I85PREP
⌘ K
AZ MLO study materials
AZ MLO study resources

AZ MLO practice questions

One question at a time. Build your AZ MLO understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Design and prepare a machine learning solution

A team is planning an Azure Machine Learning workspace. Which Azure resource must exist in the same region as the workspace to provide centralized secrets storage for datastores and compute?

  1. An Azure SQL Database server
  2. An Azure Container Registry instance
  3. An Azure Application Insights resource
  4. An Azure Key Vault instance Correct answer.

Explanation

The workspace uses an associated Key Vault to store secrets such as storage account keys and credentials referenced by datastores and compute targets. This association is created automatically when the workspace is created in the portal and must reside in the same region.

Question 2Design and prepare a machine learning solution

Which Azure Machine Learning compute type is best suited for running a training script that requires a GPU and should scale down to zero nodes when idle?

  1. Compute cluster Correct answer.
  2. Compute instance
  3. Serverless compute
  4. Attached Kubernetes cluster

Explanation

Compute clusters are designed for batch training jobs, support GPU SKUs, and can be configured with a minimum node count of zero so no cost accrues while idle. Compute instances are single-user development VMs that stay running until stopped.

Question 3Design and prepare a machine learning solution

What is the primary purpose of a datastore in Azure Machine Learning?

  1. To define a secure connection to a storage service that jobs can reference by name Correct answer.
  2. To store model evaluation metrics for comparison
  3. To schedule recurring pipeline runs
  4. To host the Docker images used by training environments

Explanation

A datastore encapsulates connection information and credentials for an Azure storage service so that scripts and jobs can reference it by name instead of embedding secrets. Data itself stays in the underlying storage account.

Question 4Design and prepare a machine learning solution

A data scientist needs an interactive notebook environment with preinstalled ML packages and persistent files, but does not need it to run unattended jobs. Which workspace resource should be provisioned?

  1. Serverless compute
  2. Inference cluster
  3. Compute instance Correct answer.
  4. Compute cluster

Explanation

A compute instance is a managed, single-user cloud workstation intended for interactive development such as notebooks and IDE sessions. It persists files in a workspace storage volume and comes with preconfigured environments.

Question 5Design and prepare a machine learning solution

When you register a dataset or data asset in Azure Machine Learning, what does the registration primarily provide?

  1. A guarantee that the data schema will never change
  2. A physical copy of the data inside the workspace
  3. Automatic encryption of the underlying storage account
  4. A named, versioned reference to data stored in an existing location Correct answer.

Explanation

Registering a data asset records metadata such as name, version, and the storage path or URI, giving teams a reusable and traceable reference. The bytes remain in the original storage location.

Question 6Design and prepare a machine learning solution

Which construct in Azure Machine Learning groups related runs so their metrics and outputs can be compared together?

  1. Model registry
  2. Environment
  3. Endpoint
  4. Experiment Correct answer.

Explanation

An experiment is the logical container that groups runs of a script so metrics, logs, and artifacts can be compared side by side. Each submission creates a run inside the experiment.

Question 7Design and prepare a machine learning solution

What does an Azure Machine Learning environment define?

  1. The retry policy for a pipeline step
  2. The network topology used by a managed online endpoint
  3. The Python packages, base image, and settings used to execute a job or deployment Correct answer.
  4. The access control list applied to a datastore

Explanation

An environment captures the runtime context-base Docker image, conda or pip dependencies, and environment variables-needed to reproduce a job or deployment. Environments are versioned and can be registered for reuse.

Question 8Design and prepare a machine learning solution

A team wants to compare several hyperparameter configurations automatically. Which Azure Machine Learning capability should they use?

  1. Managed batch endpoint
  2. Responsible AI dashboard
  3. AutoML forecasting
  4. Sweep job with a search space Correct answer.

Explanation

A sweep job runs a command job repeatedly over a defined search space using a sampling algorithm such as random or Bayesian search, optionally with early termination. This is the standard mechanism for hyperparameter tuning in Azure Machine Learning.

