Google Professional Machine Learning Engineer Practice Test 13
Question 120
During model-training validation on a Google Cloud experimental project, the following condition occurs: An existing TensorFlow training pipeline has large dense matrix operations and engineers select a TPU solely from marketing without checking model operation support. The change should preserve a meaningful independent model-quality comparison. Which outcome most directly demonstrates a successful, measurable fix?
- The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
- Scale efficiency improves and model quality remains stable without synchronization stalls dominating every step.
- GPU utilization and examples per second rise without changing model accuracy or violating data-order requirements.
- The training manifest contains traceable permitted sources and safety tests detect known prohibited outputs.
Explanation
Correct Answer
A:The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing. This is specific acceptance evidence for tpu versus gpu for tensorflow workload. TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. It confirms the controlled correction: Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
Incorrect Answers
B:Scale efficiency improves and model quality remains stable without synchronization stalls dominating every step. This result could demonstrate success for distributed gradient communication bottleneck but not for the stated tpu versus gpu for tensorflow workload issue. The direct test is: The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
C:GPU utilization and examples per second rise without changing model accuracy or violating data-order requirements. This result could demonstrate success for accelerator starvation from slow input pipeline but not for the stated tpu versus gpu for tensorflow workload issue. The direct test is: The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
D: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 tpu versus gpu for tensorflow workload issue. The direct test is: The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
Question 121
At a Google Cloud analytics team preparing a governed pilot, the following condition occurs: An international retailer manually sends each product description to an English-only classifier although the business first needs translated content in several supported languages. The team must satisfy a reproducible low-code AI assessment. What should the engineer inspect before selecting a correction?
- Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample.
- Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints.
- Inspect the training preprocessing SQL, TRANSFORM clause, model input schema, and the inference query's feature values.
- Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task.
Explanation
Correct Answer
A:Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample. This inspects evidence relevant to translate api for multilingual product descriptions. Translation and classification are different tasks; an industry language API directly addresses supported cross-language conversion. After identifying the cause, the matching correction is: Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content.
Incorrect Answers
B:Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints. This examines gemini model choice for latency versus complexity rather than the specific mechanism reported here. It is not the decisive first check for translate api for multilingual product descriptions. The relevant evidence is: Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample.
C:Inspect the training preprocessing SQL, TRANSFORM clause, model input schema, and the inference query's feature values. This examines bigquery ml transform serving parity rather than the specific mechanism reported here. It is not the decisive first check for translate api for multilingual product descriptions. The relevant evidence is: Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample.
D:Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task. This examines model garden task and license selection rather than the specific mechanism reported here. It is not the decisive first check for translate api for multilingual product descriptions. The relevant evidence is: Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample.
Question 122
During a staged restoration test at a secondary serving deployment, the following condition occurs: A text classifier was trained with a specific subword vocabulary, but production preprocessors use a newer incompatible tokenizer. The team must prove an actual user-visible improvement without weakening access controls. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Inspect quota errors, retry behavior, queue occupancy, tail latency and autoscaling upper limits.
- Examine endpoint exposure settings, caller IAM, VPC routes, private connectivity and data egress requirements.
- Review payload size distributions, request duration, response timeout policies and tolerance for deferred completion.
- Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts.
Explanation
Correct Answer
D:Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts. This inspects evidence relevant to inference preprocessing tokenization mismatch. Embedding matrices interpret token IDs according to one vocabulary; mismatched preprocessing changes input meaning. After identifying the cause, the matching correction is: Package and version the exact trained preprocessing pipeline with the served model and run equivalence tests.
Incorrect Answers
A:Inspect quota errors, retry behavior, queue occupancy, tail latency and autoscaling upper limits. This examines predictive serving rate limit and backpressure rather than the specific mechanism reported here. It is not the decisive first check for inference preprocessing tokenization mismatch. The relevant evidence is: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts.
B:Examine endpoint exposure settings, caller IAM, VPC routes, private connectivity and data egress requirements. This examines private endpoint for regulated inference traffic rather than the specific mechanism reported here. It is not the decisive first check for inference preprocessing tokenization mismatch. The relevant evidence is: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts.
C:Review payload size distributions, request duration, response timeout policies and tolerance for deferred completion. This examines online prediction timeout and batch fallback rather than the specific mechanism reported here. It is not the decisive first check for inference preprocessing tokenization mismatch. The relevant evidence is: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts.
