CertifyAIxAce every AI certification.

PMLE sample questions with answers

10 free practice questions for the Professional Machine Learning Engineer exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Architecting low-code AI solutions

Sablefin Advisory runs a Gemini-based assistant over a 300-page policy handbook, resending the whole handbook as context on every request, and the monthly bill is dominated by input tokens. Two workloads exist: an interactive chat used by 200 advisers during office hours, and an overnight job producing 60,000 compliance summaries with no deadline before 07:00. Finance wants a large reduction in spend without degrading interactive answer quality. What should you do?

  1. A.

    Put the handbook into Gemini Enterprise Agent Platform context caching for the interactive chat, and move the 60,000 overnight summaries to batch prediction.

  2. B.

    Reduce the maximum output token limit on both workloads so that every response is shorter and fewer output tokens are billed each month.

  3. C.

    Buy Provisioned Throughput for the interactive chat so that input tokens are no longer billed at all.

  4. D.

    Increase the temperature on both workloads so the model spends fewer sampling steps per response.

Show answer

Answer: A

Context caching removes the cost of resending the same handbook on every interactive call, and batch prediction serves the deadline-free overnight job at a lower price than online requests.

  • A. Caching bills the repeated handbook prefix at the lower cached-input rate without changing what the model sees, and batch prediction prices the deadline-free overnight generations below online requests.
  • B. The bill is dominated by input tokens, so capping output length saves little, and shorter answers degrade the interactive quality Finance wants preserved.
  • C. Provisioned Throughput is a capacity reservation aimed at availability; it changes how you are billed but does not make input tokens free.
  • D. Temperature controls sampling randomness, not the number of tokens processed, so it changes output style rather than cost.
Question 2Architecting low-code AI solutions

Hallowmere Pharma wants to flag possible adverse events from 22 million tabular case records held in BigQuery, of which only about 1,400 are labelled positive. The safety team has two weeks to test roughly 30 feature variants suggested by clinicians, and each variant must be scored at a fixed recall operating point before the next one is tried. Four analysts write SQL and have no ML platform of their own, the department’s ML budget is small, and nobody has yet asked for the model to be served outside BigQuery. The class imbalance must be handled rather than ignored. What should you do?

  1. A.

    Train a BigQuery ML BOOSTEDTREECLASSIFIER with AUTOCLASSWEIGHTS=TRUE for each feature variant, and compare candidates with ML.EVALUATE and ML.ROC_CURVE at the chosen operating point.

  2. B.

    Randomly downsample the negative records until the classes are balanced, then train a BigQuery ML LOGISTIC_REG model on each feature variant.

  3. C.

    Train a BigQuery ML model with MODELTYPE='AUTOMLCLASSIFIER' for each feature variant so that BigQuery performs the architecture search.

  4. D.

    Train a Gemini Enterprise Agent Platform AutoML tabular classification model for each feature variant with AUC-PR as the optimisation objective, and compare the evaluation results.

Show answer

Answer: A

The deciding constraint is iteration cadence on a small budget: a boosted tree in BigQuery ML trains in minutes per variant and AUTOCLASSWEIGHTS handles the imbalance without discarding data.

  • A. Minutes per training run at query pricing sustains thirty variants in two weeks, and AUTO_CLASS_WEIGHTS handles imbalance without discarding negatives.
  • B. Downsampling to balance against 1,400 positives throws away almost all 22 million negatives and distorts the predicted probabilities.
  • C. AUTOML_CLASSIFIER delegates to the same managed AutoML training, so staying in SQL does not avoid its per-run node-hour cost or duration.
  • D. Each AutoML run costs a node-hour training budget and takes hours, so thirty variants exceed both the two-week window and the small budget.
Question 3Architecting low-code AI solutions

Northbay Telecom trains a Gemini Enterprise Agent Platform AutoML tabular classifier that predicts whether a subscriber will churn in the next 30 days. Offline the model reports 0.93 AUC-PR, but when the campaign team used it on last month's subscribers the lift almost vanished. The dataset covers 24 months, each row is one subscriber-month, and the default random data split was used. Churn behaviour shifted noticeably after a tariff change nine months ago. You must make the offline metric predictive of live performance. What should you do?

  1. A.

    Enable a manual data split that assigns 80 percent of subscribers to training and 20 percent to test, keyed on subscriber ID.

