AI-300 Hard Questions — 25 Difficult Questions

Question 1

You are configuring a 'Managed Online Endpoint' for a very large model (10GB+). The deployment is failing during the 'pulling image' phase. What is the most likely cause?

Accepted Answer

The 'readiness_probe' timeout is too short.. The 'pulling image' phase is when the managed online endpoint pulls the model's Docker image from Azure Container Registry to the compute instance. For a 10GB+ model, the pull can take many minutes. The default readiness_probe timeout is often too short, causing the deployment to fail before the image finishes pulling and the container becomes ready. Increasing the readiness_probe timeout (and possibly the startup timeout) resolves this.

Answer

The instance count is too high.

Answer

The ACR is private.

Answer

The compute is too small.

Question 2

You need to debug a training job that is failing inside a specific Docker container. What is the best way to investigate?

Accepted Answer

Use 'az ml job stream' to view logs. The 'az ml job stream' command streams the logs of a running or completed Azure ML job in real time, including container logs from the Docker environment. This allows you to see error messages, stack traces, and output from the training script inside the container, making it the most direct way to debug a failing job.

Answer

Delete the datastore

Answer

Use the Azure ML SDK to download the container image

Answer

Submit a new training job without parameters

Question 3

Which THREE of the following are key components of a robust MLOps strategy in Azure?

Accepted Answer

Automated CI/CD pipelines. Automated CI/CD pipelines (B) are a core MLOps component because Azure DevOps or GitHub Actions can automatically build, test, and deploy ML assets via Azure Machine Learning pipelines, ensuring repeatable and reliable releases. Version control for code and models (C) is essential since Azure ML and Git track code, datasets, and model artifacts, enabling reproducibility, auditability, and rollback of experiments and deployments. Automated model monitoring (E) is also key because Azure Machine Learning model monitoring detects data drift, prediction drift, and performance degradation in production, triggering alerts or retraining. Manual model testing (A) is not a robust MLOps practice because it is error-prone, slow, and non-repeatable compared with automated testing in CI/CD. Using only local workstations (D) is incorrect because MLOps relies on shared, scalable cloud resources such as Azure Machine Learning compute, registries, and pipelines rather than isolated local environments.

Answer

Manual model testing

Answer

Using only local workstations

Question 4

You are deploying a model that requires a high-memory compute for inference. You are using a 'Managed Online Endpoint'. Where do you specify the instance type for this deployment?

Accepted Answer

In the 'Deployment' YAML file under 'instance_type'.. The instance type is defined in the 'Deployment' configuration object, typically in the 'instance_type' field.

Answer

In the 'Endpoint' YAML file.

Answer

In the 'Environment' definition.

Answer

In the 'Workspace' settings.

Question 5

A pipeline step fails because it cannot find a file in the datastore. What is the most likely cause?

Accepted Answer

Incorrect path definition in the component. The mount path or the relative path defined in the pipeline step component configuration is likely incorrect.

Answer

The workspace is too large

Answer

The compute cluster is off

Answer

The subscription is expired

Question 6

You need to implement a retraining trigger based on performance degradation. Which TWO metrics should you monitor to decide when to retrain?

Accepted Answer

Model prediction precision. Options D and E are correct because retraining triggers based on performance degradation must directly measure how well the model's predictions match ground truth: model prediction precision (D) captures the proportion of positive predictions that are actually correct, and model prediction accuracy (E) captures the overall fraction of correct predictions, both of which decline when the model drifts or degrades. Monitoring these model-quality metrics lets you set thresholds (e.g., accuracy drops below 90% or precision falls by 5%) that signal when retraining is warranted. Option A (CPU usage of the inference cluster) is an infrastructure resource metric that reflects load or scaling needs, not model correctness. Option B (network latency) measures communication delay and says nothing about prediction quality. Option C (number of active users) is a traffic/usage metric that may correlate with load but does not indicate whether the model's predictions have degraded.

Answer

CPU usage of the inference cluster

Answer

Network latency

Answer

Number of active users

Question 7

You notice your model is outputting redundant information. Which parameter specifically targets the penalty for repeating tokens?

Accepted Answer

frequency_penalty. The frequency_penalty parameter applies a penalty proportional to how many times a token has already appeared in the generated text, directly discouraging repeated tokens and reducing redundancy. Its value ranges from -2.0 to 2.0, with positive values increasing the penalty for frequent tokens.

Answer

max_tokens

Answer

presence_penalty

Answer

top_p

Question 8

You are configuring a connection to a vector store. What is the best way to handle the secret credential?

Accepted Answer

Store in Azure Key Vault and reference in the connection.. Key Vault is the centralized store for all application secrets in Azure, which should be linked to the project connection.

Answer

Hardcode the key in the flow configuration.

Answer

Pass it as a parameter in the code.

Answer

Store it in an environment variable.

Question 9

You are building a RAG application and notice that the model sometimes hallucinates information not present in the retrieved documents. Which evaluation metric should you prioritize to mitigate this?

Accepted Answer

Groundedness. Groundedness specifically assesses whether the generated response is derived from the retrieved documents.

Answer

Relevance

Answer

Performance

Answer

Fluency

Question 10

You are experiencing latency issues with a model deployed on a Managed Online Endpoint. What is the first thing you should check?

Accepted Answer

Application Insights logs. Reviewing logs in Application Insights identifies performance bottlenecks and latency issues.