Question 123
During a staged restoration test at a secondary serving deployment, the following condition occurs: A real-time scoring service reads a managed feature view that has not synchronized since yesterday while the source BigQuery table contains recent transaction updates. The team must prove an actual user-visible improvement without weakening access controls. Which statement identifies the primary reason for this failure?
- A cloud-only inference architecture cannot satisfy hard offline and low-latency edge requirements.
- Endpoint invoker permissions should follow least privilege, rather than expanding access for convenience.
- Online feature serving requires an up-to-date synchronized view; source-table changes need not appear instantly.
- Identity permission is not proof of network reachability; private serving also requires correct routing and name resolution.
Explanation
Correct Answer
C:Online feature serving requires an up-to-date synchronized view; source-table changes need not appear instantly. This explains the mechanism behind feature store freshness and online sync and why the described failure is possible. A discriminating investigation is: Inspect feature-view sync status, latest source timestamps, entity ID and online lookup freshness. The corresponding repair is: Schedule and monitor approved Feature Store synchronization at the needed frequency, with stale-feature handling.
Incorrect Answers
A:A cloud-only inference architecture cannot satisfy hard offline and low-latency edge requirements. This principle describes edge versus cloud inference network constraints, which is not the cause of the observed feature store freshness and online sync failure. The controlling explanation is: Online feature serving requires an up-to-date synchronized view; source-table changes need not appear instantly.
B:Endpoint invoker permissions should follow least privilege, rather than expanding access for convenience. This principle describes public endpoint iam oversharing, which is not the cause of the observed feature store freshness and online sync failure. The controlling explanation is: Online feature serving requires an up-to-date synchronized view; source-table changes need not appear instantly.
D:Identity permission is not proof of network reachability; private serving also requires correct routing and name resolution. This principle describes private vpc routing failure for serving, which is not the cause of the observed feature store freshness and online sync failure. The controlling explanation is: Online feature serving requires an up-to-date synchronized view; source-table changes need not appear instantly.
Question 124
During model-training validation on a Google Cloud experimental project, the following condition occurs: An existing TensorFlow training pipeline has large dense matrix operations and engineers select a TPU solely from marketing without checking model operation support. The change should preserve a meaningful independent model-quality comparison. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect parameter and activation memory per device, optimizer memory and available interconnect bandwidth.
- Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost.
- Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events.
- Inspect task requirements for retrieval ranking, available query-document pairs, embedding similarity and labeled relevance examples.
Explanation
Correct Answer
B:Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. This inspects evidence relevant to tpu versus gpu for tensorflow workload. TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. After identifying the cause, the matching correction is: Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
Incorrect Answers
A:Inspect parameter and activation memory per device, optimizer memory and available interconnect bandwidth. This examines model parallelism for memory-bound transformer rather than the specific mechanism reported here. It is not the decisive first check for tpu versus gpu for tensorflow workload. The relevant evidence is: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost.
C:Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events. This examines accelerator quota and regional availability rather than the specific mechanism reported here. It is not the decisive first check for tpu versus gpu for tensorflow workload. The relevant evidence is: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost.
D:Inspect task requirements for retrieval ranking, available query-document pairs, embedding similarity and labeled relevance examples. This examines embedding retrieval versus text classifier rather than the specific mechanism reported here. It is not the decisive first check for tpu versus gpu for tensorflow workload. The relevant evidence is: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost.
Question 125
During a production ML inference release on Google Cloud, the following condition occurs: A text classifier was trained with a specific subword vocabulary, but production preprocessors use a newer incompatible tokenizer. The service must meet its agreed quality and latency requirements. Which diagnostic check should be performed FIRST to isolate the described cause?
- Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts.
- Inspect model weights, KV or activation memory, concurrency load, GPU memory, and supported device types.
- Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model.
- Review payload size distributions, request duration, response timeout policies and tolerance for deferred completion.
Explanation
Correct Answer
A:Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts. This inspects evidence relevant to inference preprocessing tokenization mismatch. Embedding matrices interpret token IDs according to one vocabulary; mismatched preprocessing changes input meaning. After identifying the cause, the matching correction is: Package and version the exact trained preprocessing pipeline with the served model and run equivalence tests. This matters here because a text classifier was trained with a specific subword vocabulary, but production preprocessors use a newer….
