Microsoft Certified: Azure Data Scientist Associate — Practice Set 1

Question 1

You are setting up an Azure Machine Learning workspace for a data science team. You need to create a reusable compute environment where multiple team members can run interactive Jupyter notebooks with GPU support. Which compute type should you provision?

Accepted Answer

An Azure Machine Learning compute instance is a fully managed, cloud-based workstation for individual data scientists that supports Jupyter notebooks and JupyterLab. It can be configured with GPU-enabled VM sizes. Compute clusters are for scalable batch training jobs. AKS is for model deployment. ACI is used for lightweight test deployments.

Question 2

Your organization needs to store training datasets in Azure Machine Learning so that they can be versioned, tracked, and used across multiple experiments. Which Azure Machine Learning asset should you use?

Accepted Answer

Azure Machine Learning data assets (formerly datasets) provide versioning, lineage tracking, and reusability of training data across experiments. Data assets reference data stored in datastores (Azure Blob Storage, Azure Data Lake Storage, etc.) and can be registered as URI file, URI folder, or MLTable types. Datastores are connection configurations to storage services, not the data itself. Environments define software dependencies. Model registry stores trained models.

Question 3

You need to configure a data scientist's development environment in Azure Machine Learning. Which two items should you configure to ensure reproducibility of experiments across the team? Choose 2.

Accepted Answer

Reproducibility in machine learning requires (A) Azure Machine Learning environments that capture the exact software dependencies (Python packages, Docker base image) and (B) versioned data assets that capture the exact training data snapshot. Azure AD groups manage access; Cost Management handles billing; Blob lifecycle policies manage data retention — none of these contribute to ML experiment reproducibility.

Question 4

Your team wants to share a trained model and its associated data assets across multiple Azure Machine Learning workspaces in different Azure regions. Which Azure Machine Learning feature enables sharing assets across workspaces?

Accepted Answer

Azure Machine Learning registries enable sharing and versioning of ML assets (models, environments, components, data assets) across multiple workspaces and Azure regions. This supports MLOps practices such as promoting models from development to staging to production workspaces. Pipelines are for workflow orchestration within a workspace. Compute clusters are training compute resources. Datastores connect to storage services.

Question 5

You are designing a machine learning solution that will train large deep learning models. The training data is 500 GB stored in Azure Data Lake Storage Gen2. You need to select a compute type that scales automatically during training and shuts down when idle to minimize cost. Which Azure Machine Learning compute type should you use?

Accepted Answer

Azure Machine Learning compute clusters are multi-node scalable clusters that automatically scale up (add nodes) during training and scale down to zero nodes when idle, minimizing costs. They support GPU-enabled VMs for deep learning. Compute instances are single-node workstations, not autoscaling. AKS inference clusters are for model deployment, not training. ACI is for lightweight test deployments.

Question 6

You are a data scientist setting up source control for your Azure Machine Learning project. You need to configure version control for your notebooks and training scripts so that changes are tracked and the team can collaborate using pull requests. Which Azure Machine Learning feature should you configure?

Accepted Answer

Azure Machine Learning compute instances support Git integration, allowing data scientists to clone repositories from GitHub, Azure DevOps, or other Git providers directly onto the compute instance. This enables version control for notebooks and training scripts with full support for branching, commits, and pull requests. Model registry versioning tracks trained model artifacts, not code. Pipeline triggers automate job execution. Datastore access keys configure storage connectivity.

Question 7

You want to quickly find the best machine learning algorithm and preprocessing steps for a classification problem without manually testing each combination. Which Azure Machine Learning capability automates this process?

Accepted Answer

Automated machine learning (AutoML) in Azure Machine Learning automatically explores multiple algorithms, preprocessing steps, and hyperparameter combinations to find the best performing model for a given dataset and task type (classification, regression, time-series forecasting). It evaluates models and selects the best one based on a primary metric. Azure ML pipelines orchestrate workflow steps; the designer is a drag-and-drop tool; MLflow tracks experiment metrics.

Question 8

You are training a machine learning model and want to track the model's accuracy, loss values, and hyperparameters for each experiment run so you can compare runs later. Which tool integrated with Azure Machine Learning should you use?

Accepted Answer

MLflow is an open-source platform for ML lifecycle management that is deeply integrated with Azure Machine Learning. It provides experiment tracking (logging metrics, parameters, artifacts), model registry, and model serving. You use mlflow.log_metric(), mlflow.log_param(), and mlflow.log_artifact() in your training scripts. Azure Monitor and Application Insights are for application/infrastructure monitoring, not ML experiment tracking.

Question 9

You are conducting hyperparameter tuning for a neural network. You want to randomly sample hyperparameter combinations from a defined search space to find the best configuration efficiently. Which sampling method should you use?

