AWS Certified AI Practitioner AIF-C01 Practice Test 28
Question 270
A feedback program collects human preference pairs for generated customer responses. A pilot uses this approach instead: Use unlabeled random characters as preference rankings. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Uninterpretable data cannot provide reliable preference signals
- Encryption protects data but does not teach response preferences
- Preference comparisons can help train models toward human-valued behavior
- Token volume does not express which answer humans prefer
Explanation
Correct Answer
A:This is the actual failure mode of the proposed approach: Use unlabeled random characters as preference rankings. Uninterpretable data cannot provide reliable preference signals
Incorrect Answers
B:This limitation concerns Replace preference review with only storage encryption; it does not diagnose the proposed Use unlabeled random characters as preference rankings approach.
C:Preference comparisons can help train models toward human-valued behavior This explains the recommended design, but it is not the requested diagnosis of Use unlabeled random characters as preference rankings.
D:This limitation concerns Assume a token-count report is an alignment label; it does not diagnose the proposed Use unlabeled random characters as preference rankings approach.
Question 271
A model fine-tuned on a narrow support sample produces poor answers for a neglected customer segment. A pilot uses this approach instead: Repeat the same narrow examples more often. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Removing evaluation hides generalization problems
- Coverage of real user scenarios reduces harmful specialization gaps
- Parameter count does not measure group performance
- More duplicates do not add missing segment coverage
Explanation
Correct Answer
D:This is the actual failure mode of the proposed approach: Repeat the same narrow examples more often. More duplicates do not add missing segment coverage
Incorrect Answers
A:This limitation concerns Remove the holdout set to make training accuracy look better; it does not diagnose the proposed Repeat the same narrow examples more often approach.
B:Coverage of real user scenarios reduces harmful specialization gaps This explains the recommended design, but it is not the requested diagnosis of Repeat the same narrow examples more often.
C:This limitation concerns Use only model size as a fairness measure; it does not diagnose the proposed Repeat the same narrow examples more often approach.
Question 272
A task is straightforward and the team wants the model to answer without supplying any examples. During a design review, someone proposes Use only an embedding vector with no instruction. Which technical objection is specifically relevant to this proposal?
- Evaluation metrics do not tell the model the current task
- Few-shot prompting includes demonstrations of the desired mapping
- A semantic vector is not a complete requested task
- A zero-shot prompt states the task without demonstrated examples
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Use only an embedding vector with no instruction. A semantic vector is not a complete requested task
Incorrect Answers
A:This limitation concerns Assume an unrelated model evaluation result is the prompt; it does not diagnose the proposed Use only an embedding vector with no instruction approach.
B:This limitation concerns Call the prompt few-shot despite containing no examples; it does not diagnose the proposed Use only an embedding vector with no instruction approach.
D:A zero-shot prompt states the task without demonstrated examples This explains the recommended design, but it is not the requested diagnosis of Use only an embedding vector with no instruction.
Question 273
An internal prompt includes confidential API keys that users might echo back. During a design review, someone proposes Raise temperature so the model forgets the secret. Which technical objection is specifically relevant to this proposal?
- Refusal behavior cannot replace credential rotation and least privilege
- Obscurity is not credential protection
- Sampling randomness does not revoke exposed credentials
- Prompts may be logged or exposed and should not contain embedded credentials
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Raise temperature so the model forgets the secret. Sampling randomness does not revoke exposed credentials
Incorrect Answers
A:This limitation concerns Assume model refusal alone makes a leaked key invalid; it does not diagnose the proposed Raise temperature so the model forgets the secret approach.
B:This limitation concerns Rely solely on hiding secret text in a longer prompt; it does not diagnose the proposed Raise temperature so the model forgets the secret approach.
D:Prompts may be logged or exposed and should not contain embedded credentials This explains the recommended design, but it is not the requested diagnosis of Raise temperature so the model forgets the secret.
Question 274
A healthcare team is curating fine-tuning examples and many contain sensitive patient identifiers. During a design review, someone proposes Treat the model's hidden parameters as proof no patient data can leak. Which technical objection is specifically relevant to this proposal?
- Unvetted sensitive training data creates privacy and compliance risks
- Encryption does rather than replace lawful data-use authorization
- Training data governance and privacy matter before model adaptation
- Weights do not by themselves guarantee training privacy
Explanation
Correct Answer
D:This is the actual failure mode of the proposed approach: Treat the model's hidden parameters as proof no patient data can leak. Weights do not by themselves guarantee training privacy
Incorrect Answers
A:This limitation concerns Use every raw patient record without consent assessment; it does not diagnose the proposed Treat the model's hidden parameters as proof no patient data can leak approach.
B:This limitation concerns Assume file encryption alone establishes the right to use all medical records; it does not diagnose the proposed Treat the model's hidden parameters as proof no patient data can leak approach.
C:Training data governance and privacy matter before model adaptation This explains the recommended design, but it is not the requested diagnosis of Treat the model's hidden parameters as proof no patient data can leak.
