100% Latest & Updated Google Professional Machine Learning Engineer Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 790
A deployment has been completed and needs an acceptance test: Record source revision, dataset version, reviewer approval, immutable model ID and deployment event. This intervention addresses: A production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment. Which post-change check most directly proves the intended result?
- Verify alarm conditions over rolling windows and held-out candidate performance.
- Reconstruct the current deployment from audit events and original artifact digests.
- Submit old and new schema samples and assert intended acceptance/rejection.
- Re-run old failure cases and inspect per-language error metrics.
Explanation
Correct Answer
B:This is the appropriate post-change test: Reconstruct the current deployment from audit events and original artifact digests. The change under test is: Record source revision, dataset version, reviewer approval, immutable model ID and deployment event. The original condition was: A production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment. The expected behavior is: An incident investigation can attribute a release to the approved change chain.
Incorrect Answers
A:This check relates to another system: A deployed classifier's precision declines after labels arrive two weeks later while feature distributions appear mostly unchanged. It would not verify whether the current problem was fixed: A production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment. The appropriate acceptance test is: Reconstruct the current deployment from audit events and original artifact digests.
C:This check relates to another system: An upstream producer changes product category encoding and a model retrain uses the new schema but the current inference API still sends the old fields. It would not verify whether the current problem was fixed: A production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment. The appropriate acceptance test is: Reconstruct the current deployment from audit events and original artifact digests.
D:This check relates to another system: A model version appears accurate on a tiny curated test set but fails for common multilingual inputs after an update. It would not verify whether the current problem was fixed: A production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment. The appropriate acceptance test is: Reconstruct the current deployment from audit events and original artifact digests.
Question 791
An ML platform team is resolving 2 independent workstreams. Workstream A: A helpdesk bot retrieves a policy document that contains instructions to ignore previous prompts and email internal API keys to an attacker. Workstream B: A newly released model lacks documentation of training data limitations, fairness checks and approved decision boundaries. Select exactly TWO actions, one directly addressing each separate workstream.
- Publish a versioned model card with intended use, known limitations and measured risks.
- Configure appropriate supported safety filters and an escalation workflow.
- Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review.
- Treat retrieved text as untrusted data and constrain tool permissions, with Model Armor or equivalent supported defenses.
- Minimize or redact sensitive fields before invoking an appropriately governed model.
Explanation
Correct Answers
A:Publish a versioned model card with intended use, known limitations and measured risks. Confirm model-card metrics match the frozen release evaluation artifacts. A model card does not remediate a failed or biased model by itself.
D:Treat retrieved text as untrusted data and constrain tool permissions, with Model Armor or equivalent supported defenses. Test adversarial retrieved passages and confirm no secret or tool call is exposed. A single content filter cannot eliminate every indirect prompt-injection technique.
Incorrect Answers
B:Configuring appropriate supported safety filters and an escalation workflow is relevant when a youth education assistant occasionally produces violent or disallowed instructions despite a general polite system prompt; it does not establish the result required by this workstream.
C:Evaluating adversarial and stress-test input suites and apply robust preprocessing or fallback review is relevant when a document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload; it does not establish the result required by this workstream.
E:The approach “Minimize or redact sensitive fields before invoking an appropriately governed model” is relevant when a hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories; it does not establish the result required by this workstream.
Question 792
An ML engineering team must resolve this production problem: A hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories. Which action best addresses the actual technical requirement?
- Minimize or redact sensitive fields before invoking an appropriately governed model.
- Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review.
- Approve the selected model only after legal and security review of actual artifact terms.
- Publish a versioned model card with intended use, known limitations and measured risks.
Explanation
Correct Answer
A:Minimize or redact sensitive fields before invoking an appropriately governed model. Run protected PII detection tests and check audit records for redacted payloads. Removing names alone may leave reidentifying rare combinations.
Incorrect Answers
B:That measure applies when a document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload. The required intervention here is different: Minimize or redact sensitive fields before invoking an appropriately governed model. The need is a hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories.
C:That measure applies when a developer selects an open-weight foundation model from Model Garden and assumes every deployment mode permits commercial redistribution. The required intervention here is different: Minimize or redact sensitive fields before invoking an appropriately governed model. The need is a hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories.
