Google Cloud Professional Machine Learning Engineer - Full Practice Exam 1

1. A business needs text extraction from common supported documents with minimal model operations. Which approach best fits?
2. A candidate model should be measured on production-like inputs without affecting the user's served response. Which approach best fits?
3. A classifier has high training accuracy but poor representative validation performance. Which approach best fits?
4. A costly job can restart from periodic saved state after interruption. Which approach best fits?
5. A deployed model performs much worse for an important user subgroup than its aggregate score suggests. Which approach best fits?
6. A deployed model version needs traceability to its image, training data, parameters, and evaluation result. Which approach best fits?
7. A feature is computed differently in the online service than in the training pipeline. Which approach best fits?
8. A fraud dataset is highly imbalanced and a model predicts every transaction as legitimate while reporting high accuracy. Which approach best fits?
9. A generative application needs current enterprise facts without retraining the foundation model on every document update. Which approach best fits?
10. A generative endpoint's output is sometimes truncated before a required structured field is completed. Which approach best fits?
11. A grounded assistant returns fluent answers that cite irrelevant passages. Which approach best fits?
12. A model development team needs reproducible experiments with training parameters, artifacts, and comparable metrics. Which approach best fits?
13. A model experiment must be reproducible, and the team must be able to identify which data and implementation produced its metrics. Which two records are necessary? Select 2 answers.
14. A model fits on one accelerator, but training time should be reduced across several devices using independent example batches. Which approach best fits?
15. A model registry must distinguish approved production versions from experimental artifacts. Which approach best fits?
16. A model service needs predictable latency after quiet periods and accepts the cost of standing capacity. Which approach best fits?
17. A model should prioritize catching costly positive cases even if false positives require additional review. Which approach best fits?
18. A model team shares results as screenshots without recording evaluation dataset versions. Which approach best fits?
19. A model update improves accuracy but sharply increases inference cost and latency beyond the approved budget. Which approach best fits?
20. A model's apparent benefit comes from identifiers that correlate with the training sample but are unstable in deployment. Which approach best fits?
21. A model's input contains new categories outside the training vocabulary and an increased missing-value rate. Which approach best fits?
22. A model's state exceeds one device's memory, requiring its computation or parameters to be divided. Which approach best fits?
23. A newly trained model must not deploy unless it meets defined evaluation and risk thresholds. Which approach best fits?
24. A pipeline change should be tested before it retrains production models on confidential data. Which approach best fits?
25. A pipeline failure should rerun only safe incomplete work without duplicating registrations or deployments. Which approach best fits?
26. A pipeline reruns unchanged deterministic preprocessing on the same versioned input and configuration. Which approach best fits?
27. A production classifier serves requests reliably, but user outcomes may degrade as data changes. Which two monitoring areas should the team cover in addition to infrastructure health? Select 2 answers.
28. A replacement model should receive only a small initial share of production requests while the existing model remains available. Which approach best fits?
29. A serving model is slow because it reloads large artifacts for every request. Which approach best fits?
30. A smaller model meets task quality while substantially reducing latency and cost on representative workloads. Which approach best fits?
31. A specialized model requires a custom architecture and training code not supported by a low-code workflow. Which approach best fits?
32. A tabular prediction task needs managed low-code training with a suitable labeled dataset. Which approach best fits?
33. A team needs a managed standard image-understanding capability and has no labeled custom image dataset. Which approach best fits?
34. A team needs repeatable preprocessing, training, evaluation, and registration steps with tracked dependencies. Which approach best fits?
35. A team needs to compare foundation models against its task rather than selecting solely by parameter count. Which approach best fits?
36. A team needs to identify which input dataset and code produced a registered model version. Which approach best fits?
37. A time-series forecasting model is evaluated using random splits that place future observations in training and earlier ones in validation. Which approach best fits?
38. A training container must run reproducibly with its required libraries on a managed training service. Which approach best fits?
39. A training dataset changes silently between experiment runs, making comparisons unreliable. Which approach best fits?
40. A training job reads the same large immutable dataset each epoch and repeatedly waits on remote input. Which approach best fits?
41. A training pipeline uses mutable source data without recording the exact snapshot. Which approach best fits?
42. A use case requires visual content generation, not a numerical forecast from structured history. Which approach best fits?
43. An application needs the same embedding model for indexed documents and incoming queries. Which approach best fits?
44. An application requires a prediction in the user's request path within a bounded latency target. Which approach best fits?
45. An endpoint needs alerts for rising latency and error rate even when the model's statistical accuracy remains unchanged. Which approach best fits?
46. An endpoint saturates because individual requests are expensive. The team needs evidence before choosing larger hardware. Which approach best fits?
47. An experiment compares models using a test set repeatedly to choose hyperparameters. Which approach best fits?
48. An inference application retries a failed request that triggers an external business action. Which approach best fits?
49. An ML pipeline runs with an administrator identity because multiple steps need different resources. Which approach best fits?
50. An online endpoint handles variable demand and should add capacity based on supported serving metrics. Which approach best fits?
51. Automatic model promotion must retain a known-good rollback version. Which approach best fits?
52. Labels for the same task disagree across annotators. The team needs a reliable supervised-learning target. Which approach best fits?
53. Millions of records require predictions overnight, and interactive response latency is unnecessary. Which approach best fits?
54. Outcome labels become available and show worsening predictive performance despite stable service uptime. Which approach best fits?
55. Production input distributions diverge substantially from the training dataset, but outcome labels arrive much later. Which approach best fits?
56. Retraining should begin when approved freshness or quality conditions are met, rather than on every incoming prediction. Which approach best fits?
57. SQL analysts need a supported predictive model close to data already in BigQuery, avoiding an unnecessary separate training system. Which approach best fits?
58. The team needs to compare model quality and resource consumption before scaling a prototype to production-sized training. Which approach best fits?
59. Training cost rises without measurable quality improvement as a run continues. Which approach best fits?
60. Training runs are dominated by repeated manual hyperparameter guesses. Which approach best fits?

Comments — 0