standard encoding used by Azure OpenAI Service.
- D. ASCII is limited to basic English characters and cannot properly encode multi-turn chat data with varied languages and special characters.
Question 10
You are fine-tuning a base language model to analyze customer feedback. You label examples of support tickets. You must improve classification accuracy by configuring and fine-tuning the base model in Microsoft Foundry. You need to configure and run fine-tuning. What should you do first?
Show answer and explanation
Correct answer: D. Format the dataset as a JSONL file with prompt-completion pairs and upload the file.
Formatting the dataset as a JSONL file with prompt-completion pairs is the essential first step before any fine-tuning can occur. This data preparation step is prerequisite to all subsequent fine-tuning activities in Microsoft Foundry. Without properly formatted training data, the fine-tuning job cannot be configured or executed, making this the logical first action.
Why the other options are wrong
- A. Generating prompt templates with prompt flow is a variant testing activity that occurs after the base model is fine-tuned on prepared data.
- B. Deploying the base model to an endpoint is unnecessary before fine-tuning; fine-tuning can be configured and executed directly from the base model.
- C. Enabling tracing is a monitoring activity that occurs after fine-tuning begins, not a prerequisite configuration step.
Question 11
A team manages an Azure Machine Learning workspace and deploys a model to an endpoint. A deployed online endpoint shows inconsistent response times during periods of high traffic. You need to identify potential performance degradation. Which three metrics should you monitor? Each correct answer presents part of the solution. Choose three. NOTE: Each correct selection is worth one point.
Show answer and explanation
Correct answer: B, C, E. Requests per minute; Connections active; Request latency
To identify performance degradation during high traffic, you must monitor requests per minute to understand traffic volume, connections active to see resource utilization and connection pooling issues, and request latency to directly measure response time performance. These three metrics together provide visibility into traffic patterns, resource constraints, and actual performance degradation symptoms.
Why the other options are wrong
- A. Feature count is a static property of the model schema and does not change with traffic patterns or performance issues.
- D. Dataset size used during training is not a runtime metric and does not indicate current performance issues with the deployed endpoint.
Question 12
A data science team trains a classification model that predicts loan approval outcomes. Before registering the model, the team must ensure the following: Predictions must not disproportionately impact protected groups. Prediction errors can be evaluated across different data segments. You need to assess whether the model meets Responsible AI expectations. Which two approaches should you use? Each correct answer presents part of the solution. Choose two. NOTE: Each correct selection is worth one point.
Show answer and explanation
Correct answer: D, E. Evaluate feature importance for prediction transparency; Analyze error rates across defined demographic cohorts.
To ensure predictions do not disproportionately impact protected groups, you must evaluate feature importance for prediction transparency to understand what drives predictions, and analyze error rates across defined demographic cohorts to detect disparate impact. Measuring performance differences across demographic groups is essential for fairness assessment, and feature importance provides transparency into model behavior for different groups.
Why the other options are wrong
- A. Analyzing error rates across only the global cohort masks disparate performance across demographic groups and fails to detect fairness issues.
- B. Endpoint latency is a performance metric unrelated to fairness or bias assessment.
- C. Inference schema compatibility addresses technical validation, not fairness or responsible AI expectations.
Question 13
You create a binary classification model. You use the Fairlearn package to assess model fairness. You must eliminate the need to retrain the model. You need to implement the Fairlearn package. Which algorithm should you use?
Show answer and explanation
Correct answer: D. fairlearn.postprocessing.ThresholdOptimizer
The ThresholdOptimizer is a post-processing algorithm in Fairlearn that adjusts decision thresholds after model training to improve fairness metrics without retraining the model. This directly meets the requirement to eliminate the need for retraining while still addressing fairness concerns in a binary classification model.
Why the other options are wrong
- A. ExponentiatedGradient is a reductions-based algorithm that requires model retraining during the optimization process.
- B. CorrelationRemover is a preprocessing algorithm that modifies input data before training, not applicable after model creation.
- C. GridSearch is a reductions-based algorithm that requires iterative model retraining to find fair solutions.
Question 14
A team is working in Microsoft Foundry to test and compare large language model (LLM) prompt variants in a development environment. The team requires consistent inputs to evaluate prompt variants without relying on live user traffic. You need to create a controlled evaluation of input data. Which action should you perform first?