Answer

The Azure ML workspace name

Answer

The subscription limit

Answer

The storage account size

Question 11

You are debugging a prompt that is performing poorly on edge cases. You decide to use a 'Chain-of-Thought' approach. Why does this improve performance?

Accepted Answer

It enables the model to break down complex logic into sequential steps.. Chain-of-Thought (CoT) prompting instructs the model to produce intermediate reasoning steps before the final answer, which improves accuracy on multi-step logic, arithmetic, and edge cases by giving the model more compute and context to work through ambiguity. This sequential decomposition is why CoT helps on problems where a direct answer would skip critical reasoning.

Answer

It bypasses the safety alignment layer.

Answer

It forces the model to use more input tokens, reducing costs.

Answer

It reduces the context window size.

Question 12

Your team wants to perform 'Red Teaming' on your application. Which activity describes this process correctly?

Accepted Answer

Controlled adversarial testing to find safety and security gaps. Red Teaming involves adversarial testing to find edge cases, safety flaws, and security vulnerabilities.

Answer

Unit testing code snippets

Answer

Automated performance testing for latency

Answer

A/B testing two different model versions

Question 13

You need to detect 'jailbreak' attempts in your RAG application. You are implementing a custom evaluation pipeline. Which technique is most effective for identifying adversarial inputs designed to bypass system instructions?

Accepted Answer

Adversarial input classification using a dedicated judge model. Adversarial evaluation using a 'jailbreak' detection model or prompt-based evaluation is the standard approach to identify bypass attempts.

Answer

Token usage tracking

Answer

Latency threshold monitoring

Answer

Semantic similarity scoring

Question 14

You are deploying a custom model in a container. To ensure the model infrastructure is highly available, what is the best practice?

Accepted Answer

Deploy to 'Azure Kubernetes Service' with 'Availability Zones' enabled.. Deploying across multiple Availability Zones ensures that the infrastructure remains operational even if a specific datacenter fails.

Answer

Disable the 'Load Balancer' to save costs.

Answer

Use a 'Standard' tier App Service plan.

Answer

Deploy the container on a single large VM.

Question 15

You are optimizing the cost of your GenAI infrastructure. You have several agents running in Prompt Flow that are idle for large portions of the day. Which runtime configuration should be modified?

Accepted Answer

Set the 'automatic_shutdown' property on the Prompt Flow runtime compute instance.. Setting the 'idle_time_before_shutdown' or similar inactivity policies in the Prompt Flow runtime ensures compute resources are deallocated when not in use.

Answer

Change the VM SKU to a burstable instance type.

Answer

Implement a script to delete the runtime when the flow is finished.

Answer

Set the concurrency limit to zero.

Question 16

You are configuring a 'Managed Online Endpoint' with SSL termination. Where do you manage the SSL certificates?

Accepted Answer

Via the Azure front-end load balancer/gateway service.. The SSL certificates are managed at the Azure 'Front Door' or 'Application Gateway' level, or by using the built-in certificate management if using Azure-provided domains for endpoints.

Answer

In the 'Python' score script.

Answer

Inside the Endpoint configuration YAML.

Answer

In the 'workspace.json' file.

Question 17

A training job is consuming too much disk space on a compute cluster. What can you do to provide more persistent scratch space?

Accepted Answer

Mount a Datastore as a mount point. Mounting a Datastore or using Azure Files provides persistent, shared scratch space accessible by the compute cluster.

Answer

Use a smaller Docker container

Answer

Increase the OS disk size of the VM

Answer

Change the experiment name

Question 18

Which THREE settings are part of the 'InferenceConfig' object in Azure ML?

Accepted Answer

Source directory.. In Azure ML, the InferenceConfig object (created via InferenceConfig or inference_config in the Python SDK) defines how a deployed model runs, and its core parameters include the source directory (A), which points to the folder containing the model artifacts and scoring code; the environment definition (B), which specifies the Docker image and dependencies used by the inference container; and the entry script path (D), which identifies the scoring script (e.g., score.py) that contains the init() and run() functions. These three settings are exactly the fields accepted by InferenceConfig (entry_script, source_directory, environment). Option C (target storage account URL) is not part of InferenceConfig—storage targets are configured in deployment or datastore settings, not in the inference configuration. Option E (compute instance size) is also not part of InferenceConfig; compute sizing belongs to the deployment configuration (e.g., AciWebservice or AksWebservice) or the compute target definition, not to the inference configuration object.

Answer

Target storage account URL.

Answer

Compute instance size.

Question 19

You are managing a multi-workspace environment. You need to restrict the ability to create new compute clusters to only specific users. Where should you apply this control?

Accepted Answer

Azure RBAC on the Workspace. RBAC (Role-Based Access Control) at the workspace level determines which users can interact with compute resources.

Answer

Network Security Group

Answer

Azure Storage Access Policies

Answer

Azure ML Datastore configuration

Question 20

You have an automated deployment pipeline. You want to run an integration test on the model after deployment. Which tool is best suited for this?

Accepted Answer

Azure DevOps pipelines.. 'Azure DevOps' (or GitHub Actions) is the industry-standard tool for orchestrating post-deployment testing as part of a CI/CD pipeline.

Answer

A manual 'Endpoint' health check.

Answer

The training script.

Answer

Azure Machine Learning Studio 'Tests' tab.