Question 9Design and prepare a machine learning solution

You need to run a training script that reads a large CSV from Azure Blob Storage without copying the file to the compute node first. Which configuration achieves this?

  1. Embed the CSV contents as a job environment variable
  2. Download the CSV into the script's working directory at submission time
  3. Register the blob container as a datastore and pass a URI input to the job Correct answer.
  4. Upload the CSV to the model registry before training

Explanation

URI-based data inputs let a job stream or read data directly from its storage location, avoiding a full local copy. Registering the container as a datastore supplies the credentials and path resolution.

Question 10Design and prepare a machine learning solution

A workspace must be created without public internet access, and compute must reach a storage account through a private endpoint. Which configuration is required?

  1. A managed virtual network with the workspace in private link mode Correct answer.
  2. A public workspace with firewall rules allowing all Azure services
  3. An attached Kubernetes cluster in a different region
  4. A compute instance with a public IP address

Explanation

Network isolation for Azure Machine Learning relies on private endpoints and, for managed compute, a managed virtual network configured in isolation mode. This keeps traffic to storage and other dependencies on private addresses.

Question 11Design and prepare a machine learning solution

A pipeline step must run only after a data preparation step succeeds, and the pipeline should re-run when the input data asset version changes. Which two features support this?

  1. Pipeline dependencies and data asset versioning Correct answer.
  2. Compute cluster idle scale-down and low priority nodes
  3. Endpoint traffic splitting and autoscaling
  4. Model registry tags and descriptions

Explanation

Pipeline steps declare dependencies so downstream steps wait for upstream success, and referencing a specific data asset version lets the pipeline detect changes and re-run. Together they provide ordering and change detection.

Question 12Design and prepare a machine learning solution

A team wants to reuse an existing on-premises Kubernetes cluster for training jobs managed by Azure Machine Learning. What must be true?

  1. The cluster must run only Windows nodes
  2. The cluster must be in the same subscription as the workspace by definition
  3. The cluster must be converted to a managed compute cluster
  4. The cluster must have Azure Arc enabled and be attached to the workspace Correct answer.

Explanation

On-premises or other-cloud Kubernetes clusters are attached through Azure Arc so Azure Machine Learning can target them as compute. Once attached, jobs can be submitted to the cluster's namespaces.

Question 13Design and prepare a machine learning solution

Which statement correctly describes the relationship between a workspace and its associated Application Insights resource?

  1. It is optional and provides monitoring telemetry for deployed endpoints Correct answer.
  2. It is mandatory and stores all training data
  3. It is required only when using AutoML
  4. It replaces the need for a Log Analytics workspace

Explanation

Application Insights is an optional workspace dependency used to collect telemetry and diagnostics from online endpoints and web services. Training data is never stored there.

Question 14Design and prepare a machine learning solution

A data scientist wants to train a model on tabular data with minimal code and have Azure Machine Learning automatically perform featurization and algorithm selection. Which approach fits best?

  1. Write a custom sweep job over scikit-learn pipelines
  2. Create a batch endpoint with a scoring script
  3. Deploy a managed online endpoint and send training data to it
  4. Submit an AutoML job with a classification task Correct answer.

Explanation

Automated machine learning handles preprocessing, algorithm selection, and hyperparameter tuning for supported tasks such as classification and regression. It is the lowest-code path for tabular modeling.

Question 15Design and prepare a machine learning solution

Which Azure Machine Learning asset should be used to track which model version is approved for production and to store its metadata?

  1. Data asset in a datastore
  2. Environment version
  3. Compute cluster
  4. Registered model in the model registry Correct answer.

Explanation

The model registry stores model versions with lineage, tags, and descriptions, allowing teams to mark and promote approved versions. Deployment can then reference a specific registered version.

Question 16Design and prepare a machine learning solution

A pipeline needs a step that runs a reusable registered component with typed inputs and outputs. What is the benefit of using a registered component instead of an inline script step?

  1. It automatically tunes hyperparameters
  2. It provides a versioned, reusable interface with defined input and output ports Correct answer.
  3. It guarantees the step runs faster on GPU compute
  4. It removes the need for an environment

Explanation

Components package code, environment, and a typed interface so they can be versioned, shared, and composed across pipelines. Input and output ports make data flow explicit between steps.

