Machine Learning Operations Engineer Associate (AI-300) - Full Practice Exam 1

1. A candidate model passes release criteria. Which two artifacts are important for controlled deployment? Select TWO.
2. A deep-learning job must train across multiple nodes. What should be configured together?
3. A demand-forecast model is evaluated on time-series data. Which split strategy helps avoid learning from future information?
4. A deployment still needs a protected external API secret for an agent tool. Where should the secret be managed?
5. A fine-tuned model improves one task but regresses on important safety and generalization cases. What should the release process do?
6. A foundation-model version changes while the service remains reachable. Which release assumption is unsafe?
7. A generated answer is fluent and relevant to the topic but contains claims unsupported by its retrieved evidence. Which quality dimension most directly reveals the problem?
8. A model-based evaluator rates responses highly, but sampled human review finds systematic errors. What is the appropriate response?
9. A model must respond synchronously to individual application requests with low latency. Which endpoint pattern fits?
10. A model scores millions of records overnight and consumers can wait for output files. Which endpoint pattern should be evaluated?
11. A new model should receive a small share of live traffic before broad rollout. What should the deployment support?
12. A platform hosts experimental and production agents. Researchers must not alter production deployments accidentally. Which resource and identity design fits?
13. A platform wants reproducible ML workspace infrastructure deployed by a trusted GitHub workflow without storing long-lived Azure secrets. Which combination should it evaluate?
14. A private AI deployment cannot reach its knowledge store even though the model endpoint works. What should be investigated?
15. A proposed foundation model must run in a required Azure region with a particular deployment mode. What should be verified before implementation?
16. A support corpus contains exact error codes as well as conceptual descriptions. Which retrieval strategy should be evaluated?
17. A team compares prompt variants. Which experimental practice is appropriate?
18. A team compares supported serverless API and managed compute deployment options. Which considerations should guide the decision?
19. A team compares two retrieval configurations in production. What makes the comparison more trustworthy?
20. A team must compare training attempts by hyperparameters, metrics, and produced artifacts. Which capability should it use?
21. A team needs a management boundary for Azure ML compute, experiments, data references, and model assets. Which resource is the foundation?
22. A team uses automated ML to compare candidate models. What must it still define and review?
23. A training job uses a managed identity. Which two access concerns must be validated? Select TWO.
24. A training script should reference a versioned dataset without embedding storage credentials into the source. Which design should be evaluated?
25. A trusted CI workflow deploys Foundry infrastructure. The organization wants to avoid long-lived stored application secrets. Which approach fits?
26. A workflow prepares data, trains a model, evaluates it, and conditionally registers the candidate. Which orchestration approach fits?
27. A workload can use several models with different cost and latency. Which selection process is appropriate?
28. An agent evaluation includes normal tasks but no attempts to misuse its tools. Which coverage should be added?
29. An application has sustained predictable high inference demand and requires stable capacity. Which supported deployment option should be evaluated against measured workload demand?
30. An assistant must use rapidly changing private policies. The team proposes fine-tuning as the only update mechanism. Which alternative should be evaluated first for factual freshness?
31. An evaluation run maps the wrong dataset column into the user prompt field. What should happen before interpreting its scores?
32. An ML workspace is restricted using private connectivity. Training jobs still cannot retrieve their data and images. What must the design investigate?
33. Detailed traces contain private source passages and API secrets. What should observability design do?
34. End-to-end response latency increases. The team must distinguish retrieval time, model inference, and tool execution. Which instrumentation is most useful?
35. Feature distributions are stable, but the relationship between inputs and outcomes changes and accuracy falls. Which issue should be investigated?
36. Hyperparameter experiments are expensive and the team has a fixed compute budget. Which design should be evaluated?
37. Inference requests remain constant, but total cost rises. Prompts now include more retrieved content and answers are longer. Which telemetry should be examined?
38. Monitoring triggers a retraining workflow. What must happen before replacing the production model?
39. Multiple training pipelines need the same validated preprocessing step with defined inputs and outputs. Which Azure ML asset should be considered?
40. Relevant passages are buried inside very large indexed documents. Which retrieval optimization should be evaluated?
41. Reviewers need to inspect proposed changes to AI infrastructure before the release applies them. Which pipeline step is useful?
42. Several workspaces need controlled sharing of approved model and component versions. Which capability should the platform evaluate?
43. Synthetic examples will be used for model customization. What should the data preparation process include?
44. The distribution of incoming prediction features differs substantially from training data. Which condition is being observed?
45. The team changes to an incompatible embedding model. What must the index migration plan address?
46. Training features are computed one way, but the production API computes them differently. What risk should be investigated?
47. Training runs occasionally require many nodes but the cluster should not retain peak idle capacity. Which compute pattern should be evaluated?
48. Two identical training scripts produce different behavior because their libraries differ. Which asset should be versioned and pinned?
49. Two releases use the same model but different system instructions. The team wants reproducible evaluation and rollback. What should it version?
50. Two runs have different validation results. Which evidence is needed to interpret the difference?

Comments — 0