D:That measure applies when a newly released model lacks documentation of training data limitations, fairness checks and approved decision boundaries. The required intervention here is different: Minimize or redact sensitive fields before invoking an appropriately governed model. The need is a hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories.
Question 793
An ML platform team is resolving 2 independent workstreams. Workstream A: A lending model's overall false-negative rate remains stable but denied applications from one monitored group increase sharply. Workstream B: A public GenAI endpoint receives malicious prompts asking it to reveal internal policies or run unauthorized tool calls. Select exactly TWO actions, one directly addressing each separate workstream.
- Use validated model attribution as a descriptive explanation and avoid unsupported causal claims.
- Apply approved Model Armor or supported safety protections alongside least-privilege tool design.
- Monitor and investigate group-conditioned error and approval metrics with responsible review.
- Treat retrieved text as untrusted data and constrain tool permissions, with Model Armor or equivalent supported defenses.
- Use synthetic or appropriately approved data and an authorized governed test environment.
Explanation
Correct Answers
B:Apply approved Model Armor or supported safety protections alongside least-privilege tool design. Execute injection tests and confirm unauthorized actions remain blocked. A security model is no replacement for permission boundaries and output verification.
C:Monitor and investigate group-conditioned error and approval metrics with responsible review. Measure subgroup confidence bounds and document reviewed mitigation decisions. Statistical differences require context, sample sizes and legal interpretation.
Incorrect Answers
A:Using validated model attribution as a descriptive explanation and avoid unsupported causal claims is relevant when a credit score's explanation shows one income-related field as influential and executives claim that changing that field would causally change approval; it does not establish the result required by this workstream.
D:The approach “Treat retrieved text as untrusted data and constrain tool permissions, with Model Armor or equivalent supported defenses” is relevant when a helpdesk bot retrieves a policy document that contains instructions to ignore previous prompts and email internal API keys to an attacker; it does not establish the result required by this workstream.
E:Using synthetic or appropriately approved data and an authorized governed test environment is relevant when an analyst pastes private customer complaints into a public playground for a quick prompt test without company authorization; it does not establish the result required by this workstream.
Question 794
A Google Cloud change request describes the following issue: A credit score's explanation shows one income-related field as influential and executives claim that changing that field would causally change approval. Which action best addresses the actual technical requirement?
- Configure appropriate supported safety filters and an escalation workflow.
- Minimize or redact sensitive fields before invoking an appropriately governed model.
- Use validated model attribution as a descriptive explanation and avoid unsupported causal claims.
- Monitor and investigate group-conditioned error and approval metrics with responsible review.
Explanation
Correct Answer
C:Use validated model attribution as a descriptive explanation and avoid unsupported causal claims. Compare local attribution stability under near-identical legitimate examples. Correlated features can redistribute attribution and mislead local narratives.
Incorrect Answers
A:That measure applies when a youth education assistant occasionally produces violent or disallowed instructions despite a general polite system prompt. The required intervention here is different: Use validated model attribution as a descriptive explanation and avoid unsupported causal claims. The need is a credit score's explanation shows one income-related field as influential and executives claim that changing that field would causally change approval.
B:That measure applies when a hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories. The required intervention here is different: Use validated model attribution as a descriptive explanation and avoid unsupported causal claims. The need is a credit score's explanation shows one income-related field as influential and executives claim that changing that field would causally change approval.
D:That measure applies when a lending model's overall false-negative rate remains stable but denied applications from one monitored group increase sharply. The required intervention here is different: Use validated model attribution as a descriptive explanation and avoid unsupported causal claims. The need is a credit score's explanation shows one income-related field as influential and executives claim that changing that field would causally change approval.
Question 795
An ML platform team is resolving 2 independent workstreams. Workstream A: A newly released model lacks documentation of training data limitations, fairness checks and approved decision boundaries. Workstream B: A public model endpoint accepts unlimited probing requests and logs reveal a client scraping large numbers of generated examples. Select exactly TWO actions, one directly addressing each separate workstream.
- Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review.
- Minimize or redact sensitive fields before invoking an appropriately governed model.
- Approve the selected model only after legal and security review of actual artifact terms.
- Enforce authenticated quotas and abuse monitoring with scoped content controls.
- Publish a versioned model card with intended use, known limitations and measured risks.