Question 17Design and prepare a machine learning solution

You are choosing between file-based and URI-based data inputs for a job. Which consideration most strongly favors a URI-based (mount or direct read) approach?

  1. The data must be embedded in the Docker image
  2. The script requires random writes to the input files
  3. The dataset is very large and copying it to compute would be slow and costly Correct answer.
  4. The compute cluster has no network access to storage

Explanation

URI-based inputs avoid duplicating large datasets onto compute nodes, reducing startup time and storage cost. They are read directly from the source storage during the run.

Question 18Design and prepare a machine learning solution

A team wants reproducible training runs. Which practice most directly supports reproducibility?

  1. Using a compute cluster with autoscaling enabled
  2. Pinning the environment version and data asset version in the job definition Correct answer.
  3. Logging only the final accuracy metric
  4. Running jobs interactively on a compute instance

Explanation

Reproducibility requires fixing the code, environment, and data versions referenced by a job so the same inputs can be replayed. Pinning environment and data asset versions captures those dependencies explicitly.

Question 19Design and prepare a machine learning solution

Which identity type allows a compute cluster to access a storage account without storing credentials in the workspace?

  1. A local user account on each node
  2. A SAS token hard-coded in the environment file
  3. A shared account key embedded in the training script
  4. A workspace managed identity assigned appropriate RBAC roles Correct answer.

Explanation

Managed identities let Azure services authenticate to other resources through Microsoft Entra ID, with authorization granted via RBAC roles. This avoids embedding secrets in code or configuration.

Question 20Design and prepare a machine learning solution

A pipeline must process data that arrives daily and should start automatically when new data lands in a storage account. Which Azure Machine Learning feature supports this?

  1. A compute instance startup script
  2. A model registry webhook
  3. A managed online endpoint with autoscaling
  4. A schedule triggered by a storage event via Event Grid Correct answer.

Explanation

Schedules can trigger pipelines on a time basis or in response to events such as blob creation delivered through Event Grid. This enables event-driven retraining when new data arrives.

Question 21Design and prepare a machine learning solution

An organization requires that all training jobs use approved base images and that dependency changes are auditable. Which design satisfies this with the least operational overhead?

  1. Use the default image and patch dependencies manually per node
  2. Install packages at runtime with pip inside every script
  3. Curate versioned environments in a shared registry and reference them by version in jobs Correct answer.
  4. Let each data scientist build a custom image at job submission time

Explanation

Registered, versioned environments centralize approved dependencies and create an audit trail of what each job used. Referencing a specific version prevents drift and ad-hoc image builds.

Question 22Design and prepare a machine learning solution

A team wants to minimize cost for a long training job that can tolerate interruptions, while ensuring the job eventually completes. Which configuration is most appropriate?

  1. Use dedicated nodes with autoscaling set to a high maximum
  2. Use a compute instance left running overnight
  3. Use low-priority nodes with the job configured to restart on preemption Correct answer.
  4. Use serverless compute with no timeout

Explanation

Low-priority nodes are cheaper but can be preempted, so the job must be resilient-typically through checkpointing or automatic restart. This balances cost savings with completion guarantees.

Question 23Design and prepare a machine learning solution

You must design a solution where feature engineering code runs identically during training and batch inference. Which approach best guarantees this?

  1. Rewrite the transformations separately for inference to optimize speed
  2. Store transformed features only and discard the transformation code
  3. Package the transformation logic in a registered component reused by both the training pipeline and the inference pipeline Correct answer.
  4. Perform feature engineering manually in a notebook before each run

Explanation

Reusing a versioned component for both training and inference ensures the same transformation logic and dependencies are applied, preventing training-serving skew. The component interface makes the contract explicit.

Question 24Design and prepare a machine learning solution

A regulated workload requires that training data never leaves a virtual network and that only specific subnet ranges can reach the workspace. Which combination of design choices is required?

