AWS Machine Learning Engineer Associate (MLA-C02) - Full Practice Exam 1 - Beta Blueprint

1. A Bedrock agent should use an internal tool, but fulfillment must remain in the existing application service for authorization and transaction control. Which feature fits?
2. A binary classifier reports 99% accuracy on a dataset with 1% positives but predicts every example as negative. Which conclusion is correct?
3. A categorical feature gains new values after deployment. The training encoder fails on unknown categories. What is the best preparation improvement?
4. A chatbot safety test suite contains only ordinary questions. What additional testing is needed?
5. A company needs to reduce an FM adaptation job memory footprint while changing a limited set of model parameters. Which technique should be evaluated for supported frameworks?
6. A company scores several million records once per night and does not need an always-running endpoint. Which SageMaker option is most appropriate?
7. A company wants an FM to answer from policies updated every day and cite the current source. Which approach is the best initial baseline?
8. A custom inference container fails endpoint health checks before receiving traffic. What should the engineer inspect first?
9. A customer churn dataset contains a cancellation timestamp that is populated only after the customer leaves. Offline validation is excellent, but predictions at signup fail. What is the most likely issue?
10. A customer-facing Bedrock Flow uses a mutable draft and changes unexpectedly when developers experiment. What should production use?
11. A dataset contains several records from each device. Random row splitting gives strong validation, but the model fails on new devices. Which correction best tests generalization to unseen devices?
12. A demand forecasting model will predict next month from past sales. Which validation split best reflects this deployment?
13. A deployed model receives inputs with a substantially different distribution from its training baseline. Labels are not yet available. Which monitoring capability addresses this observation?
14. A deployment pipeline rebuilds an inference image from floating dependencies after model approval. Production behavior differs from testing. Which practice best prevents this?
15. A developer tunes a prompt repeatedly against one benchmark and then reports that same benchmark as final quality. What should be added?
16. A drift alert automatically starts retraining. The candidate is worse on an important cohort. What should happen before deployment?
17. A feature engineering job repeatedly scans a large text dataset to select a few columns for one date. Which data organization should be evaluated?
18. A fraud classifier produces 80 true positives, 20 false positives, and 40 false negatives. What is its precision for the positive class?
19. A fraud model needs historical features as they existed at each transaction time. A join currently uses each customer latest feature values. What is required?
20. A Glue job incrementally processes source files. Repeated full reprocessing is expensive, and stable supported source state exists. Which capability should be evaluated?
21. A GPU endpoint is expensive but shows low GPU utilization and repeated small requests. What should the engineer evaluate before adding more GPUs?
22. A labeling project has low agreement between annotators on an important class. What should happen before scaling annotation?
23. A lake contains shared training data, but teams should access only approved tables and columns. Which governance service should be evaluated?
24. A long training job can tolerate interruption if progress is saved. Which two features support a cost-conscious SageMaker training design? Select 2 answers.
25. A model endpoint meets its CPU target but users experience frequent failed predictions. Which operating practice best aligns monitoring with the product?
26. A model is trained on one-hot encoded columns, but serving sends raw categorical strings directly to the estimator. What is the primary issue?
27. A model loads a large artifact during endpoint startup. A rollout times out before the container becomes ready. Which response is most appropriate?
28. A model must follow a specific response style, but the business facts are supplied separately at inference time. Which customization objective is most appropriate?
29. A model registry version is created after training. What information best supports reproducibility and review?
30. A model response cache serves one tenant answer to another tenant after a key collision. What is the essential correction?
31. A multi-step workflow must run a SageMaker job, wait for completion, then call another AWS service with defined retries. Which service is appropriate for cross-service orchestration?
32. A new model version must receive a small amount of production traffic before a full rollout. The team needs automatic rollback on bad endpoint metrics. Which deployment capability fits?
33. A new source changes a numeric column into free text and the preprocessing job silently emits nulls. Which control should catch this before training?
34. A pipeline computes normalization statistics using the complete dataset before splitting into training and test sets. What should change?
35. A pipeline should deploy only models whose validation metric exceeds an approved threshold. Which design is most appropriate?
36. A platform team wants repeatable ML infrastructure in multiple environments without manual console drift. Which two practices are appropriate? Select 2 answers.
37. A prediction pipeline needs near-real-time device events with multiple consumers and replay of retained data. Which ingestion service fits?
38. A preprocessing change alters feature order but the production estimator remains unchanged. Which deployment practice best prevents this class of failure?
39. A pretrained image model already recognizes general visual features, but the company has a small labeled dataset for its own product categories. Which approach should be evaluated first?
40. A production model version causes a quality regression. The previous version and its runtime artifacts remain available. What is the most appropriate response?
41. A protected training job must not make arbitrary internet calls. Which configuration should be evaluated according to supported job requirements?