Accepted Answer

Random sampling randomly selects hyperparameter combinations from the defined search space. It is computationally efficient and can find good configurations without exhaustively testing all combinations. Grid sampling evaluates every combination in the search space (exhaustive, expensive). Bayesian sampling uses prior results to intelligently choose the next set of hyperparameters (more efficient but requires more sequential runs). There is no 'sequential sampling' in Azure ML.

Question 10

You are using AutoML in Azure Machine Learning to train a classification model. After the run completes, you review the results and find a warning about class imbalance in the training data. According to responsible AI guidelines in Azure Machine Learning, what does this warning indicate?

Accepted Answer

Class imbalance occurs when one or more target classes have significantly fewer training samples than others (e.g., 95% 'non-fraud' vs. 5% 'fraud' in a fraud detection dataset). This can cause the model to be biased toward predicting the majority class, leading to poor recall on minority classes. Azure Machine Learning AutoML detects and flags this as a responsible AI concern. Mitigation strategies include oversampling, undersampling, or using weighted loss functions.

Question 11

You are tuning a machine learning model's hyperparameters using Azure Machine Learning's sweep job. Which two early termination policies help stop poorly-performing runs early to save compute costs? Choose 2.

Accepted Answer

The Bandit policy terminates runs where the primary metric falls below a specified slack factor from the best performing run. The Median stopping policy terminates runs whose primary metric falls below the median of all running trials. Both are early termination policies in Azure Machine Learning sweep jobs. Truncation selection policy is also valid (terminates the bottom X% of runs), but the question asks for two from those listed. Grid search and cross-validation are not early termination policies.

Question 12

You are a data scientist working in a Jupyter notebook on an Azure Machine Learning compute instance. You need to access a large feature store to retrieve precomputed features for model training. Which Azure Machine Learning capability should you use to retrieve features from the feature store?

Accepted Answer

Azure Machine Learning feature store allows teams to create, share, and manage ML features. In a notebook, you retrieve features by specifying a feature retrieval specification that defines which feature sets and features to retrieve for a given set of entity IDs. This ensures consistent feature computation between training and inference. Azure ML pipeline components orchestrate multi-step workflows. Azure Data Factory handles data movement. Azure Synapse Spark pools are a different compute environment.

Question 13

You have trained a machine learning model and want to register it in Azure Machine Learning with full MLflow format support, including the model signature. Which command registers an MLflow model in Azure Machine Learning from a training script?

Accepted Answer

Using mlflow.<flavor>.log_model() (e.g., mlflow.sklearn.log_model()) with the registered_model_name parameter both logs the model as an artifact and registers it in the MLflow model registry (which maps to Azure Machine Learning's model registry). The model signature (input/output schema) should be defined using mlflow.models.infer_signature(). This is the recommended pattern for registering MLflow models during training.

Question 14

You need to deploy a trained machine learning model to an endpoint that can handle incoming REST API requests for real-time predictions. The endpoint must scale automatically based on traffic. Which Azure Machine Learning deployment type should you use?

Accepted Answer

Managed online endpoints in Azure Machine Learning provide a fully managed infrastructure for hosting models that respond to real-time REST API requests. They support auto-scaling based on request volume, multiple deployment slots (blue/green deployments), and traffic splitting. Batch endpoints are for asynchronous batch scoring. ACI is used for non-production test deployments without auto-scaling. Compute clusters run training jobs, not serve predictions.

Question 15

You have a multi-step machine learning pipeline in Azure Machine Learning. Step 1 preprocesses raw data and outputs a cleaned dataset. Step 2 trains a model using the cleaned dataset. Which Azure Machine Learning feature allows you to pass data between these pipeline steps?

Accepted Answer

In Azure Machine Learning pipelines, data is passed between steps using the Output type on the producing step and the Input type on the consuming step. The pipeline framework manages the data transfer through Azure Machine Learning-managed storage. This ensures clean data lineage tracking between pipeline steps. Environment variables, connection strings, and SAS tokens are infrastructure configurations, not the correct mechanism for inter-step data passing in Azure ML pipelines.

Question 16

Your organization has 10 million records to score using a trained model. The scoring does not need to happen in real time — results can be delivered within a few hours. Which Azure Machine Learning deployment type is most cost-efficient for this scenario?

Accepted Answer

Batch endpoints in Azure Machine Learning are designed for asynchronous, large-scale batch scoring jobs. They invoke a batch scoring job that runs on a compute cluster, processes all input data, and writes results to a specified output location. This is highly cost-efficient for large-scale scoring that does not require real-time latency. Online endpoints are for real-time low-latency scoring and would be expensive and unnecessary for batch workloads.

Question 17

You are building a responsible AI evaluation for a classification model in Azure Machine Learning. Which two metrics should you review to assess potential fairness issues across demographic groups? Choose 2.

Accepted Answer

Demographic parity difference measures whether the model produces similar prediction rates across demographic groups. Equalized odds difference measures whether the model has similar true positive and false positive rates across groups. Both are fairness metrics available in Azure Machine Learning's Responsible AI dashboard. Overall accuracy (B) and MAE (D) are performance metrics but do not assess fairness across subgroups. Training duration (E) is an operational metric.

