Google Professional Machine Learning Engineer Practice Test 50
Question 490
During a postdeployment AI safety and reliability review, the following condition occurs: An LLM summarizer produces fluent answers containing unsupported policy details, but the team measures only response latency and vocabulary similarity. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect attachment ingestion, trust boundaries, malicious prompt examples and deployed Model Armor filter results.
- Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation.
- Inspect retrieved evidence, claim-level factual consistency, reference answers and human review samples.
- Compare live and offline feature SQL, window definitions, sample entity values and prediction-time cutoffs.
Explanation
Correct Answer
C:Inspect retrieved evidence, claim-level factual consistency, reference answers and human review samples. This inspects evidence relevant to generative ai factual grounding evaluation. Fluent text or semantic similarity is not proof an answer is supported by the retrieved source material. After identifying the cause, the matching correction is: Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly. This matters here because an LLM summarizer produces fluent answers containing unsupported policy details, but the team measures only response….
Incorrect Answers
A:Inspect attachment ingestion, trust boundaries, malicious prompt examples and deployed Model Armor filter results. This examines model armor untrusted content screening rather than the specific mechanism reported here. It is not the decisive first check for generative ai factual grounding evaluation. The relevant evidence is: Inspect retrieved evidence, claim-level factual consistency, reference answers and human review samples.
B:Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation. This examines model monitoring false-alert seasonality rather than the specific mechanism reported here. It is not the decisive first check for generative ai factual grounding evaluation. The relevant evidence is: Inspect retrieved evidence, claim-level factual consistency, reference answers and human review samples.
D:Compare live and offline feature SQL, window definitions, sample entity values and prediction-time cutoffs. This examines training-serving feature skew detection rather than the specific mechanism reported here. It is not the decisive first check for generative ai factual grounding evaluation. The relevant evidence is: Inspect retrieved evidence, claim-level factual consistency, reference answers and human review samples. This matters here because an LLM summarizer produces fluent answers containing unsupported policy details, but the team measures only response….
Question 491
During a regulated MLOps pipeline deployment, the following condition occurs: An automated retraining job overwrites the previous serving model before verifying quality under the most recent data distribution. The workflow must be repeatable and protect the last approved production model. Which technical explanation best accounts for the reported behavior?
- Continuous training needs statistically and operationally justified policy thresholds rather than reactive promotion of noise.
- Unpinned tags do not establish a trustworthy relationship between code, model artifacts and production rollout.
- Silent casting can hide producer errors and produce invalid model inputs across pipeline boundaries.
- Continuous training should not imply uncontrolled continuous deployment of every new model.
Explanation
Correct Answer
D:Continuous training should not imply uncontrolled continuous deployment of every new model. This explains the mechanism behind continuous training rollback protection and why the described failure is possible. A discriminating investigation is: Review champion model retention, candidate metrics, change approvals and automatic rollback triggers. The corresponding repair is: Retain the last healthy model and require canary validation before switching all prediction traffic.
Incorrect Answers
A:Continuous training needs statistically and operationally justified policy thresholds rather than reactive promotion of noise. This principle describes canary failure feedback into ct policy, which is not the cause of the observed continuous training rollback protection failure. The controlling explanation is: Continuous training should not imply uncontrolled continuous deployment of every new model.
B:Unpinned tags do not establish a trustworthy relationship between code, model artifacts and production rollout. This principle describes cloud build release version promotion, which is not the cause of the observed continuous training rollback protection failure. The controlling explanation is: Continuous training should not imply uncontrolled continuous deployment of every new model.
C:Silent casting can hide producer errors and produce invalid model inputs across pipeline boundaries. This principle describes serving schema validation component, which is not the cause of the observed continuous training rollback protection failure. The controlling explanation is: Continuous training should not imply uncontrolled continuous deployment of every new model.
Question 492
During a production ML inference release on Google Cloud, the following condition occurs: Two releases use the same container tag latest, but the registry retargets it to a different dependency build between deployments. The service must meet its agreed quality and latency requirements. Which configuration or procedure is the most appropriate correction?
- Select supported regional placement or routing close to users while respecting residency and model availability.
- Pin releases to reviewed immutable image digests and record image plus model artifact versions.
