PMLE sample questions with answers
The other 310 questions are in the practice tool, with answers revealed as you go.
Start 20 questionsArchitecting low-code AI solutions
- A regional bank wants a chat assistant that answers account and product questions. It should follow the bank's policy wording and draw on its published help articles. The team wants to configure this through a console and plain-language guidance, not by writing an intent classifier. Which Google Cloud offering is most appropriate?
- A copywriting group wants to compare several large language models side by side and iterate on prompt wording before choosing one for its product catalog text. Which Google Cloud environment supports that work?
- A retailer wants shoppers to photograph an item and be shown matching SKUs from a 50,000-item catalog using the Vision API's product search feature. Which option lists the setup steps in the correct order?
- A product manager wants to send one prompt to several foundation models in a browser, read the responses side by side, and then choose a model to build on. Which Vertex AI surface is intended for this?
- A copywriting group wants to compare candidate text models from Google, open-source communities, and third-party vendors on the same prompts before choosing one. Which catalog in Vertex AI serves that purpose?
- A media company needs subtitles in 40 languages. Most target languages are common, but one regional language has little public training text and its output quality is poor. What is the most economical plan?
- A telecom wants a chatbot built on a hosted Gemini model to discuss only its plans and devices, and to decline everything else courteously. Which configuration is the most practical?
- A help-desk app receives messages in several languages and sends each one to a language-specific classifier. Before routing, the app needs the language code for each message. Which Google Cloud capability returns that?
Collaborating to manage data and models
- A payments company produces many builds of its fraud scoring model on Vertex AI. Auditors want a record of every build, the one currently serving live traffic, and the evaluation numbers for each build. Which capability should the team use?
- A recommendation model is trained from nightly batch aggregates but scored using live event streams. You need the same feature definitions and values at training and serving time, with correct point-in-time history for training. What design is right?
- An auditor asks which service records, for every pipeline run, the input datasets, the hyperparameters, and the resulting model versions, so that any artifact can be traced back to its origins. Which service answers that?
- A research group wants notebook instances that can query a warehouse dataset but must have no route to the public internet. Which setup meets both requirements?
Scaling prototypes into ML models
- A lender must tell each rejected applicant the main factors behind that specific decision. The model is a gradient boosted classifier deployed on Vertex AI. Which technique provides this?
- A data science lead notices that her classifier scores 0.98 accuracy on the examples used to fit it but only 0.71 on a holdout set drawn from the same population. Which diagnosis fits best?
- An engineer fine-tuning a transformer language model on a single accelerator keeps hitting out-of-memory errors, and the batch size they need is out of reach. Which two changes most directly lower peak memory while keeping training quality?
- While training a churn predictor, the loss on the training split keeps falling, but the loss on a held-out validation split has risen over the last several evaluation rounds. What is the most appropriate stopping rule?
- In a housing price model, floor area, room count and built-up area are almost perfectly correlated, and the coefficient estimates swing wildly between retrains. Which remedy addresses this redundancy?
- A research team trained a model in JAX and now wants to serve it on Vertex AI. None of the prebuilt prediction images include JAX. What is the best way to deploy it?
- A vision model needs seven days per run on one accelerator, but your deadline requires each run to finish within one day. The architecture and dataset must stay the same. What change would most directly meet the deadline?
- You trained a binary classifier and now want to inspect its confusion matrix and precision-recall metrics for version 3 in the Vertex AI console. Where should you look?
- A fraud team's classifier flags very few transactions, and fraud is rare in the data. The business says missing a fraudulent transaction is far worse than a false alarm. Which metric should headline the report?
- A convolutional classifier scores 97% on training batches but only 79% on held-out images. The team wants better generalization without changing the dataset. What is the best next step?
- A multilingual support team is evaluating one language model for three jobs: answering product questions, condensing long tickets, and tagging customer mood. Which evaluation plan gives the most complete picture?
Serving and scaling models
- A fintech app calls a fraud scorer for every card swipe and must answer within 100 milliseconds at the 95th percentile, with traffic that spikes during evenings. Which hosting choice best fits?
- A retailer's recommendation endpoint must keep p99 latency under 50 ms during a flash sale that could bring ten times normal load. Which configuration best meets this requirement?
- An ML team has produced a new version of a recommendation model. They want only a small slice of live requests to reach it at first, with that share raised step by step as dashboards confirm quality. Which rollout approach fits?
- A marketing team must score 40 million customer rows once a week and write the scores back to storage. No one needs answers in real time. Which option is the most economical?
- A logistics firm scores about 100 million shipment rows that live in a warehouse table. A TensorFlow model needs roughly 45 minutes to run on that volume, and the output must be ready before the morning reports. Which option keeps spend lowest while finishing on time?
- A lender runs a credit model trained six months ago that has a strong production record. A newer candidate trained last week scores better on the hold-out set. Leadership wants evidence on live traffic before any customer-facing change. Which rollout is safest?
- After a rollout, p95 latency on a serving deployment has climbed sharply, yet accelerator and CPU graphs on the serving nodes look mostly idle. The model call alone benchmarks quickly when run in isolation. What should be checked first?
Automating and orchestrating ML pipelines
- After writing a Kubeflow pipeline in Python, which artifact must you generate before Vertex AI can run it?
- A retail company drops daily sales files into a storage bucket. Each new file should kick off the training workflow automatically, with no polling and no human action. What is the best design?
- A team applies normalization and vocabulary lookups to features during training, then reimplements them in the serving service. Predictions drift because the two copies diverge. What key advantage does TensorFlow Transform offer over preprocessing written separately for training and serving?
- A fraud-scoring pipeline ingests transaction batches whose columns change: vendors add new fields and retire old ones. The team wants new columns flagged for review rather than silently breaking training, while still tracking gradual value shifts. Which TFX components support this?
- A team wants retraining to start automatically whenever a new data batch passes validation, with a new model promoted to an endpoint only if evaluation thresholds are met. Which pairing fits this continuous training and delivery pattern?
Monitoring AI solutions
- A regulated lender needs a durable record of every scoring call's input, output and deployed model version. Which endpoint capability provides this?
- Regulators require that each declined loan application come with the per-feature contributions that drove the score, computed at the moment of the request. Which capability provides this?
- A drift alert has just fired on several key inputs. What belongs first in the incident runbook?
- A fintech retrains its risk model weekly. A candidate should replace the production model only when its ROC AUC beats the incumbent by a set margin and no subgroup error rate gets worse by more than a small tolerance. The whole decision must happen automatically inside the pipeline. What construct enforces this?
- Attribution monitoring flags that the influence of a single input, 'territory_code', has risen markedly over the last seven days compared with its baseline. Which interpretation is most reasonable?