Explanation
Correct Answers
D:Enforce authenticated quotas and abuse monitoring with scoped content controls. Replay throttled request patterns and confirm enforcement and audit events. Rate limits alone cannot prevent every low-rate or distributed extraction attack.
E:Publish a versioned model card with intended use, known limitations and measured risks. Confirm model-card metrics match the frozen release evaluation artifacts. A model card does not remediate a failed or biased model by itself.
Incorrect Answers
A:Evaluating adversarial and stress-test input suites and apply robust preprocessing or fallback review is relevant when a document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload; it does not establish the result required by this workstream.
B:The approach “Minimize or redact sensitive fields before invoking an appropriately governed model” is relevant when a hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories; it does not establish the result required by this workstream.
C:The approach “Approve the selected model only after legal and security review of actual artifact terms” is relevant when a developer selects an open-weight foundation model from Model Garden and assumes every deployment mode permits commercial redistribution; it does not establish the result required by this workstream.
Question 796
An engineer is choosing an appropriate design for this workload: A document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload. Which action best addresses the actual technical requirement?
- Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review.
- Enforce authenticated quotas and abuse monitoring with scoped content controls.
- Minimize or redact sensitive fields before invoking an appropriately governed model.
- Publish a versioned model card with intended use, known limitations and measured risks.
Explanation
Correct Answer
A:Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review. Measure error and abstention rates under fixed corruption severities. Tests cannot certify robustness against all future adversarial examples.
Incorrect Answers
B:That measure applies when a public model endpoint accepts unlimited probing requests and logs reveal a client scraping large numbers of generated examples. The required intervention here is different: Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review. The need is a document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload.
C:That measure applies when a hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories. The required intervention here is different: Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review. The need is a document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload.
D:That measure applies when a newly released model lacks documentation of training data limitations, fairness checks and approved decision boundaries. The required intervention here is different: Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review. The need is a document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload.
Question 797
An ML platform team is resolving 2 independent workstreams. Workstream A: A public GenAI endpoint receives malicious prompts asking it to reveal internal policies or run unauthorized tool calls. Workstream B: An analyst pastes private customer complaints into a public playground for a quick prompt test without company authorization. Select exactly TWO actions, one directly addressing each separate workstream.
- Apply approved Model Armor or supported safety protections alongside least-privilege tool design.
- Treat retrieved text as untrusted data and constrain tool permissions, with Model Armor or equivalent supported defenses.
- Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review.
- Use synthetic or appropriately approved data and an authorized governed test environment.
- Monitor and investigate group-conditioned error and approval metrics with responsible review.
Explanation
Correct Answers
A:Apply approved Model Armor or supported safety protections alongside least-privilege tool design. Execute injection tests and confirm unauthorized actions remain blocked. A security model is no replacement for permission boundaries and output verification.
D:Use synthetic or appropriately approved data and an authorized governed test environment. Review model invocation logs for restricted strings and ensure access is limited. A confidentiality sentence in a prompt cannot override actual provider data settings.
Incorrect Answers
B:The approach “Treat retrieved text as untrusted data and constrain tool permissions, with Model Armor or equivalent supported defenses” is relevant when a helpdesk bot retrieves a policy document that contains instructions to ignore previous prompts and email internal API keys to an attacker; it does not establish the result required by this workstream.
C:Evaluating adversarial and stress-test input suites and apply robust preprocessing or fallback review is relevant when a document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload; it does not establish the result required by this workstream.
E:Monitoring and investigate group-conditioned error and approval metrics with responsible review is relevant when a lending model's overall false-negative rate remains stable but denied applications from one monitored group increase sharply; it does not establish the result required by this workstream.
Question 798
An ML engineering team must resolve this production problem: A youth education assistant occasionally produces violent or disallowed instructions despite a general polite system prompt. Which action best addresses the actual technical requirement?
- Treat retrieved text as untrusted data and constrain tool permissions, with Model Armor or equivalent supported defenses.
- Configure appropriate supported safety filters and an escalation workflow.
- Use validated model attribution as a descriptive explanation and avoid unsupported causal claims.
- Publish a versioned model card with intended use, known limitations and measured risks.
Explanation
Correct Answer
B:Configure appropriate supported safety filters and an escalation workflow. Track unsafe outputs and false-positive blocks on reviewed prompts. Aggressive filters can reject harmless educational requests.