  1. Service endpoints on the default virtual network with no restrictions
  2. Private endpoints for workspace and storage plus network rules restricting access to approved subnets Correct answer.
  3. A workspace in a different region with cross-region replication
  4. Public endpoints with IP allowlists covering all Azure datacenters

Explanation

Keeping data inside the virtual network requires private endpoints for the workspace and dependent services, combined with firewall rules that permit only approved subnets. This prevents public exposure of data and control-plane traffic.

Question 25Design and prepare a machine learning solution

A team needs to compare model performance across many runs and later reproduce the best run exactly. Which logging and tracking practice supports both goals?

  1. Log only the best metric value to a shared spreadsheet
  2. Store outputs only on the local compute instance disk
  3. Overwrite a single run each time a new model is trained
  4. Log metrics and parameters per run in the experiment and record the run ID with the environment and data versions Correct answer.

Explanation

Per-run metrics and parameters enable comparison, while capturing the run ID with environment and data versions allows exact reproduction. This creates a complete lineage record for each candidate model.

Question 26Design and prepare a machine learning solution

A data scientist registers a dataset in Azure Machine Learning. Which statement best describes what a registered data asset provides to the workspace?

  1. A live connection string that grants every workspace user read and write access to the backing storage account
  2. A physical copy of the data that Azure Machine Learning stores inside the workspace's managed container registry
  3. A cached snapshot of the data that is automatically refreshed on a fixed hourly schedule
  4. A named, versioned reference to data stored in a backing datastore, with metadata such as schema and lineage Correct answer.

Explanation

A registered data asset is a versioned pointer to data that stays in its backing storage; the workspace stores the reference plus metadata like schema and lineage. It does not copy the bytes into the workspace and does not itself grant permissions, because access is governed by the datastore credentials and identity.

Question 27Design and prepare a machine learning solution

Which Azure Machine Learning datastore type can be used to reference data in an Azure SQL Database for use with managed identity authentication?

  1. Azure Data Lake Storage Gen2 datastore
  2. Azure SQL Database datastore Correct answer.
  3. Azure Blob Storage datastore
  4. Azure File Share datastore

Explanation

Azure Machine Learning supports datastores for several backing services, and Azure SQL Database is one of the supported datastore types. Each datastore type maps to its own service and credential mechanism, so the SQL type is the one that references a relational database.

Question 28Design and prepare a machine learning solution

A team wants a compute target that scales nodes automatically between defined minimum and maximum bounds and supports idle scale-down. Which Azure Machine Learning compute type fits this requirement?

  1. Azure Machine Learning compute cluster Correct answer.
  2. Compute instance
  3. Attached Azure HDInsight cluster
  4. Managed online endpoint

Explanation

An Azure Machine Learning compute cluster is the managed training compute target that can autoscale between a minimum and maximum node count and shrink when idle. Compute instances are single-user development machines, and endpoints serve inference rather than training scale-out.

Question 29Design and prepare a machine learning solution

When you define a job in Azure Machine Learning, what is the role of the command specified in the job?

  1. It registers the trained model into the workspace registry after completion
  2. It configures the retention period for metrics logged during the run
  3. It declares the entry script and arguments that the job executes on the compute target Correct answer.
  4. It selects the virtual machine size that the compute cluster will provision for the run

Explanation

The command in a command job specifies the program to run, such as a Python entry script, along with its arguments and the environment context. Compute size, metric retention, and model registration are configured through separate properties or steps, not by the command itself.

Question 30Design and prepare a machine learning solution

A pipeline must pass a trained model file from a training step to a separate evaluation step. Which mechanism should you use so the dependency is explicit and tracked?

  1. Write the file to a shared blob path and hard-code the path in both scripts
  2. Store the model in the workspace default datastore root and let the evaluation step discover it
  3. Rely on both steps running on the same compute node and reading a local temporary folder
  4. Declare a pipeline output on the training step and map it as an input to the evaluation step Correct answer.

Explanation

Typed step outputs and inputs create an explicit, tracked data dependency between pipeline steps, so the platform can order execution and record lineage. Hard-coded paths or shared local folders hide the dependency and break when steps run on different nodes.