Incorrect Answers
B:Inspect model weights, KV or activation memory, concurrency load, GPU memory, and supported device types. This examines gpu cost and memory fit for large embedding model rather than the specific mechanism reported here. It is not the decisive first check for inference preprocessing tokenization mismatch. The relevant evidence is: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts.
C:Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. This examines model traffic rollback on high error rate rather than the specific mechanism reported here. It is not the decisive first check for inference preprocessing tokenization mismatch. The relevant evidence is: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts.
D:Review payload size distributions, request duration, response timeout policies and tolerance for deferred completion. This examines online prediction timeout and batch fallback rather than the specific mechanism reported here. It is not the decisive first check for inference preprocessing tokenization mismatch. The relevant evidence is: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts. This matters here because a text classifier was trained with a specific subword vocabulary, but production preprocessors use a newer….
Question 126
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A service account running an AI workload can read all analytics datasets and copy confidential training data to an unrelated storage bucket. The fix must retain legitimate model usefulness and appropriate controls. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema.
- Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers.
- Inspect subgroup confusion matrices, sample sizes, training representation and evaluation slices over time.
- Inspect attachment ingestion, trust boundaries, malicious prompt examples and deployed Model Armor filter results.
Explanation
Correct Answer
B:Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers. This inspects evidence relevant to training data exfiltration through broad access. Least-privilege access and destination controls reduce unauthorized model-data disclosure risk. After identifying the cause, the matching correction is: Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads.
Incorrect Answers
A:Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema. This examines ground truth association for quality monitoring rather than the specific mechanism reported here. It is not the decisive first check for training data exfiltration through broad access. The relevant evidence is: Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers.
C:Inspect subgroup confusion matrices, sample sizes, training representation and evaluation slices over time. This examines model feature bias by subgroup rather than the specific mechanism reported here. It is not the decisive first check for training data exfiltration through broad access. The relevant evidence is: Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers.
D: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 training data exfiltration through broad access. The relevant evidence is: Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers.
Question 127
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: An imbalanced defect dataset is oversampled before train/test splitting, so near-duplicate synthetic records appear in both partitions. Engineers need an auditable, resource-aware correction to the issue. What should the engineer inspect before selecting a correction?
- Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss.
- Inspect sampling order, example hashes and class distributions in training and held-out folds.
- Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators.
- Determine whether failures arise from changing reference knowledge, stable behavior patterns or inadequate retrieval coverage.
Explanation
Correct Answer
B:Inspect sampling order, example hashes and class distributions in training and held-out folds. This inspects evidence relevant to training class imbalance fold leakage. Resampling before splitting leaks information and makes minority-class performance appear artificially high. After identifying the cause, the matching correction is: Split raw records first and resample only inside training folds, leaving validation cohorts untouched.
Incorrect Answers
A:Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss. This examines early stopping on validation divergence rather than the specific mechanism reported here. It is not the decisive first check for training class imbalance fold leakage. The relevant evidence is: Inspect sampling order, example hashes and class distributions in training and held-out folds.
C:Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators. This examines distributed data parallel gradient synchronization rather than the specific mechanism reported here. It is not the decisive first check for training class imbalance fold leakage. The relevant evidence is: Inspect sampling order, example hashes and class distributions in training and held-out folds.
D:Determine whether failures arise from changing reference knowledge, stable behavior patterns or inadequate retrieval coverage. This examines foundation model prompt versus fine-tuning rather than the specific mechanism reported here. It is not the decisive first check for training class imbalance fold leakage. The relevant evidence is: Inspect sampling order, example hashes and class distributions in training and held-out folds.
Question 128
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: Two experiments use the same model configuration but show different accuracy due to untracked data shuffling and nondeterministic seeds. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which technical explanation best accounts for the reported behavior?
- Trial metrics and artifact provenance must be linked to support valid experiment selection and release audits.
- LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration.
- Single-machine notebook memory is limited; warehouse operations should execute near large warehouse-resident data.
- A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation.
Explanation
Correct Answer
D:A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation. This explains the mechanism behind experiment random seed and data order variance and why the described failure is possible. A discriminating investigation is: Inspect random seeds, sample ordering, framework determinism settings and environment versions recorded per run. The corresponding repair is: Log reproducibility settings and rerun comparable trials with fixed split and controlled randomness before attributing changes.