- Benchmark a Cloud Run custom container against a managed Agent Platform endpoint and choose for workload fit.
- Deploy a supported private endpoint or constrained private network path with least-privilege caller authorization.
Explanation
Correct Answer
B:Pin releases to reviewed immutable image digests and record image plus model artifact versions. This is the targeted corrective response for model artifact image digest reproducibility. Mutable tags can point to new image content, defeating reproducibility even if the deployment text is unchanged. Verify the diagnosis using: Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. Confirm success using: Identical release manifests resolve to identical image bytes across reproducible deployments.
Incorrect Answers
A:Select supported regional placement or routing close to users while respecting residency and model availability. This action addresses serving region proximity versus model compute, not the failure in the described model artifact image digest reproducibility case. The targeted correction is: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
C:Benchmark a Cloud Run custom container against a managed Agent Platform endpoint and choose for workload fit. This action addresses cloud run versus managed inference endpoint, not the failure in the described model artifact image digest reproducibility case. The targeted correction is: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
D:Deploy a supported private endpoint or constrained private network path with least-privilege caller authorization. This action addresses private endpoint for regulated inference traffic, not the failure in the described model artifact image digest reproducibility case. The targeted correction is: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
Question 493
During model-training validation on a Google Cloud experimental project, the following condition occurs: A demand model trains on moving averages computed after the prediction cutoff, so the held-out forecast benefits from future sales updates. The change should preserve a meaningful independent model-quality comparison. Which result is the strongest postchange evidence of success?
- The training manifest contains traceable permitted sources and safety tests detect known prohibited outputs.
- Held-out predictions are currency-valued and their absolute or squared errors match the financial evaluation contract.
- Every evaluation feature is available at its simulated prediction time and later-period scores remain stable.
- The deployed artifact corresponds to the best held-out epoch rather than the final overfit epoch.
Explanation
Correct Answer
C:Every evaluation feature is available at its simulated prediction time and later-period scores remain stable. This is specific acceptance evidence for temporal training label leakage in rolling windows. Temporal features formed after prediction time leak future observations into model training and evaluation. It confirms the controlled correction: Recompute lagged windows using only prior observations and validate with rolling-origin splits.
Incorrect Answers
A:The training manifest contains traceable permitted sources and safety tests detect known prohibited outputs. This result could demonstrate success for fine-tuning image generator data rights but not for the stated temporal training label leakage in rolling windows issue. The direct test is: Every evaluation feature is available at its simulated prediction time and later-period scores remain stable.
B:Held-out predictions are currency-valued and their absolute or squared errors match the financial evaluation contract. This result could demonstrate success for regression versus classification product choice but not for the stated temporal training label leakage in rolling windows issue. The direct test is: Every evaluation feature is available at its simulated prediction time and later-period scores remain stable.
D:The deployed artifact corresponds to the best held-out epoch rather than the final overfit epoch. This result could demonstrate success for early stopping on validation divergence but not for the stated temporal training label leakage in rolling windows issue. The direct test is: Every evaluation feature is available at its simulated prediction time and later-period scores remain stable.
Question 494
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: Two ML engineers rerun the same notebook and obtain incompatible model artifacts because one environment silently installs a newer Python library release. The handoff must preserve privacy, lineage and testable behavior. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
- Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project.
- Standardize stable entity identifiers and validated feature definitions across training and online serving pathways.
- Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs.
- Push aggregations into BigQuery or stream bounded samples to the notebook rather than materializing the whole table.
Explanation
Correct Answers
A:Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes. This is the targeted corrective response for notebook dependency lock reproducibility. Notebook cells alone do not capture interpreter and package state; reproducibility depends on the execution environment. Verify the diagnosis using: Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. Confirm success using: Independent notebook runs resolve the same approved libraries and produce comparable metric and artifact outputs. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
D:Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. This inspects evidence relevant to notebook dependency lock reproducibility. Notebook cells alone do not capture interpreter and package state; reproducibility depends on the execution environment. After identifying the cause, the matching correction is: Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
Incorrect Answers
B:Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project. This is an action or diagnostic relevant to workbench credentials embedded in repository, not one of the required two steps. In this scenario verify: Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. Then correct: Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
C:Standardize stable entity identifiers and validated feature definitions across training and online serving pathways. This is an action or diagnostic relevant to feature store online and offline entity alignment, not one of the required two steps. In this scenario verify: Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. Then correct: Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
E:Push aggregations into BigQuery or stream bounded samples to the notebook rather than materializing the whole table. This is an action or diagnostic relevant to notebook memory exhaustion on full data extraction, not one of the required two steps. In this scenario verify: Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. Then correct: Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
Question 495
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A CI/CD workflow always deploys the most recent checkpoint even when its validation F1 is below the previous approved baseline. The team requires an idempotent pipeline behavior and traceable evidence. Which targeted engineering change most directly addresses the failure?