Question 275
An evaluation team wants scalable subjective feedback on FM answers before human sampling. A pilot uses this approach instead: Treat judge outputs as infallible ground truth. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Cost does not judge whether the response is correct
- Self-reported confidence may be poorly calibrated
- Automated judges can be biased and should be calibrated
- A judging model can scale rubric-based scoring but may have systematic bias
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Treat judge outputs as infallible ground truth. Automated judges can be biased and should be calibrated
Incorrect Answers
A:This limitation concerns Use the price per token as the sole quality score; it does not diagnose the proposed Treat judge outputs as infallible ground truth approach.
B:This limitation concerns Use only the model's confidence in its own answer; it does not diagnose the proposed Treat judge outputs as infallible ground truth approach.
D:A judging model can scale rubric-based scoring but may have systematic bias This explains the recommended design, but it is not the requested diagnosis of Treat judge outputs as infallible ground truth.
Question 276
An FM must output a compact JSON object with specified fields. After selecting the recommended approach, which evidence MOST DIRECTLY demonstrates that the actual required outcome has been achieved?
- A component test invokes this alternate approach successfully: Increase temperature to guarantee JSON validity
- A reviewed design record approves this competing approach: Use unrelated example documents without naming desired keys
- A system inventory lists this competing approach as active: Provide no format guidance and rely on natural prose
- Tests parse the generated response against the required schema
Explanation
Correct Answer
D:Tests parse the generated response against the required schema. This checks the business result rather than just a configured feature.
Incorrect Answers
A:Random sampling is not a schema validator A component invocation does not validate the necessary end-to-end outcome.
B:Irrelevant examples do not define the output contract Design approval is not evidence that the alternative meets the actual requirement.
C:Unguided output may omit required fields An active setting cannot independently prove the requested business outcome.
Question 277
A large FM is too costly to serve, but a smaller model could learn some of its task behavior. Which option BEST meets the business requirement?
- Expand the context window as the entire model compression method
- Evaluate model distillation from a stronger teacher into a smaller student
- Assume distillation means indexing source PDFs in a vector DB
- Replace the small student model with a larger untested teacher in production
Explanation
Correct Answer
B:Distillation can transfer useful behavior to a more efficient model The specified scenario makes this the most appropriate approach.
Incorrect Answers
A:More context does not reduce model parameters This prevents the approach from meeting the stated requirement.
C:Retrieval indexing is not teacher-student model training This prevents the approach from meeting the stated requirement.
D:A larger teacher alone does not solve serving-cost constraints This prevents the approach from meeting the stated requirement.
Question 278
A multi-step support agent appears fluent but regularly fails to complete the requested customer workflow. After selecting the recommended approach, which evidence MOST DIRECTLY demonstrates that the actual required outcome has been achieved?
- A component test invokes this alternate approach successfully: Measure retrieval relevance without tracking whether actions finished
- Traces confirm correct final state and acceptable cost per completed task
- A reviewed design record approves this competing approach: Treat configured tools as proof every action executed
- A system inventory lists this competing approach as active: Score only conversational word count
Explanation
Correct Answer
B:Traces confirm correct final state and acceptable cost per completed task. This checks the business result rather than just a configured feature.
Incorrect Answers
A:Retrieval quality alone is not end-to-end task success A component invocation does not validate the necessary end-to-end outcome.
C:Tool availability is not successful authorized execution Design approval is not evidence that the alternative meets the actual requirement.
D:Longer responses can mask unfinished work An active setting cannot independently prove the requested business outcome.
Question 279
A customer service bot needs accurate answers from an internal policy collection that changes weekly. Before approving a rollout, an architect suggests Rely only on public pretraining for private documents. Which mismatch should be documented for this precise design?
- Static weights may not reflect weekly document changes
- Pretrained data does not contain confidential current policies by default
- Randomness does not provide current policy evidence
- Knowledge Bases can retrieve relevant documents to ground generated replies
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Rely only on public pretraining for private documents. Pretrained data does not contain confidential current policies by default
Incorrect Answers
A:This limitation concerns Fine-tune on one static snapshot as the only update mechanism; it does not diagnose the proposed Rely only on public pretraining for private documents approach.
C:This limitation concerns Increase temperature until the model recalls current internal clauses; it does not diagnose the proposed Rely only on public pretraining for private documents approach.
D:Knowledge Bases can retrieve relevant documents to ground generated replies This explains the recommended design, but it is not the requested diagnosis of Rely only on public pretraining for private documents.
Question 280
An assistant has a strict 300 millisecond response budget and a fixed monthly inference allowance. A pilot uses this approach instead: Assume largest available model is always most responsive. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Larger parameter count does not guarantee lower latency
- A model that meets the task with less compute can be preferable
- Training corpus size does not establish served latency
- Cheap incorrect answers fail the application requirement
Explanation
Correct Answer
A:This is the actual failure mode of the proposed approach: Assume largest available model is always most responsive. Larger parameter count does not guarantee lower latency
Incorrect Answers
B:A model that meets the task with less compute can be preferable This explains the recommended design, but it is not the requested diagnosis of Assume largest available model is always most responsive.
C:This limitation concerns Choose solely by advertised training dataset size; it does not diagnose the proposed Assume largest available model is always most responsive approach.
D:This limitation concerns Select by price without testing task accuracy; it does not diagnose the proposed Assume largest available model is always most responsive approach.