Incorrect Answers
A:Trial metrics and artifact provenance must be linked to support valid experiment selection and release audits. This principle describes pipeline experiment versus notebook-only tracking, which is not the cause of the observed experiment random seed and data order variance failure. The controlling explanation is: A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation.
B:LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration. This principle describes llm-as-a-judge rubric reliability, which is not the cause of the observed experiment random seed and data order variance failure. The controlling explanation is: A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation.
C:Single-machine notebook memory is limited; warehouse operations should execute near large warehouse-resident data. This principle describes notebook memory exhaustion on full data extraction, which is not the cause of the observed experiment random seed and data order variance failure. The controlling explanation is: A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation.
Question 129
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: A neural network's training loss continues falling for fifty epochs while validation loss starts rising after epoch ten. Engineers need an auditable, resource-aware correction to the issue. Which TWO complementary choices jointly establish the evidence and appropriate correction? Select TWO.
- Use suitable text embeddings with vector similarity search and reranking rather than a binary spam decision.
- Adapt the training program to the managed data input contract or stage files to an explicit configurable path.
- Configure early stopping with a justified patience window and restore the best validated checkpoint.
- Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss.
- Review parameter ranges, total trial count, training budget and the metric used to rank completed experiments.
Explanation
Correct Answers
C:Configure early stopping with a justified patience window and restore the best validated checkpoint. This is the targeted corrective response for early stopping on validation divergence. More training steps can increase overfitting even when training-set loss keeps improving. Verify the diagnosis using: Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss. Confirm success using: The deployed artifact corresponds to the best held-out epoch rather than the final overfit epoch. This matters here because while preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs:….
D:Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss. This inspects evidence relevant to early stopping on validation divergence. More training steps can increase overfitting even when training-set loss keeps improving. After identifying the cause, the matching correction is: Configure early stopping with a justified patience window and restore the best validated checkpoint. This matters here because while preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs:….
Incorrect Answers
A:Use suitable text embeddings with vector similarity search and reranking rather than a binary spam decision. This is an action or diagnostic relevant to embedding retrieval versus text classifier, not one of the required two steps. In this scenario verify: Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss. Then correct: Configure early stopping with a justified patience window and restore the best validated checkpoint.
B:Adapt the training program to the managed data input contract or stage files to an explicit configurable path. This is an action or diagnostic relevant to custom training input schema mismatch, not one of the required two steps. In this scenario verify: Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss. Then correct: Configure early stopping with a justified patience window and restore the best validated checkpoint.
E:Review parameter ranges, total trial count, training budget and the metric used to rank completed experiments. This is an action or diagnostic relevant to hyperparameter search budget limit, not one of the required two steps. In this scenario verify: Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss. Then correct: Configure early stopping with a justified patience window and restore the best validated checkpoint.
Question 130
During a staged restoration test at a secondary serving deployment, the following condition occurs: A team deploys a model using a mutable storage path named production-latest and cannot reconstruct which evaluation approved its weights. The team must prove an actual user-visible improvement without weakening access controls. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles.
- Review immutable artifact IDs, Model Registry versions, evaluation reports, approvals and deployed revision metadata.
- Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model.
- Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests.
Explanation
Correct Answer
B:Review immutable artifact IDs, Model Registry versions, evaluation reports, approvals and deployed revision metadata. This inspects evidence relevant to model registry version provenance. Mutable nicknames do not establish immutable model provenance for controlled production releases. After identifying the cause, the matching correction is: Promote an explicitly approved model version from Agent Platform Model Registry and record deployment lineage.
Incorrect Answers
A:Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. This examines dynamic batch serving for gpu saturation rather than the specific mechanism reported here. It is not the decisive first check for model registry version provenance. The relevant evidence is: Review immutable artifact IDs, Model Registry versions, evaluation reports, approvals and deployed revision metadata.
C:Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. This examines model traffic rollback on high error rate rather than the specific mechanism reported here. It is not the decisive first check for model registry version provenance. The relevant evidence is: Review immutable artifact IDs, Model Registry versions, evaluation reports, approvals and deployed revision metadata.
D:Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. This examines model artifact image digest reproducibility rather than the specific mechanism reported here. It is not the decisive first check for model registry version provenance. The relevant evidence is: Review immutable artifact IDs, Model Registry versions, evaluation reports, approvals and deployed revision metadata.