Question 31Design and prepare a machine learning solution

You need a job to authenticate to a storage account using the identity of the compute cluster rather than a stored key. Which configuration accomplishes this?

  1. Embed the storage account key in the environment Dockerfile so the container can read it
  2. Store the storage key as a plaintext variable in the training script
  3. Enable anonymous public read access on the storage container used by the job
  4. Assign a managed identity to the workspace or compute and grant it the appropriate role on the storage account Correct answer.

Explanation

Managed identities let Azure Machine Learning compute obtain tokens for Microsoft Entra ID without secrets in code, and role assignment on the storage account grants the needed data access. Embedding keys or enabling public access weakens security and is not identity-based authentication.

Question 32Design and prepare a machine learning solution

A team wants the environment for a training job to be rebuilt from the same specification later. Which practice best supports this?

  1. Reference a mutable base image tag such as latest so the newest packages are always used
  2. Allow each engineer to build a personal conda file that is not checked into source control
  3. Pin package versions and base image digests in a registered environment and reference it by name and version Correct answer.
  4. Install dependencies at runtime with an unpinned pip command inside the entry script

Explanation

Reproducibility requires that the environment specification be immutable and versioned, which pinning package versions and image digests in a registered environment provides. Mutable tags and runtime installs can silently change dependencies between runs.

Question 33Design and prepare a machine learning solution

A pipeline must run a step only when a preceding data validation step reports success, and must stop the run otherwise. Which pipeline feature should you configure?

  1. Disable logging on the validation step so failures are not surfaced
  2. Configure the training step to retry indefinitely until validation eventually passes
  3. Set the validation step to run in parallel with the training step to save time
  4. Use step dependencies so the training step depends on validation and the pipeline halts on failure Correct answer.

Explanation

Pipeline steps support explicit dependencies, so a downstream step runs only after its upstream dependency completes successfully; a failed upstream step stops the run by default. Parallel execution or infinite retries would not enforce the gating behavior.

Question 34Design and prepare a machine learning solution

A data scientist needs to log a custom metric from inside a training script so it appears in the run record. Which approach is correct?

  1. Print the metric to stdout and rely on the platform to parse arbitrary console text
  2. Send the metric to the workspace using an HTTP request to the studio web page
  3. Write the metric to a file named metrics.json in the current working directory
  4. Call the run logging API, such as mlflow.log_metric or the Azure ML run log method, from the script Correct answer.

Explanation

Metrics are recorded in the run when the script calls a supported logging API, such as MLflow tracking or the Azure Machine Learning run logging methods. Plain console output or arbitrary files are not automatically promoted to tracked metrics.

Question 35Design and prepare a machine learning solution

An organization requires that training jobs run inside a virtual network and that the workspace's managed compute reach storage without traversing the public internet. Which combination should the design include?

  1. A private endpoint for the workspace and a private endpoint for the storage account, with managed virtual network integration Correct answer.
  2. Public network access enabled on storage with IP firewall rules limited to the corporate office range
  3. Storage account keys distributed to each compute node through environment variables
  4. A service endpoint on the storage account with anonymous access allowed for Azure services

Explanation

Private endpoints place the workspace and storage on private IP addresses within a virtual network, and managed virtual network integration keeps compute traffic inside that network. Public access paths, even with IP restrictions, still traverse public networking and do not meet the isolation requirement.

Question 36Design and prepare a machine learning solution

A team must choose between a managed online endpoint and a batch endpoint for a scoring workload that processes a large file nightly. Which choice and rationale are correct?

  1. Batch endpoint, because it deploys models to edge devices for offline scoring
  2. Batch endpoint, because it is designed for long-running scoring over large datasets with asynchronous job completion Correct answer.
  3. Managed online endpoint, because it keeps a model resident and responds to individual requests with low latency
  4. Managed online endpoint, because it can process arbitrarily large files in a single synchronous request

Explanation

Batch endpoints are built for asynchronous, long-running scoring jobs over large datasets, which matches a nightly bulk file workload. Online endpoints are optimized for low-latency synchronous requests and are not intended for very large single payloads.