- Add a data-validation component that quarantines nonconforming inputs and fails downstream training until reviewed.
- Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling.
- Require candidate evaluation against a governed baseline with an enforced acceptance threshold before promotion.
- Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives.
Explanation
Correct Answer
C:Require candidate evaluation against a governed baseline with an enforced acceptance threshold before promotion. This is the targeted corrective response for model evaluation gate before deployment. Evaluation without a blocking release decision cannot protect production from quality regressions. Verify the diagnosis using: Inspect model-metric artifacts, comparison rules, registry status and deployment dependencies. Confirm success using: An intentionally weak candidate cannot deploy, whereas an approved passing model can proceed with traceable evidence.
Incorrect Answers
A:Add a data-validation component that quarantines nonconforming inputs and fails downstream training until reviewed. This action addresses data validation threshold gate, not the failure in the described model evaluation gate before deployment case. The targeted correction is: Require candidate evaluation against a governed baseline with an enforced acceptance threshold before promotion.
B:Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling. This action addresses airflow task idempotency on retry, not the failure in the described model evaluation gate before deployment case. The targeted correction is: Require candidate evaluation against a governed baseline with an enforced acceptance threshold before promotion.
D:Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives. This action addresses pipeline notification alert on silent failure, not the failure in the described model evaluation gate before deployment case. The targeted correction is: Require candidate evaluation against a governed baseline with an enforced acceptance threshold before promotion.
Question 496
During a production ML inference release on Google Cloud, the following condition occurs: An internal fraud scorer sends sensitive feature vectors to a publicly accessible prediction endpoint even though clients are inside a controlled VPC. The service must meet its agreed quality and latency requirements. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect startup logs, HTTP listening port, health-check expectations and request and response schema.
- Inspect trained class-to-index metadata, endpoint output tensors and gateway mapping version.
- Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures.
- Examine endpoint exposure settings, caller IAM, VPC routes, private connectivity and data egress requirements.
Explanation
Correct Answer
D:Examine endpoint exposure settings, caller IAM, VPC routes, private connectivity and data egress requirements. This inspects evidence relevant to private endpoint for regulated inference traffic. Network exposure and IAM controls are complementary; a public endpoint adds attack surface even with authentication. After identifying the cause, the matching correction is: Deploy a supported private endpoint or constrained private network path with least-privilege caller authorization. This matters here because an internal fraud scorer sends sensitive feature vectors to a publicly accessible prediction endpoint even though….
Incorrect Answers
A:Inspect startup logs, HTTP listening port, health-check expectations and request and response schema. This examines inference container health interface mismatch rather than the specific mechanism reported here. It is not the decisive first check for private endpoint for regulated inference traffic. The relevant evidence is: Examine endpoint exposure settings, caller IAM, VPC routes, private connectivity and data egress requirements.
B:Inspect trained class-to-index metadata, endpoint output tensors and gateway mapping version. This examines custom postprocessing label mapping drift rather than the specific mechanism reported here. It is not the decisive first check for private endpoint for regulated inference traffic. The relevant evidence is: Examine endpoint exposure settings, caller IAM, VPC routes, private connectivity and data egress requirements.
C:Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures. This examines feature online entity key mismatch rather than the specific mechanism reported here. It is not the decisive first check for private endpoint for regulated inference traffic. The relevant evidence is: Examine endpoint exposure settings, caller IAM, VPC routes, private connectivity and data egress requirements.
Question 497
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A shared prototype notebook can read entire sensitive production BigQuery datasets because its attached credentials are inherited from a broadly privileged user. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which targeted engineering change most directly addresses the failure?