  2. B.

    Configure a chronological data split on the subscriber-month timestamp so that training, validation and test sets cover successive time ranges.

  3. C.

    Switch the optimisation objective from AUC-PR to log loss so the probabilities are better calibrated.

  4. D.

    Increase the AutoML training budget in node-hours so the search explores more model architectures.

Show answer

Answer: B

A random split over 24 months of subscriber-months leaks future periods into training, so only a chronological split produces a test metric that reflects live use.

  • A. Grouping by subscriber removes identity leakage but still shuffles months, so recent behaviour continues to leak into training.
  • B. A chronological split trains on the past and tests on the future, which mirrors production and reveals the post-tariff regime change.
  • C. Changing the objective alters calibration, not the split; the inflated score comes from the data partition, not from the metric.
  • D. More node-hours refine a model that is being measured wrongly; a better architecture cannot fix an evaluation that leaks the future.
Question 4Architecting low-code AI solutions

Ashlar Home Furnishings wants to generate lifestyle imagery for its product listings. Demand is spiky: nothing for several days, then roughly 4,000 images in one afternoon when a range launches. Two engineers maintain the entire data platform, neither has operated GPU infrastructure, and the company holds no accelerator quota in its region and does not intend to request any. Prompting with reference images already produces acceptable results, so the model weights do not need to change, and marketing will not accept paying for capacity that sits idle between launches. What should you do?

  1. A.

    Deploy an open-weight image generation model from Model Garden to a Gemini Enterprise Agent Platform endpoint with GPU replicas and autoscaling.

  2. B.

    Self-host an open-weight image generation model on a GKE cluster with GPU node pools behind an inference service the team builds.

  3. C.

    Fine-tune an open-weight image generation model on the product catalogue and deploy the tuned version to a dedicated endpoint.

  4. D.

    Call a managed image generation model offered as a service in Model Garden through the Gemini Enterprise Agent Platform API, paying per request with no endpoint to provision or operate.

Show answer

Answer: D

With no weight customisation needed, no accelerator quota and long idle periods, a managed model-as-a-service offering removes both the idle cost and the GPU operations the team cannot take on.

  • A. A deployed model keeps at least one GPU replica warm, so it bills for idle accelerators between launches and requires quota the company will not request.
  • B. Running GPU node pools and a bespoke inference service is the heaviest operational burden, which two engineers with no GPU experience cannot sustain.
  • C. Tuning is unnecessary because prompting with reference images already meets the requirement, and it adds a training workload plus a dedicated endpoint.
  • D. A model served as a service needs no endpoint, no accelerator quota and no GPU operations, and per-request billing costs nothing during multi-day idle periods.
Question 5Architecting low-code AI solutions

Meridian Airways is launching a Gemini-powered rebooking assistant that must stay responsive during irregular operations, when traffic rises roughly tenfold for several hours with no warning. In a load test the service began returning 429 resource-exhausted errors at the peak rate, and the airline's operations agreement requires predictable throughput during those windows rather than best-effort capacity. Cost is secondary to reliability during disruptions. The assistant runs in a single region and its model version is pinned for change-control reasons. What should you do?

  1. A.

    Deploy the assistant in three additional regions and round-robin requests across the regional Gemini endpoints.

  2. B.

    Add exponential backoff with jitter to the client and retry every 429 response until the request succeeds, accepting the added latency and the growing queue of waiting passengers during irregular operations.

  3. C.

    Switch the assistant to a smaller Gemini Flash-class model so each request consumes less quota.

  4. D.

    Purchase Gemini Enterprise Agent Platform Provisioned Throughput for the model and region, sized for the disruption peak, and let overflow traffic fall back to pay-as-you-go.

Show answer

Answer: D

Provisioned Throughput reserves a guaranteed, fixed rate of generative AI capacity, which is what an availability commitment during unpredictable peaks requires.

