Q1.A development team is carrying out interactive experimentation during the model-building phase.
At the same time, it also runs scalable jobs for model training.
The team needs to be able to scale compute resources up as needed while keeping costs as low as possible.
Even within the same workspace, the compute behavior required differs by workload.
You need to configure the appropriate compute target for each workload.
Which compute target should you choose?
To answer, move the appropriate compute target to the matching workload type.
Each compute target may be used once, more than once, or not at all.
You may need to move the split bar between panes or scroll to review the content.
Drag each option to the matching requirement’s drop zone, or tap to place it. Click a filled zone to send it back. (The same option may be used more than once.)
Azure Databricks cluster
Azure Kubernetes Service
Azure virtual machine
Azure Machine Learning compute cluster
Azure Machine Learning compute instance
Show answer
An Azure Machine Learning compute instance is suited to interactive experimentation.
This is because it is a fully managed personal workstation for machine learning development, equipped with tools such as Jupyter Notebook.
An Azure Machine Learning compute cluster is suited to scalable training jobs.
It automatically adds nodes when a job is submitted, and if you set the minimum node count to 0 it scales down to zero nodes when idle, which keeps costs low.
The sample answer in the image shows an Azure Databricks cluster, but current official guidance recommends the compute cluster, which is Azure Machine Learning’s managed training target.
What is an Azure Machine Learning compute instance
Create an Azure Machine Learning compute cluster
Q2.A machine learning model is deployed to an Azure Machine Learning production environment and continuously serves predictions to a business application.
The model was trained on a historical dataset representing the input patterns expected at the time of deployment.
The responsible team needs to reliably meet the following points.
• Be able to detect changes that occur in the distribution of the input data.
• Automatically initiate an appropriate action when a predefined threshold is exceeded.
You need to configure model monitoring to satisfy these requirements.
Which configuration should you choose for each requirement?
To answer, select the appropriate options in the answer area.
Show answer
Data drift monitoring compares baseline historical data with the production input data and detects whether the distribution of features is changing over time.
Increasing the compute size only adjusts processing capacity and is not a capability that captures statistical changes in the input data.
Manual retraining is an action taken after detection, not the monitoring itself.
When a monitoring metric exceeds a configured threshold, you configure an alert to notify the responsible personnel or a downstream automation process.
Note that the legacy dataset data drift monitoring (v1) has been retired, and the data drift signal of model monitoring (v2) is now used.
Azure Machine Learning model monitoring
Monitor the performance of models deployed to production
Q3.You deploy a new model version to a managed online endpoint.
You need to test this model with 10% of the total traffic and automatically roll back if latency or the error rate rises above a threshold.
What should you configure?
Show answer
With a managed online endpoint, you can assign traffic ratios to multiple deployments within the same endpoint.
By allocating 10% to the new version and 90% to the existing version, you can achieve a canary release that safely validates the model with real requests.
In addition, you monitor latency and failure rate with Azure Monitor and raise an alert when a threshold is exceeded.
For automatic rollback, you configure the alert’s action group to trigger a service such as Azure Functions or Logic Apps and shift traffic back to the existing model.
A separate endpoint cannot use the built-in traffic split, and a batch endpoint is not intended for validating real-time inference.
Deploy and score a machine learning model by using an online endpoint
Monitor online endpoints
Q4.In a retrieval-augmented generation (RAG) solution, relevant content is not consistently retrieved from the knowledge source, resulting in incomplete answers.
You need to improve the accuracy of the RAG without changing the embedding model currently in use.
Operating costs must be kept to a minimum.
Select two actions you should perform.
Each correct answer forms part of the solution.
Show answer
Adjusting the chunk size and overlap to match the document structure prevents important information from being split at chunk boundaries and prevents meaning from being diluted in chunks that are too large.
In addition, a reranker reorders the candidates obtained from the initial vector search by semantic relevance and pushes noisy candidates lower.
Both are methods that improve search accuracy through limited re-evaluation without changing the embedding model.
Raising the token limit does not directly improve search quality, and changing the vector length may require compatibility with the embedding model and re-indexing.
Chunk large documents for RAG and vector search in Azure AI Search
Semantic ranking in Azure AI Search
Q5.A data science team has completed multiple training runs within a single experiment using MLflow.
The team wants to save the selected model to Azure Machine Learning so it can later be version-controlled and deployed.
The model needs to be centrally version-controlled so it can be reused across multiple environments.
You need to version-control the trained model.
Select two actions you should perform.
Each correct answer forms part of the solution.
Show answer
First, identify the model artifacts logged to the MLflow run and retrieve the files needed for registration, such as the model itself, its dependencies, and the MLmodel metadata.
Next, when you register the model in Azure Machine Learning, it becomes a managed asset with a name and version and can be deployed to online or batch endpoints.
Tagging the experiment helps with searching and organizing, but it does not create a version of the model itself.
Exporting to a local location does not provide central version control either.
MLOps model management with Azure Machine Learning
Share models, components, and environments across workspaces by using registries
Q6.You are training a machine learning model with Azure Machine Learning.
