PMLE mock exam

Question 1, Architecting low-code AI solutionsAn analyst team that knows SQL needs a churn model from data already in BigQuery, as fast as possible. What fits?

• AA hand-built PyTorch model
• BFine-tuning a Gemini model
• CBigQuery ML
• DA custom TensorFlow job on GKE

Question 2, Collaborating within and across teams to manage data and modelsFeatures must be defined once and reused consistently across several models. What do you use?

• AFeature Store on Agent Platform
• BA Cloud Run service that recomputes features
• CModel Armor with shared feature rules
• DA shared CSV file in Cloud Storage

Question 3, Scaling prototypes into ML modelsA large model will not fit into a single GPU's memory during training. Which strategy addresses this?

• AData parallelism across devices
• BGradient accumulation on one GPU
• CModel parallelism across devices
• DLowering the learning rate

Question 4, Scaling prototypes into ML modelsWhen is fine-tuning a foundation model the wrong first step?

• AWhen prompting alone has been tried and has stopped improving results
• BWhen the model only lacks access to current company documents
• CWhen output needs a very specific style that prompts cannot hold
• DWhen a task needs consistent domain behavior across many requests

Question 5, Serving and scaling modelsYou want to compare a new model against the current one with real traffic and limited risk. What do you do?

• AAdd more replicas and move all traffic at once
• BSend a small share of traffic in a canary
• CReplace the current model and watch the dashboards
• DTest only on the held-out training data

Question 6, Automating and orchestrating ML pipelinesOnline predictions are worse than offline tests, though the model is unchanged. Preprocessing is coded separately in training and serving. What is the likely cause?

• ATraining-serving skew
• BToo few epochs during training
• CSlow hardware on the serving nodes
• DConcept drift in the live data

Question 7, Monitoring AI solutionsThe relationship between inputs and the target has changed, even though input distributions look stable. What is this?

• AData drift
• BConcept drift
• CLabel leakage
• DQuota exhaustion

Question 8, Monitoring AI solutionsA Gemini app must screen prompts for malicious input and sensitive data. Which tool is named for this?

• AColab Enterprise
• BCloud Build
• CDataflow
• DModel Armor