  • A. Multi-region fan-out multiplies best-effort quota without guaranteeing any of it, and adds data residency and consistency complexity.
  • B. Backoff smooths bursts but cannot create capacity; during a multi-hour tenfold surge it turns 429s into unacceptable queueing delay.
  • C. A smaller model cuts cost and latency but still draws on shared capacity, so the same surge will exhaust it and cost is explicitly secondary.
  • D. Provisioned Throughput reserves deterministic capacity for a model and region, and overflow can spill to on-demand instead of failing.
Question 6Architecting low-code AI solutions

Calderon Legal Services stores 140,000 reviewed contract clauses in BigQuery, each with a house-style risk summary written by a paralegal. Zero-shot Gemini summaries are accurate but ignore the firm's fixed three-sentence format and its risk vocabulary, and prompt engineering has not fixed the formatting reliably. The data governance officer forbids copying clause text out of BigQuery, and the analytics team writes SQL rather than Python. Roughly 8,000 new clauses are reviewed each month, and the tuned behaviour must be callable from the same nightly SQL job that already produces the review queue. What should you do?

  1. A.

    Export the clause pairs to Cloud Storage as JSONL and start a Gemini Enterprise Agent Platform supervised tuning job from the Google Cloud console.

  2. B.

    Use ML.GENERATE_EMBEDDING on the clauses and train a BigQuery ML logistic regression model to produce the summaries.

  3. C.

    Run supervised fine-tuning of the Gemini model from BigQuery ML by creating a remote model whose CREATE MODEL statement selects the clause and paralegal summary columns as prompt and label.

  4. D.

    Create a BigQuery remote model over Gemini and call ML.GENERATE_TEXT with a longer system instruction that restates the three-sentence rule.

Show answer

Answer: C

BigQuery ML can launch supervised tuning of a Gemini model directly from a table of prompt and label columns, keeping the clause text inside BigQuery and the workflow in SQL.

  • A. Exporting clause text to Cloud Storage violates the governance rule that the text must not leave BigQuery.
  • B. Embeddings plus logistic regression produce a class label or a similarity score, not free-form generated summaries.
  • C. BigQuery ML tunes a remote Gemini model from prompt and label columns, so the clause text never leaves BigQuery and the team stays in SQL.
  • D. The scenario states prompt engineering has already failed to enforce the format; a longer instruction repeats an approach that did not work.
Question 7Architecting low-code AI solutions

An analytics engineer at Vellmark Insurance is building a BigQuery ML boosted-tree classifier that scores motor policies for renewal risk. The raw table holds skewed premium amounts, a free-text vehicle band, and a signup timestamp. During a review you discover that the notebook applies bucketing and log scaling in a hand-written SQL view before training, and that a second, slightly different view is used when the scoring query runs each night. The team wants one definition of the features and no drift between training and scoring. What should you do? (Choose TWO.)

Choose 2.

  1. A.

    Use the built-in preprocessing functions such as ML.BUCKETIZE, ML.QUANTILEBUCKETIZE and ML.FEATURECROSS inside that TRANSFORM clause instead of hand-written CASE expressions.

  2. B.

    Keep both views but add a scheduled query that compares their row counts every night and alerts the team when the counts differ.

  3. C.

    Enable ENABLEGLOBALEXPLAIN in the model options so ML.GLOBAL_EXPLAIN rewrites the scoring features to match training.

  4. D.

    Materialise the training view into a static table and point both training and scoring at that table.

  5. E.

    Move every preprocessing expression into the TRANSFORM clause of the CREATE MODEL statement so the same transformations are stored with the model.

Show answer

Answer: A, E

The TRANSFORM clause stores preprocessing inside the model so ML.PREDICT replays it automatically, and BigQuery ML's preprocessing functions express those steps declaratively.

  • A. BigQuery ML preprocessing functions such as ML.QUANTILE_BUCKETIZE fit statistics during training and reapply the same ones at prediction time.
  • B. Row-count monitoring detects nothing about differing expressions; two views can agree on row count and still compute different features.
  • C. ML.GLOBAL_EXPLAIN reports feature attributions; it never modifies or aligns input features.
  • D. A frozen table stops new policies from being scored and still leaves the transformation logic outside the model where it can diverge again.
  • E. TRANSFORM serialises preprocessing into the model, so ML.PREDICT replays it on raw input and the second view disappears.
Question 8Architecting low-code AI solutions

Bewcastle Leisure forecasts weekly visitor numbers with a BigQuery ML ARIMA_PLUS model. The commercial team accepts the numbers but the finance director will not sign off on a capital plan built on them until she can see how much of next quarter's rise is underlying trend, how much is the usual summer pattern and how much comes from the holiday effects the model was told about. The team works in SQL and does not want to rebuild the model. What should you do?