You use HyperDrive for hyperparameter tuning.
In past model training and tuning runs, many models showed similar performance.
You need to choose an early termination policy that satisfies the following requirements.
• When evaluating the current run, consider the performance of all past runs.
• Avoid comparing the current run only with the single best-performing run so far.
Which two early termination policies should you use?
Each correct answer forms part of the solution.
Show answer
The median stopping policy evaluates the current run against the running average of the primary metric reported across all runs, so it can consider the performance of all past runs.
The truncation selection policy compares all runs at each evaluation interval and terminates the runs that fall into a specified lowest percentage by performance.
Therefore, it is not a method that uses only the single best run as the baseline.
The bandit policy, on the other hand, compares the current run with the best run and terminates it if it is outside the allowed range, so it does not fit the stated requirements.
Tune hyperparameters for your model (v2)
sweep Package
Q7.A team is deploying a machine learning model to a production inference endpoint in Azure Machine Learning.
The team wants a way to safely validate a new model version without affecting existing consumers.
You need to propose a deployment strategy that can test the new model version in a controlled manner.
What should you configure?
Show answer
A blue-green deployment is appropriate, in which the existing blue deployment and a new green deployment run side by side within the same managed online endpoint and only a portion of the traffic is allocated to the new version.
For example, if you allocate 90% to the existing deployment and 10% to the new model, you can validate performance with real requests while limiting the impact on consumers.
Even if a problem occurs, you can quickly recover by shifting traffic back to the existing deployment.
If you want to validate without returning the new model’s responses to consumers, you can also use traffic mirroring, which duplicates a portion of the requests while returning responses only from the existing model.
Merely registering the model version does not validate it with production traffic, and a separate endpoint cannot use the built-in traffic control within a single endpoint.
Perform a safe rollout of a new deployment for real-time inference
Online endpoints for real-time inference
Q8.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 goal.
Some questions might have more than one correct solution, while others might have none.
After you answer a question in this section, you cannot return to it.
As a result, these questions do not appear on the review screen.
You manage an Azure Machine Learning workspace.
A 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 this machine learning model.
When submitting the script as a training job, you need to specify a command that passes the path to the dataset as a parameter value.
As the solution, you specify the following command.
python train.py –training_data training_data
Does this solution meet the goal?
Show answer
When passing a command job’s input data to a script, you must reference the input name defined in the job’s inputs by using the Azure Machine Learning expression ${{inputs.training_data}}.
An appropriate command example is “python script.py –training_data ${{inputs.training_data}}”; the command shown passes the literal string training_data instead of the input expression and runs train.py instead of the specified script.py, so it does not meet the goal.
At runtime, ${{inputs.training_data}} is replaced with the actual local path where the input data is mounted or downloaded.
In the current SDK v2, you use input types such as uri_file for a single file, uri_folder for a folder, and MLTable for tabular data.
Access data in a job
CLI (v2) core YAML syntax
Q9.You use Azure Machine Learning to deploy a model as a real-time web service.
You need to author the service’s entry script so that the model is loaded when the service starts and is used to score newly received data.
Which functions should the script include?
To answer, move the appropriate functions to the correct actions.
Each function may be used once, more than once, or not at all.
You may need to move the split bar between panes or scroll to review the content.
Drag each option to the matching requirement’s drop zone, or tap to place it. Click a filled zone to send it back. (The same option may be used more than once.)
init()
score()
predict()
run()
Show answer
The init() function is called only once when the inference container starts, and it performs initialization such as loading the registered model and storing it in a global variable.
Because the model does not have to be loaded on every request, inference latency is reduced.
The run() function is called each time an inference request arrives at the online endpoint; it receives the input data, performs prediction or scoring with the model, and returns the result.
main(), score(), and predict() can be defined as optional helper functions, but the entry points required in a standard scoring script for online inference are init() and run().
Debug scoring scripts by using the Azure Machine Learning inference HTTP server
Deploy and score a machine learning model by using an online endpoint
Q10.You have a Microsoft Foundry project that includes a connected Azure OpenAI Service model.
You have a set of text files stored locally on your computer.
You need to set up a flow that generates responses based on the content of these local files.
You need to implement the solution.
Which four actions should you perform, and in which order?
To answer, move the appropriate actions from the list of actions to the answer area and arrange them in the correct order.
Drag (or tap) the required number of items from “Options” on the left to the “Answer area” on the right. Within the answer area, drag to reorder.
- Create a flow.
- Create a data asset.
- Create a connection.
- Create a Foundry Search resource.
- Create a vector index.
- Drag here
Show answer
First, prepare the Foundry Search resource that serves as the search foundation, then create a vector index that ingests the local files and performs chunking, embedding, and indexing.
The vector index enables you to search for file content that is semantically close to the user’s question.
Next, create a connection so that the project and flow can use that index.
Finally, create a flow and configure the RAG processing that passes the relevant information retrieved by search into the prompt of the Azure OpenAI model.
Creating a data asset is one available option, but it is not included in the four-step procedure presented in this question.
Create and consume a vector index in the Microsoft Foundry portal (classic)
Add a new connection to a project