42. A RAG index embeds documents with one model but queries with an unrelated embedding model. Similarity results become unreliable. What should change?
43. A RAG pipeline splits procedural documents into tiny fragments and loses required prerequisite context. Which preparation change should be evaluated?
44. A RAG source document contains instructions to ignore user permissions and reveal other project data. What is the correct defense?
45. A real-time endpoint has rising latency and errors during peaks. Which two investigations best distinguish capacity pressure from application failures? Select 2 answers.
46. A recovery environment contains the model artifact but lacks the execution role, container image, and endpoint configuration. What is the correct assessment?
47. A regression task penalizes a few very large errors more strongly than many small ones. Which metric should be considered alongside the business loss?
48. A SaaS system serves many similarly structured models with modest individual traffic. Keeping a separate endpoint for each is costly. Which approach should be evaluated?
49. A SageMaker execution role has administrator permissions but only needs one training prefix and one output prefix. What is the best improvement?
50. A SageMaker processing role can list an S3 bucket but cannot read SSE-KMS-encrypted training objects. What should be checked?
51. A screening workflow must reduce missed positive cases and can tolerate more manual review. Which threshold experiment is most appropriate?
52. A serving requirement includes resilience to a single instance failure and continued capacity during recovery. What should the deployment plan validate?
53. A small model has intermittent traffic and can tolerate startup latency. The team wants to avoid managing a continuously running instance fleet. Which option should be evaluated if its constraints fit?
54. A source dataset is withdrawn from approved use. What should the ML team determine?
55. A tabular model requires handling missing numeric values. The team plans median imputation. Which implementation is most appropriate?
56. A team compares FMs for a narrow extraction task. Which two measurements best support model selection? Select 2 answers.
57. A team must reproduce last month training run after source files have changed. Which preparation practice is most important?
58. A team needs to compare training runs with different datasets, parameters, and metrics. Which SageMaker capability should be evaluated?
59. A team wants to compare a candidate model on real request traffic without returning its predictions to users. Which feature should be evaluated?
60. A team wants to measure production prediction accuracy after outcomes arrive several days later. What must the monitoring pipeline support?
61. A team wants to search learning rate and tree-depth settings without manually launching every training job. Which SageMaker capability fits?
62. A text classifier performs well on short examples but fails on long documents because its tokenizer truncates the decisive section. What should be investigated?
63. A training job fails immediately because the configured algorithm expects a supported data format different from the supplied files. What is the best correction?
64. A training pipeline must validate required columns and numeric ranges before starting an expensive job. Which design is best?
65. A training run is reproducible in code but results vary because random seeds and dependency versions differ. What should be recorded?
66. A training run produces unstable loss after the learning rate is increased substantially. What should the engineer test first?
67. A tuning job minimizes an objective metric, but the metric is validation accuracy. The selected model is unexpectedly poor. What should be corrected?
68. An AI agent proposes a tool call for another user account. What must the executor do before acting?
69. An application needs low-latency feature lookup for live predictions and historical feature data for training. Which SageMaker capability is most relevant?
70. An application needs synchronous predictions with sustained low latency throughout the day. Which SageMaker inference option should be evaluated first?
71. An approved model must be promoted from experimentation to production with an auditable review gate. Which mechanism is most relevant?
72. An endpoint experiences predictable daily peaks. What should guide autoscaling configuration?
73. An endpoint returns errors after clients send a new unsupported feature shape. What should be implemented?
74. An existing SageMaker Clarify customer needs to understand which features influence predictions and assess unwanted bias across selected groups. Which capability should the team evaluate?
75. An image classifier has few labeled examples and must be robust to realistic orientation changes. Which preparation strategy should be evaluated?
76. An inference log contains sensitive raw inputs that are not needed for debugging. Which two actions are appropriate? Select 2 answers.
77. An ML dataset includes personal identifiers not required for the prediction task. Which two preparation measures are appropriate? Select 2 answers.
78. Bedrock request costs rise after a prompt update. Which monitoring is most useful for diagnosis?
79. Development training jobs continue running after experiments are abandoned. What controls best reduce avoidable spend?
80. Inference requests have large inputs and processing times beyond the normal interactive budget. Users can retrieve results later. Which SageMaker mode should be evaluated?
81. Input distributions remain stable, but the relationship between features and outcomes changes and quality declines. Which issue is most consistent with this?
82. Only 1% of examples belong to the positive class. A small random split leaves almost no positives in validation. What should the data engineer evaluate?
83. Rare-event oversampling is proposed for a classifier. Which implementation best preserves credible evaluation?
84. Training loss decreases steadily while validation loss rises after several epochs. Which two changes should be evaluated? Select 2 answers.
85. Two models have similar average quality, but one performs poorly on an important low-volume customer cohort. Which selection process is appropriate?

Comments — 0