Question 18

You are a data scientist troubleshooting a failing Azure Machine Learning pipeline run. The second step of the pipeline fails with an out-of-memory error. You need to investigate the error logs. Where should you look first in Azure Machine Learning Studio?

Accepted Answer

When an Azure Machine Learning pipeline step fails, the error details and stdout/stderr logs are available in Azure Machine Learning Studio under Jobs > [Pipeline Run] > [Step] > Logs tab. This shows the exact error message, stack trace, and output from the script. Azure Monitor Metrics shows aggregate cluster metrics but not script-level logs. Storage accounts hold output artifacts. Key Vault stores secrets.

Question 19

Your organization wants to deploy a large language model (LLM) for a customer-facing chatbot. You need to select a pre-trained foundation model and test it in a playground before deployment. Which Azure service provides a model catalog with foundation models and a playground for testing?

Accepted Answer

Azure AI Foundry (formerly Azure AI Studio, now Microsoft Foundry) provides a model catalog with hundreds of foundation models from Microsoft, OpenAI, Meta, Mistral, and other providers. It includes a playground where you can test deployed models with prompts before integrating them into applications. Azure ML designer is for classical ML pipelines. Azure Synapse and Databricks AutoML are for data analytics and classical ML.

Question 20

You need to improve a language model's responses for your organization's internal HR policy questions. The model frequently answers incorrectly because it lacks knowledge of your specific HR policies. Which optimization technique adds your policy documents to a vector index so the model can retrieve relevant context at inference time?

Accepted Answer

Retrieval Augmented Generation (RAG) is a technique where relevant documents are retrieved from a vector index at inference time and included in the model's context window. This allows the model to answer questions about domain-specific knowledge (like HR policies) without retraining. Azure AI Search is commonly used as the vector index. Fine-tuning modifies model weights using training data. Prompt engineering adjusts the prompt structure. Model distillation creates smaller models.

Question 21

You are using prompt flow in Azure AI Foundry to build a multi-step LLM application. You want to test multiple variants of a system prompt to see which produces better responses. Which prompt flow feature supports this comparison?

Accepted Answer

Prompt flow in Azure AI Foundry supports prompt variants, which allow you to define multiple versions of a prompt node and run batch evaluations to compare which variant produces the best results on a test dataset. You can define variants using the variant property and run them in parallel batch runs. Deployment endpoint scaling is for production serving. Connection configuration sets up model API connections. Compute session restart is an operational action.

Question 22

You are preparing data for Retrieval Augmented Generation (RAG) using Azure AI Search. Which two steps are essential in preparing documents for a RAG vector index? Choose 2.

Accepted Answer

RAG document preparation requires (A) chunking: splitting documents into smaller passages that fit within the model's context window, and (B) generating vector embeddings: converting each chunk into a high-dimensional numeric vector representation using an embedding model. These vectors are stored in Azure AI Search as a vector index for semantic similarity retrieval. ZIP compression, PowerPoint conversion, and language translation are not required steps.

Question 23

You have fine-tuned a language model to answer questions in your organization's domain-specific terminology. Before promoting the model to production, you need to measure whether the fine-tuned model produces more accurate and contextually relevant answers than the base model. Which approach should you use?

Accepted Answer

Evaluating a fine-tuned model requires running systematic batch evaluations using prompt flow with a ground truth test dataset. Evaluation metrics such as relevance, coherence, fluency, and groundedness are computed to objectively compare model versions. Manually reviewing five samples is not statistically reliable. Model file size is not an accuracy indicator. Training loss measures training convergence but not real-world output quality.

Question 24

You need to fine-tune a language model for a specific task. Before running the fine-tuning job, you must prepare the training dataset. Which format does Azure AI Foundry require for supervised fine-tuning data?

Accepted Answer

Azure AI Foundry fine-tuning requires training data in JSONL (JSON Lines) format. For chat models, each line is a JSON object with a 'messages' array containing system, user, and assistant turns. For completion models, each line has 'prompt' and 'completion' fields. This is the standard format for OpenAI-compatible fine-tuning APIs. CSV, Parquet, and XML are not supported formats for fine-tuning.

Question 25

You are optimizing an LLM application using prompt flow. You suspect that a complex reasoning task is failing because the model is not producing intermediate reasoning steps. Which prompt engineering technique should you apply to the system prompt to encourage step-by-step reasoning?

Accepted Answer

Chain-of-thought (CoT) prompting instructs the model to produce intermediate reasoning steps before arriving at the final answer. By including phrases like 'Let's think step by step' or providing few-shot examples with explicit reasoning chains in the system prompt, you can significantly improve the model's performance on complex reasoning tasks. Temperature 0.0 reduces randomness but does not add reasoning structure. Increasing max_tokens just allows longer outputs.

DP-100 — Azure Data Scientist Associate