About these questions
These five questions are original. We wrote them in the style of the AI-300 skills outline, one for each domain. They are not real exam questions, and the real exam will word things differently and may use other formats, such as case studies or interactive parts. Our question-writing page explains how we write ours.
How to use these
Try each question without looking at your notes, and pick an answer before you open the explanation. Then read the explanation even when you were right, because the reason the closest wrong option fails is where the learning is. Note any domain where you slipped and revisit it on the domains explained page.
Question 1 (Design and implement an MLOps infrastructure)
A research team keeps its image files in an Azure Files share. A new company rule says datastores in workspace ws-research must never hold stored secrets and must use identity-based access only. You need to make the image files usable in ws-research. What should you do?
- A. Create an azure_file datastore that stores an account key
- B. Create an azure_file datastore that uses the compute cluster’s managed identity and has no credentials
- C. Copy the files to Blob Storage or Azure Data Lake Storage Gen2 and register an identity-based datastore
- D. Create an azure_file datastore that stores a SAS token with a short lifetime
Show the answer and explanation
Correct answer: C. Identity-based access is supported for Blob Storage and Data Lake Storage Gen2 datastores, so no secret has to be saved in the workspace. An azure_file datastore accepts only an account key or a SAS token, and both are stored with the datastore. The closest wrong option is D, because a short lifetime makes the token safer but it is still a stored secret. Option B fails because Azure Files datastores do not support identity-based access, so the datastore cannot be created without credentials.
Question 2 (Implement machine learning model lifecycle and operations)
Deployment v1 on a managed online endpoint named fraud-ep serves all live traffic. You have created deployment v2 with a retrained model. You want v2 to process a copy of 20% of production requests so you can compare its latency and errors, while every client keeps receiving the response from v1. What should you configure?
- A. Mirrored traffic of 20% to v2
- B. A live traffic split of v1=80 and v2=20
- C. A second endpoint for v2 that receives 20% of calls through a load balancer
- D. Data collection on v1, with the saved requests sent to v2 once a week
Show the answer and explanation
Correct answer: A. Mirrored traffic (also called shadow traffic) copies a share of live requests to the other deployment and discards its responses, so clients only ever see v1. That lets you watch v2 under real load. The closest wrong option is B: a live split makes v2 answer 20% of the clients, which exposes them to an untested model. Option C has the same problem, and option D is delayed replay, which does not show how v2 behaves under live conditions.
Question 3 (Design and implement a GenAIOps infrastructure)
A support chatbot runs on a provisioned deployment named chat-ptu in a Foundry resource. During rare traffic peaks, requests that exceed the capacity of chat-ptu return HTTP 429 errors. You must send only the overflow requests to a Standard deployment of the same model and version, without changing the app. What should you configure?
- A. A Global Batch deployment of the same model
- B. Spillover from chat-ptu to the Standard deployment
- C. An Azure reservation for more provisioned throughput units
- D. A model router deployment
Show the answer and explanation
Correct answer: B. Spillover lets a provisioned deployment pass requests it cannot serve to a Standard deployment of the same model and version, so peaks do not turn into errors. The closest wrong option is C. A reservation lowers the price of provisioned capacity but does not add capacity or route any overflow. Global Batch is for asynchronous jobs, not live chat, and a model router chooses between different models rather than handling extra load.
Question 4 (Implement generative AI quality assurance and observability)
Your test file qa-set.jsonl stores each user question in a field named prompt_text. The relevance evaluator you use expects its input to be called query. You call evaluate() from the Azure AI evaluation SDK. How should you connect the field to the evaluator?
- A. Pass output_path=”prompt_text” to evaluate()
- B. Rename the evaluator’s key in the evaluators dictionary to prompt_text
- C. Set a column mapping of prompt_text to ${target.query} in the evaluator configuration
- D. Set a column mapping of query to ${data.prompt_text} in the evaluator configuration
Show the answer and explanation
Correct answer: D. The mapping key is the evaluator’s input name and the value points to its source. The ${data.field} form reads a column from the dataset, so query maps to ${data.prompt_text}. Option C is the closest wrong answer: it reverses the direction and uses the target prefix, which refers to the output of a target function, not to the dataset. The key in the evaluators dictionary only labels results, and output_path sets where results are written.
Question 5 (Optimize generative AI systems and model performance)
In a fine-tuning job, the training loss keeps falling every epoch, but the validation loss starts to rise after epoch three. You need to reduce this behavior in the next run. Which two changes should you make? Choose two.
- A. Increase the learning rate multiplier
- B. Reduce the number of epochs
- C. Increase the number of epochs
- D. Lower the learning rate multiplier
- E. Remove the validation file from the job
Show the answer and explanation
Correct answers: B and D. Falling training loss with rising validation loss is overfitting: the model memorizes the training set instead of generalizing. Fewer epochs stop training sooner, and a lower learning rate multiplier makes each update gentler. Option C is the closest wrong answer, since more epochs deepen the memorization. Option A makes updates more aggressive, and option E hides the problem by removing the very metrics that reveal it.
More guides for AI-300
- AI-300 Exam Format, Scoring and Retakes
- AI-300 Study Plan: 4 and 8 Weeks
- AI-300 Prerequisites and Skills Check
- AI-300 Renewal and Retakes Explained
- Is AI-300 Worth It? An Honest Look
- AI-300 Cheat Sheet: Key Facts to Review
- AI-200 vs AI-300: Which Exam First?
Next step
Want to see how you do on more of these? Our AI-300 practice tests give you 250 original questions in five timed sets. For a view of where people struggle, read is AI-300 hard?.