Microsoft AI-300 - Questions & Answers

A shared Azure Machine Learning workspace is the first standardization boundary because it centralizes compute, datastores, jobs, models, and other project assets. A registry is useful after assets exist and need cross-workspace sharing, but it is not the first place to manage project execution. Other choices focus on deployment or Foundry rather than the workspace-level MLOps requirement.

Managed compute targets with autoscaling isolate training workloads while scaling down when demand drops, which controls cost. A fixed-size cluster or dedicated clusters can leave unused capacity running, and a single shared cluster weakens workload isolation. The exam rule is to match compute configuration to both isolation and cost requirements.

Azure Machine Learning compute clusters support training jobs and can use low-priority virtual machines for discounted compute. Azure Container Instances and local deployment are not the standard managed training target for this requirement, and AKS is generally used for Kubernetes-based serving scenarios.
The other choices do not combine managed training execution with low-priority VM economics.

Splitting traffic between deployments is the safe rollout strategy for testing a new model version while existing users continue to use the current deployment. Replacing the endpoint or routing all traffic to the new deployment increases risk, and a batch endpoint does not serve the same real-time production pattern. The requirement is controlled validation before full rollout.

Fine-tuning chat models requires training data in chat-style JSONL records that represent the conversational messages and expected assistant output. Plain CSV, unstructured text, or unrelated data formats do not preserve the message roles required for supervised tuning. The other alternatives miss the structure needed to train a model on prompt and response behavior.

Multi-turn chat fine-tuning data must preserve the sequence of system, user, and assistant messages so the model can learn context across turns. A flat record without roles or conversation order does not satisfy the training requirement. The other formats may store text, but they do not represent the conversational structure needed for a chat model.

Classification accuracy for labeled support-ticket examples is improved by fine-tuning with task-specific labeled examples that teach the model the target categories. Retrieval or monitoring alone does not change the model behavior for the classification task. Other alternatives may help evaluation or operations, but the requirement asks for model adaptation through fine-tuning.

Microsoft Foundry supports comparing prompt variants in a development workflow so teams can evaluate behavior before release. This directly matches prompt experimentation and performance comparison.
Other choices such as production-only edits, storage containers, or endpoint configuration do not provide the focused prompt comparison capability required by the scenario.

An Azure Machine Learning registry is designed for sharing validated assets such as models, components, and environments across workspaces. That avoids duplicating assets between development and production. Other alternatives can store files or run deployments, but they do not provide the governed cross-workspace asset reuse required by the scenario.

Private networking and disabling or restricting public network access keep an Azure Machine Learning workspace off the public internet. Compute changes, model registration, or application logging do not create the required network boundary. The operational rule is to enforce connectivity at the workspace networking layer when sensitive training data must remain private.