Incorrect Answers
A:That measure applies when a helpdesk bot retrieves a policy document that contains instructions to ignore previous prompts and email internal API keys to an attacker. The required intervention here is different: Configure appropriate supported safety filters and an escalation workflow. The need is a youth education assistant occasionally produces violent or disallowed instructions despite a general polite system prompt.
C:That measure applies when a credit score's explanation shows one income-related field as influential and executives claim that changing that field would causally change approval. The required intervention here is different: Configure appropriate supported safety filters and an escalation workflow. The need is a youth education assistant occasionally produces violent or disallowed instructions despite a general polite system prompt.
D:That measure applies when a newly released model lacks documentation of training data limitations, fairness checks and approved decision boundaries. The required intervention here is different: Configure appropriate supported safety filters and an escalation workflow. The need is a youth education assistant occasionally produces violent or disallowed instructions despite a general polite system prompt.
Question 799
An ML platform team is resolving 2 independent workstreams. Workstream A: A public model endpoint accepts unlimited probing requests and logs reveal a client scraping large numbers of generated examples. Workstream B: A hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories. Select exactly TWO actions, one directly addressing each separate workstream.
- Apply approved Model Armor or supported safety protections alongside least-privilege tool design.
- Publish a versioned model card with intended use, known limitations and measured risks.
- Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review.
- Minimize or redact sensitive fields before invoking an appropriately governed model.
- Enforce authenticated quotas and abuse monitoring with scoped content controls.
Explanation
Correct Answers
D:Minimize or redact sensitive fields before invoking an appropriately governed model. Run protected PII detection tests and check audit records for redacted payloads. Removing names alone may leave reidentifying rare combinations.
E:Enforce authenticated quotas and abuse monitoring with scoped content controls. Replay throttled request patterns and confirm enforcement and audit events. Rate limits alone cannot prevent every low-rate or distributed extraction attack.
Incorrect Answers
A:Applying approved Model Armor or supported safety protections alongside least-privilege tool design is relevant when a public GenAI endpoint receives malicious prompts asking it to reveal internal policies or run unauthorized tool calls; it does not establish the result required by this workstream.
B:The approach “Publish a versioned model card with intended use, known limitations and measured risks” is relevant when a newly released model lacks documentation of training data limitations, fairness checks and approved decision boundaries; it does not establish the result required by this workstream.
C:Evaluating adversarial and stress-test input suites and apply robust preprocessing or fallback review is relevant when a document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload; it does not establish the result required by this workstream.
Question 800
A Google Cloud change request describes the following issue: A developer selects an open-weight foundation model from Model Garden and assumes every deployment mode permits commercial redistribution. Which action best addresses the actual technical requirement?
- Approve the selected model only after legal and security review of actual artifact terms.
- Evaluate adversarial and stress-test input suites and apply robust preprocessing or fallback review.
- Minimize or redact sensitive fields before invoking an appropriately governed model.
- Monitor and investigate group-conditioned error and approval metrics with responsible review.
Explanation
Correct Answer
A:Approve the selected model only after legal and security review of actual artifact terms. Record license approval and audit model artifact hashes in the release manifest. A catalog listing is not a blanket grant for derivative commercial use.
Incorrect Answers
B:That measure applies when a document classifier gives confident incorrect outputs for slightly modified or visually corrupted forms that users can upload. The required intervention here is different: Approve the selected model only after legal and security review of actual artifact terms. The need is a developer selects an open-weight foundation model from Model Garden and assumes every deployment mode permits commercial redistribution.
C:That measure applies when a hospital chatbot sends patient names and visit identifiers to a general assistant although its task only needs deidentified diagnostic categories. The required intervention here is different: Approve the selected model only after legal and security review of actual artifact terms. The need is a developer selects an open-weight foundation model from Model Garden and assumes every deployment mode permits commercial redistribution.
D:That measure applies when a lending model's overall false-negative rate remains stable but denied applications from one monitored group increase sharply. The required intervention here is different: Approve the selected model only after legal and security review of actual artifact terms. The need is a developer selects an open-weight foundation model from Model Garden and assumes every deployment mode permits commercial redistribution.
Professional Machine Learning Engineer Practice Tests