- Record input dataset versions, training runs, code revisions and output artifact links in the managed metadata and registry workflow.
- Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data.
- Redact or tokenize unnecessary identifying fields before labeling and limit access to approved mappings.
- Standardize stable entity identifiers and validated feature definitions across training and online serving pathways.
Explanation
Correct Answer
B:Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data. This is the targeted corrective response for colab enterprise notebook data access scope. Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access. Verify the diagnosis using: Review notebook execution identity, effective IAM roles, dataset access policies and audit records for protected tables. Confirm success using: The notebook still trains on approved sample tables while restricted production datasets deny unauthorized reads. This matters here because within a jointly maintained Google Cloud data science workspace, the following condition occurs: A shared prototype….
Incorrect Answers
A:Record input dataset versions, training runs, code revisions and output artifact links in the managed metadata and registry workflow. This action addresses ml metadata lineage for production model, not the failure in the described colab enterprise notebook data access scope case. The targeted correction is: Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data.
C:Redact or tokenize unnecessary identifying fields before labeling and limit access to approved mappings. This action addresses data minimization for labeling inputs, not the failure in the described colab enterprise notebook data access scope case. The targeted correction is: Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data.
D:Standardize stable entity identifiers and validated feature definitions across training and online serving pathways. This action addresses feature store online and offline entity alignment, not the failure in the described colab enterprise notebook data access scope case. The targeted correction is: Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data.
Question 498
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: A custom training image expects a local CSV under the developer's home directory but managed training supplies data through configured Cloud Storage paths. Engineers need an auditable, resource-aware correction to the issue. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Use supported structured output and schema validation with graceful retry or rejection for nonconforming responses.
- Adapt the training program to the managed data input contract or stage files to an explicit configurable path.
- Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators.
- Reduce the learning rate and use justified parameter-efficient tuning or partial freezing with a controlled evaluation set.
- Inspect task input configuration, container logs, expected file patterns and staged object paths.
Explanation
Correct Answers
B:Adapt the training program to the managed data input contract or stage files to an explicit configurable path. This is the targeted corrective response for custom training input schema mismatch. Notebook filesystem locations are not portable contracts for managed container training jobs. Verify the diagnosis using: Inspect task input configuration, container logs, expected file patterns and staged object paths. Confirm success using: A clean managed training run reads the intended versioned dataset without notebook-local path assumptions. This matters here because while preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs:….
E:Inspect task input configuration, container logs, expected file patterns and staged object paths. This inspects evidence relevant to custom training input schema mismatch. Notebook filesystem locations are not portable contracts for managed container training jobs. After identifying the cause, the matching correction is: Adapt the training program to the managed data input contract or stage files to an explicit configurable path. This matters here because while preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs:….
Incorrect Answers
A:Use supported structured output and schema validation with graceful retry or rejection for nonconforming responses. This is an action or diagnostic relevant to constrained generative json outputs, not one of the required two steps. In this scenario verify: Inspect task input configuration, container logs, expected file patterns and staged object paths. Then correct: Adapt the training program to the managed data input contract or stage files to an explicit configurable path.
C:Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators. This is an action or diagnostic relevant to distributed data parallel gradient synchronization, not one of the required two steps. In this scenario verify: Inspect task input configuration, container logs, expected file patterns and staged object paths. Then correct: Adapt the training program to the managed data input contract or stage files to an explicit configurable path.
D:Reduce the learning rate and use justified parameter-efficient tuning or partial freezing with a controlled evaluation set. This is an action or diagnostic relevant to fine-tuning catastrophic forgetting, not one of the required two steps. In this scenario verify: Inspect task input configuration, container logs, expected file patterns and staged object paths. Then correct: Adapt the training program to the managed data input contract or stage files to an explicit configurable path. This matters here because while preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs:….
Question 499
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A reporting team exports millions of BigQuery rows to a laptop for daily scoring although a trained BigQuery ML model already exists in the dataset. The owners require measurable quality and controlled operating cost. Which TWO choices together provide the decisive diagnostic evidence and the most direct acceptance test? Select TWO.
- Inspect the target column type, business loss metric, and sample prediction schema alongside model_type in the BigQuery model metadata.
- Stable clusters reveal coherent customer behavior without pretending an unsupervised assignment is known supervised ground truth.