Question 37Design and prepare a machine learning solution

A pipeline needs to reuse the same preprocessing logic for training and for scheduled batch inference. Which design most directly ensures the logic stays identical?

  1. Document the preprocessing steps in a wiki so engineers can reimplement them later
  2. Package the preprocessing as a registered component and invoke that component in both pipelines Correct answer.
  3. Rewrite the preprocessing in the batch pipeline using a different library for performance
  4. Copy the preprocessing code into each pipeline's entry script and keep the copies in sync manually

Explanation

A registered component encapsulates the code and environment as a versioned, reusable unit, so referencing it in both pipelines guarantees the same implementation runs. Manual copies or reimplementations drift over time and undermine consistency.

Question 38Design and prepare a machine learning solution

A team wants to minimize cost for a long training job that can tolerate interruptions, while ensuring the job eventually completes. Which configuration should they choose?

  1. A dedicated compute cluster with the maximum node count set to one and no autoscale
  2. A compute instance sized for the largest expected memory footprint
  3. A managed online endpoint with a single instance
  4. A low-priority compute cluster with autoscale and a minimum node count of zero Correct answer.

Explanation

Low-priority clusters use spare capacity at reduced cost and can be preempted, so autoscaling with a zero minimum avoids idle spend and the job is retried until it completes. Dedicated always-on nodes cost more, and instances or endpoints are not appropriate for this training pattern.

Question 39Design and prepare a machine learning solution

A regulated workload requires that only approved base images are used and that every dependency change is auditable. Which environment strategy best satisfies this?

  1. Let each job pull the newest public image at runtime and install packages with unpinned versions
  2. Disable environment versioning and rebuild images on each run from the latest sources
  3. Allow engineers to build ad hoc images in a personal container registry and reference them by tag
  4. Maintain curated, versioned environments built from approved base images and require jobs to reference them by version Correct answer.

Explanation

Curated, versioned environments built from approved base images give an auditable record of exactly what ran and prevent unapproved dependencies. Runtime pulls, personal registries, and unversioned rebuilds all break the audit trail and allow uncontrolled changes.

Question 40Design and prepare a machine learning solution

You are designing a solution where feature engineering must run identically during training and batch inference, and the team also wants to version the feature logic independently of the model. Which design best meets both goals?

  1. Perform feature engineering manually before upload and store only the transformed data
  2. Embed feature logic in each model's scoring script and version it together with the model
  3. Put feature logic in a shared notebook that analysts run interactively before each job
  4. Implement feature logic as a versioned registered component and reference it from both training and inference pipelines Correct answer.

Explanation

A versioned registered component isolates the feature logic as a reusable, independently versioned asset that both training and inference pipelines invoke, ensuring identical behavior. Embedding the logic in scoring scripts couples it to the model, and manual or notebook steps are not reproducible.

Question 41Explore data and train models

In Azure Machine Learning, which asset type is used to store a connection string or a storage account key so that it can be referenced securely from a training script without hard-coding the value?

  1. An environment specification file
  2. A registered model in the model registry
  3. A datastore credential stored in a workspace key vault Correct answer.
  4. A compute cluster configuration

Explanation

Secrets such as storage keys and connection strings belong in the workspace key vault and are referenced through a datastore or a key vault secret. This keeps the value out of source code and out of job logs while still allowing scripts to retrieve it at runtime with identity-based access.

Question 42Explore data and train models

You have a CSV file in Azure Blob Storage and want to register it as a tabular dataset (V2 data asset) in Azure Machine Learning. Which approach accomplishes this?

  1. Create a data asset of type mltable or uri_file through the studio or SDK, pointing at the blob path Correct answer.
  2. Upload the CSV into the workspace key vault
  3. Convert the CSV to a registered model version
  4. Attach the blob container as an AKS inference cluster

Explanation

Data assets in the V2 experience are created by specifying a type such as uri_file, uri_folder, or mltable together with a path to the storage location. Creating the asset records versioned metadata and lineage so jobs can reference it by name and version rather than by raw path.

Question 43Explore data and train models

Which Azure Machine Learning compute type is designed to run a training job that needs a GPU and that should scale down to zero nodes when no jobs are queued?