Show answer and explanation
Correct answer: A. Generate synthetic interaction data.
Generating synthetic interaction data is the first action to create a controlled evaluation environment with consistent inputs. Synthetic data allows the team to systematically test and compare prompt variants without depending on live user traffic, ensuring reproducible and controlled evaluation conditions. This provides the dataset foundation needed for all subsequent evaluation activities.
Why the other options are wrong
- B. Configuring content filters is a safety mechanism applied after evaluation setup, not a prerequisite for creating evaluation data.
- C. Applying a blocklist is a content filtering mechanism that occurs after evaluation infrastructure is established, not a first action.
- D. Enabling observability metrics is a monitoring activity that follows evaluation setup rather than being a prerequisite for creating controlled evaluation inputs.
Question 15
Note: This question is part of a series of questions that present the same scenario. Each question in the series contains a unique solution that might meet the stated goals. Some question sets might have more than one correct solution, while others might not have a correct solution. After you answer a question in this section, you will NOT be able to return to it. As a result, these questions will not appear on the review screen. You manage an Azure Machine Learning workspace. The Python script named script.py reads an argument named training_data. The training_data argument specifies the path to the training data in a file named dataset1.csv. You plan to run the script.py Python script as a command job that trains a machine learning model. You need to provide the command to pass the path for the dataset as a parameter value when you submit the script as a training job. Solution: python script.py dataset1.csv Does the solution meet the goal?
Show answer and explanation
Correct answer: B. No
When passing arguments to a Python script in Azure Machine Learning command jobs, you must use the proper argument flag syntax. The script expects a named argument called 'training_data', so the command should be 'python script.py --training_data dataset1.csv' or similar flag-based syntax, not simply 'python script.py dataset1.csv'. Positional arguments without flags will not correctly map to the named parameter the script is expecting.
Why the other options are wrong
- A. Simply passing the filename as a positional argument does not properly pass it as the named 'training_data' parameter that the script requires.
Question 16
A data science team completes multiple training runs within an experiment by using MLflow. The team wants to store a selected model in Azure Machine Learning so that it can be versioned and deployed later. The model must be versioned centrally for reuse across environments. You need to version the trained model. Which two actions should you perform? Each correct answer presents part of the solution. Choose two. NOTE: Each correct selection is worth one point.
Show answer and explanation
Correct answer: A, B. Locate and capture the model artifacts from the outputs of the training run; Register the model in the Azure Machine Learning workspace.
To version a trained model centrally in Azure Machine Learning, you must first locate and capture the model artifacts from the training run outputs, then register the model in the Azure Machine Learning workspace. Registration creates a versioned entry in the workspace's model registry, enabling tracking, deployment, and reuse across environments. These two steps together complete the process of centralizing the model for version control and deployment.
Why the other options are wrong
- C. Tagging the training experiment with a name documents the experiment but does not accomplish model versioning or registration.
- D. Exporting model files to local storage removes centralization and defeats the purpose of using Azure Machine Learning as the central versioning system.
Question 17
A team deploys a model to a real-time endpoint in Azure Machine Learning. You deploy some updates to the endpoint. The endpoint returns errors after the new deployment is released. You need to restore the service as quickly as possible. What should you do first?
Show answer and explanation
Correct answer: A. Roll back traffic to the previous deployment.
When a deployment causes errors, the fastest way to restore service is to roll back traffic to the previous deployment. Azure Machine Learning endpoints support traffic-weighted deployments, allowing immediate traffic redirection to the stable previous version without requiring deletion, redeployment, or infrastructure changes.
Why the other options are wrong
- B. Deleting and redeploying the endpoint is time-consuming and unnecessary when a previous stable version exists; rollback is faster.
- C. Changing authentication type addresses security configuration, not the functional errors caused by the deployment update.
- D. Increasing compute size addresses capacity, not the functional errors introduced by the code or model in the new deployment.
That was 17 of 170.
The full Microsoft AI-300 pack has all 170 questions, each with the answer, the explanation and why the other options are wrong, plus a questions-only copy for timed runs. US$39, paid once, with free monthly updates and a pass-or-your-money-back guarantee.
Not ready yet? Get the first 20 free as one PDF by email. Get the Half Pack →