- Classification F1 remains above the acceptance threshold while end-to-end latency and cost per thousand requests fall.
- Review row volume, model location, permissions and the current SQL report dependency on the local scoring script.
- The scheduled report processes the full authorized partition and stores traceable predictions without a local export.
Explanation
Correct Answers
D:Review row volume, model location, permissions and the current SQL report dependency on the local scoring script. This inspects evidence relevant to bigquery ml sql batch prediction integration. In-database prediction avoids unnecessary bulk data transfer when model and features are already in BigQuery. After identifying the cause, the matching correction is: Use ML.PREDICT on the appropriate BigQuery ML model within a controlled SQL batch job, writing results to governed tables.
E:The scheduled report processes the full authorized partition and stores traceable predictions without a local export. This is specific acceptance evidence for bigquery ml sql batch prediction integration. In-database prediction avoids unnecessary bulk data transfer when model and features are already in BigQuery. It confirms the controlled correction: Use ML.PREDICT on the appropriate BigQuery ML model within a controlled SQL batch job, writing results to governed tables.
Incorrect Answers
A:Inspect the target column type, business loss metric, and sample prediction schema alongside model_type in the BigQuery model metadata. This relates to bigquery ml linear regression for continuous value but does not establish the requested before-and-after evidence. The actual diagnostic is: Review row volume, model location, permissions and the current SQL report dependency on the local scoring script. The specific acceptance result is: The scheduled report processes the full authorized partition and stores traceable predictions without a local export.
B:Stable clusters reveal coherent customer behavior without pretending an unsupervised assignment is known supervised ground truth. This relates to bigquery ml kmeans customer segmentation but does not establish the requested before-and-after evidence. The actual diagnostic is: Review row volume, model location, permissions and the current SQL report dependency on the local scoring script. The specific acceptance result is: The scheduled report processes the full authorized partition and stores traceable predictions without a local export.
C:Classification F1 remains above the acceptance threshold while end-to-end latency and cost per thousand requests fall. This relates to gemini model choice for latency versus complexity but does not establish the requested before-and-after evidence. The actual diagnostic is: Review row volume, model location, permissions and the current SQL report dependency on the local scoring script. The specific acceptance result is: The scheduled report processes the full authorized partition and stores traceable predictions without a local export.
Question 500
While reproducing a training orchestration incident in isolated staging, the following condition occurs: An operations team retrains its classifier every hour based on a clock even when data have not changed, wasting compute and risking unreviewed model swaps. The team requires an idempotent pipeline behavior and traceable evidence. What acceptance result verifies the intended behavior after remediation?
- Representative raw records produce equal feature vectors offline and online, including unseen-category behavior.
- A deliberately incompatible payload prevents the release and a compatible fixture completes an end-to-end prediction.
- No-op periods create no unnecessary models and validated drift events produce candidates requiring approval.
- One completed batch starts one intended pipeline run and unrelated object uploads do not trigger training.
Explanation
Correct Answer
C:No-op periods create no unnecessary models and validated drift events produce candidates requiring approval. This is specific acceptance evidence for data drift triggered retraining policy. Retraining on a timer alone does not imply better models and may use incomplete or unrepresentative labels. It confirms the controlled correction: Trigger candidate retraining only when justified evidence and sufficient labeled data are available, with evaluation gates.
Incorrect Answers
A:Representative raw records produce equal feature vectors offline and online, including unseen-category behavior. This result could demonstrate success for training serving preprocessing artifact sharing but not for the stated data drift triggered retraining policy issue. The direct test is: No-op periods create no unnecessary models and validated drift events produce candidates requiring approval.
B:A deliberately incompatible payload prevents the release and a compatible fixture completes an end-to-end prediction. This result could demonstrate success for continuous delivery integration contract tests but not for the stated data drift triggered retraining policy issue. The direct test is: No-op periods create no unnecessary models and validated drift events produce candidates requiring approval.
D:One completed batch starts one intended pipeline run and unrelated object uploads do not trigger training. This result could demonstrate success for eventbridge equivalent event trigger on storage but not for the stated data drift triggered retraining policy issue. The direct test is: No-op periods create no unnecessary models and validated drift events produce candidates requiring approval.