AMAADOR ACADEMYCertificatesVerify a certificate

PRO-GCP-ML

Final check — Google Professional Machine Learning Engineer — Exam-Prep

Answer the 45 questions. You need 80% or more to pass. You can retry as many times as you like.

Study the course →

Question 1A team can solve a routing decision with a fixed, well-understood business rule and has no data to learn from. What is the best recommendation?
Question 2Which metric is most appropriate for a highly imbalanced fraud-detection problem (1% positives)?
Question 3A SQL-fluent analytics team needs a model quickly on tabular data already in BigQuery. Which service fits best?
Question 4What is the primary purpose of Vertex AI Feature Store?
Question 5Training-serving skew is most often caused by which of the following?
Question 6You need to score 50 million rows once per night where latency does not matter. Which serving mode is most cost-effective?
Question 7Which Vertex AI capability automates hyperparameter search using Bayesian optimization?
Question 8A team wants a high-quality image classifier but has little ML expertise and a modest budget. What is the best first choice?
Question 9Which explainability method is best suited to a differentiable deep neural network?
Question 10What does the BigQuery ML TRANSFORM clause provide?
Question 11Which scenario best justifies distributed training across multiple workers?
Question 12Prediction drift refers to which phenomenon?
Question 13To prevent label leakage when building training data with historical features, you should:
Question 14Which combination best secures a sensitive Vertex AI prediction workload?
Question 15In Google's MLOps maturity model, an automated training pipeline triggered by new data corresponds to which level?
Question 16Which service is best for serverless streaming and batch ETL to prepare ML data?
Question 17A robust deployment pipeline should deploy a new model version only when:
Question 18Which approach should you try FIRST to improve a foundation model's output before tuning?
Question 19Removing a protected attribute (e.g., gender) from features is insufficient for fairness because:
Question 20Which Vertex AI component is the central catalogue tracking model versions, lineage and metrics?
Question 21For a common, well-defined task like detecting objects in images with minimal effort, what is typically the fastest route?
Question 22Why should production preprocessing live inside a pipeline component or the model graph rather than a notebook cell?
Question 23Which model type in BigQuery ML is designed for time-series forecasting?
Question 24Vertex AI Vizier with trial early-stopping primarily helps you:
Question 25Traffic splitting on a Vertex AI endpoint is most useful for:
Question 26Which technique reduces model serving latency and size while keeping acceptable accuracy?
Question 27Slice-based evaluation across subgroups is primarily used to:
Question 28With limited labeled images, which approach typically gives the best accuracy for the least data and compute?
Question 29Which option best reduces cost for a fault-tolerant, restartable training job?
Question 30Detected feature/prediction drift in production should ideally:
Question 31Which is the correct ordering from least to most operational control/effort?
Question 32What is a primary benefit of caching steps in a Vertex AI / Kubeflow pipeline?
Question 33To de-identify PII before it enters training data, which Google Cloud service is appropriate?
Question 34Which Vertex AI capability lets you track and compare runs, parameters and metrics across experimentation?
Question 35Model Garden on Vertex AI is best described as:
Question 36Which method is best suited for explaining a tree-ensemble (non-differentiable) tabular model?
Question 37You should point a production endpoint at a model version ALIAS rather than a fixed version number because:
Question 38Grounding (RAG) a generative model on your own authoritative data primarily helps to:
Question 39Which is the correct definition of training-serving skew (as opposed to drift)?
Question 40A recommendation system on user-item interactions in BigQuery is best built with:
Question 41Which choice best advances a project from MLOps level 0 toward level 1?
Question 42When two valid options both meet the requirements, the PMLE exam usually favors the one that is:
Question 43Model cards are primarily used to:
Question 44Which is the best reason to choose TPUs over GPUs for a training job?
Question 45What is the recommended way to confirm exam readiness in this course?