  1. A.

    Call ML.GLOBAL_EXPLAIN on the model to obtain the mean absolute attribution of each component across the forecast horizon.

  2. B.

    Call ML.FORECAST with a higher confidence level so the returned prediction interval shows the contribution of each component.

  3. C.

    Retrain the model with DECOMPOSETIMESERIES set to FALSE and compare the two sets of forecasts to isolate the seasonal effect.

  4. D.

    Call ML.EXPLAIN_FORECAST on the model, which returns the forecast together with its trend, seasonal and holiday components for each time point.

Show answer

Answer: D

ML.EXPLAIN_FORECAST is the extended form of ML.FORECAST that returns the decomposed trend, seasonal and holiday components alongside the forecast values.

  • A. ML.GLOBAL_EXPLAIN gives feature attributions for supervised models; it does not apply to ARIMA time-series components.
  • B. The confidence level controls the width of the prediction interval and says nothing about which component drives the level.
  • C. Differencing two retrained models approximates by experiment what ML.EXPLAIN_FORECAST returns directly, at far higher cost.
  • D. ML.EXPLAIN_FORECAST extends ML.FORECAST with the trend, seasonal and holiday decomposition the director is asking for.
Question 9Architecting low-code AI solutions

An assistant at Winterbourne Housing answers tenant questions. It follows the required tone and structure well, but it invents repair policy details and quotes charges that changed last month. The policy corpus is updated weekly by a different team. An engineer proposes another supervised fine-tuning round on the latest policy documents so the model "knows" the current rules. What should you advise?

  1. A.

    Lower the temperature and the top-p value so the model is less likely to invent details, and leave the weekly policy corpus out of the design entirely because sampling settings already constrain the output.

  2. B.

    Ground the assistant on the policy corpus rather than tuning on it, because grounding connects output to verifiable sources and supplies facts that change weekly, while tuning teaches behaviour.

  3. C.

    Increase the number of tuning epochs so that the policy documents are memorised more thoroughly by the model weights.

  4. D.

    Proceed with the fine-tuning round, and repeat it weekly as the policy corpus changes, because tuned knowledge is more reliable than retrieved context.

Show answer

Answer: B

Facts that change weekly belong in grounding, which tethers output to verifiable sources and returns citations; tuning changes behaviour and cannot keep pace with a moving corpus.

  • A. Sampling parameters do not supply missing facts, and the newest models' reference recommends omitting them entirely.
  • B. Grounding supplies changing facts with citations and leaves the tuned behaviour intact, which is the split the symptoms describe.
  • C. More epochs makes the model more confident in facts that are already going stale, and does not add auditability.
  • D. Weekly retuning is slow and costly, still produces no citations, and bakes facts into weights that change again.
Question 10Architecting low-code AI solutions

A tuning lead at Broughton Astley Insurance has run a supervised fine-tuning job on a Gemini model and wants to answer two governance questions before promoting anything: whether an earlier, less over-fitted state of the run can be selected instead of the final one, and whether next quarter's additional examples will require starting again from the base model. Select TWO statements that correctly describe the platform's tuning features. (Choose TWO.)

Choose 2.

  1. A.

    Only the final state of a tuning run is retained, so an earlier state can be recovered only by rerunning with fewer epochs.

  2. B.

    Tuning checkpoints let you save tuning progress, compare checkpoint performance and choose the best performing checkpoint.

  3. C.

    Every additional example requires tuning again from the base model, because a tuned model cannot be used as the starting point for further tuning.

  4. D.

    Checkpoints are available for every model in Model Garden, including self-deployed open models and partner models.

  5. E.

    Continuous tuning lets you continue tuning an already tuned model or model checkpoint by adding more epochs or training examples.

Show answer

Answer: B, E

Tuning checkpoints let you compare and select an intermediate state, and continuous tuning lets a tuned model or checkpoint be extended with more epochs or examples.

  • A. Checkpoints retain intermediate states precisely so that an earlier one can be selected without rerunning.
  • B. Checkpoints save tuning progress and allow comparison and selection of the best performing intermediate state.
  • C. Continuous tuning exists so that a tuned model can be the starting point for further tuning.
  • D. Checkpoint support is documented for specific Gemini models, not for every model in Model Garden.
  • E. Continuous tuning continues from an already tuned model or checkpoint with more epochs or examples.

Keep going with 531 more PMLE questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all PMLE papers