  1. A compute instance sized for interactive work
  2. An attached Azure Databricks workspace
  3. A compute cluster with autoscaling and a GPU SKU Correct answer.
  4. A Kubernetes inference cluster

Explanation

Compute clusters are the managed compute target for running jobs, and with autoscaling enabled they can shrink to zero nodes when idle. Selecting a GPU virtual machine SKU provides GPU acceleration for the training workload while idle time costs nothing.

Question 44Explore data and train models

Which metric should be used to evaluate a binary classification model when the business cost of a false negative is much higher than the cost of a false positive?

  1. Recall for the positive class Correct answer.
  2. R-squared
  3. Overall accuracy
  4. Mean absolute error

Explanation

Recall measures the proportion of actual positives that the model correctly identifies, so a high recall means few false negatives. Accuracy can look strong even when positives are frequently missed, which is exactly the failure mode that matters when misses are expensive.

Question 45Explore data and train models

What is the primary purpose of splitting a dataset into training and validation subsets?

  1. To guarantee the model reaches perfect training accuracy
  2. To estimate how well the model generalizes to data it has not seen Correct answer.
  3. To increase the number of features available to the model
  4. To reduce the storage footprint of the dataset

Explanation

Holding out data lets you measure performance on examples the model did not learn from, which approximates generalization. Without a holdout, training metrics alone can be misleadingly optimistic because the model can memorize the training rows.

Question 46Explore data and train models

In Azure Machine Learning automated ML, what does the primary metric setting control?

  1. Which algorithm families are excluded from the run
  2. The number of compute nodes used per trial
  3. How many features are dropped during preprocessing
  4. Which metric the trials are ranked by when selecting the best model Correct answer.

Explanation

The primary metric is the score used to compare all completed trials and choose the best model for the task type. Choosing a metric aligned with business goals, such as recall or AUC, steers model selection toward the behavior you care about.

Question 47Explore data and train models

Which Azure Machine Learning feature lets you interactively explore a dataset's summary statistics and column distributions inside the studio?

  1. The endpoint deployment log stream
  2. The compute cluster node list
  3. The Data preview and profile view of a data asset Correct answer.
  4. The model registry details page

Explanation

Data assets in the studio expose a preview of rows plus a generated profile containing per-column statistics such as type, missing counts, and distributions. This supports quick exploratory analysis before committing to feature engineering or model training.

Question 48Explore data and train models

You want a training script to read a registered data asset by name and version instead of a hard-coded storage path. Which mechanism should the script use?

  1. Embed the storage account key directly in the script
  2. Pass the data asset as a job input and read the mounted or downloaded path provided to the script Correct answer.
  3. Read the workspace ARM template at runtime
  4. Query the model registry for the data asset

Explanation

When a data asset is passed as a job input, Azure Machine Learning resolves it and provides the script with a mounted or downloaded local path. This keeps the script portable across versions and environments and preserves lineage between the run and the data used.

Question 49Explore data and train models

A model shows very high accuracy on the training set but much lower accuracy on the validation set. Which conclusion is most appropriate?

  1. The model is likely overfitting and needs more regularization or more training data Correct answer.
  2. The model is underfitting and needs a more complex architecture
  3. The training metric was computed incorrectly
  4. The validation set is corrupt and should be discarded

Explanation

A large gap between training and validation performance is the classic signature of overfitting, where the model memorizes training noise. Reducing model capacity, adding regularization, or increasing data volume typically narrows the gap.

Question 50Explore data and train models

Your dataset has a numeric column with a heavily right-skewed distribution and a few extreme outliers. Which preprocessing step most directly reduces the influence of those outliers before training a linear model?

  1. Winsorizing or clipping the values at chosen percentile bounds Correct answer.
  2. Converting the column to a string type
  3. One-hot encoding the numeric column
  4. Removing the column's mean from every value

Explanation

Clipping extreme values to percentile bounds limits the leverage any single outlier can have on fitted coefficients. Centering alone shifts the distribution but leaves the extreme magnitudes intact, so it does not address the problem.