Amazon AWS Certified AI Practitioner AIF-C01 Practice Test 33
Question 320
The initial scenario was: An AI team wants a foundation model to learn specialized vocabulary from a large corpus beyond simple response examples. The original control is already in place. Now domain facts begin changing hourly. Which ADDITIONAL action BEST addresses the new condition?
- Controlled evaluations compare domain understanding and task quality after adaptation
- Assume changing one prompt permanently updates every model weight
- Prefer live retrieval for freshness-sensitive facts
- Compare continuous pretraining on domain text with instruction tuning and retrieval
Explanation
Correct Answer
C:Domain facts begin changing hourly. The appropriate added action is: Prefer live retrieval for freshness-sensitive facts.
Incorrect Answers
A:This describes evidence for the original approach, not a new action: Controlled evaluations compare domain understanding and task quality after adaptation.
B:This original shortcut remains unsuitable: Inference prompts do not by themselves retrain model parameters.
D:This was the original control and is already implemented. Continued pretraining adapts knowledge representations while instruction tuning targets following examples.
Question 321
A model's answers are improved by collecting ranked human preferences between alternative safe responses. Which option BEST meets the stated AI/ML or AWS requirement?
- Add subgroup fairness evaluation and human auditing
- Assume every human preference is a verified factual answer key
- Build a representative curated training and validation dataset
- Describe reinforcement learning from human feedback as preference-guided alignment
Explanation
Correct Answer
D:Human preference signals can shape model behavior beyond next-token pretraining. In this situation, choose: Describe reinforcement learning from human feedback as preference-guided alignment.
Incorrect Answers
A:This is appropriate only after a distinct change: The organization then needs independent bias verification. It is not the best initial response to a model's answers are improved by collecting ranked human preferences between alternative safe responses.
B:Preferences express judged quality and do not guarantee truth on every fact. The proposal, Assume every human preference is a verified factual answer key, does not satisfy this need.
C:This can be valid for a different task: A data scientist wants to fine-tune an FM using only the most recent 20 examples from one customer segment. The current request is different and needs the selected approach.
Question 322
A model's answers are improved by collecting ranked human preferences between alternative safe responses. A colleague suggests: Assume every human preference is a verified factual answer key. Which finding BEST explains why the suggestion should be rejected?
- Preferences express judged quality and do not guarantee truth on every fact
- A narrow sample cannot represent all relevant user populations
- Compression can reduce capability and must be tested for target use cases
- The shortcut is safe because its terminology is familiar to the team
Explanation
Correct Answer
A:Preferences express judged quality and do not guarantee truth on every fact. This is the direct defect in the proposed action: Assume every human preference is a verified factual answer key.
Incorrect Answers
B:This finding is associated with a different problem: A data scientist wants to fine-tune an FM using only the most recent 20 examples from one customer segment.
C:This concern belongs to another situation: An operations team considers using a teacher model to train a smaller student to lower inference expense.
D:Familiar wording is not proof that an action is appropriate. The actual defect is: Preferences express judged quality and do not guarantee truth on every fact.
Question 323
A model's answers are improved by collecting ranked human preferences between alternative safe responses. After implementing Describe reinforcement learning from human feedback as preference-guided alignment, which observation BEST verifies the intended result?
- Held-out evaluations include different segments and representative failure cases
- Evaluation shows improved rated helpfulness while factual checks remain separate
- The team schedules a future test of whether evaluation shows improved rated helpfulness while factual checks remain separate, but has no result
- The team approves Describe reinforcement learning from human feedback as preference-guided alignment but does not run an operational test
Explanation
Correct Answer
B:The observed result, Evaluation shows improved rated helpfulness while factual checks remain separate, provides direct evidence. Human preference signals can shape model behavior beyond next-token pretraining.
Incorrect Answers
A:This observation would be evidence for another case: A data scientist wants to fine-tune an FM using only the most recent 20 examples from one customer segment, not the current scenario.
C:Verify the actual outcome: Evaluation shows improved rated helpfulness while factual checks remain separate.
D:The required observable condition is: Evaluation shows improved rated helpfulness while factual checks remain separate.
Question 324
The initial scenario was: A model's answers are improved by collecting ranked human preferences between alternative safe responses. The original control is already in place. Now the organization then needs independent bias verification. Which ADDITIONAL action BEST addresses the new condition?
- Evaluation shows improved rated helpfulness while factual checks remain separate
- Assume every human preference is a verified factual answer key
- Add subgroup fairness evaluation and human auditing
- Describe reinforcement learning from human feedback as preference-guided alignment
Explanation
Correct Answer
C:The organization then needs independent bias verification. The appropriate added action is: Add subgroup fairness evaluation and human auditing.
Incorrect Answers
A:This describes evidence for the original approach, not a new action: Evaluation shows improved rated helpfulness while factual checks remain separate.
B:This original shortcut remains unsuitable: Preferences express judged quality and do not guarantee truth on every fact.
D:This was the original control and is already implemented. Human preference signals can shape model behavior beyond next-token pretraining.
Question 325
A data scientist wants to fine-tune an FM using only the most recent 20 examples from one customer segment. Which option BEST meets the stated AI/ML or AWS requirement?
- Investigate coverage gaps and collect governed examples for that segment
- Build a representative curated training and validation dataset
- Use the 20 biased examples as the sole evidence of universal model quality
- Describe model distillation as teacher-to-student behavior transfer
Explanation
Correct Answer
B:A tiny skewed sample risks overfitting and weak generalization. In this situation, choose: Build a representative curated training and validation dataset.
Incorrect Answers
A:This is appropriate only after a distinct change: A new segment repeatedly fails the deployed model's responses. It is not the best initial response to a data scientist wants to fine-tune an fm using only the most recent 20 examples from one customer segment.
C:A narrow sample cannot represent all relevant user populations. The proposal, Use the 20 biased examples as the sole evidence of universal model quality, does not satisfy this need.
D:This can be valid for a different task: An operations team considers using a teacher model to train a smaller student to lower inference expense. The current request is different and needs the selected approach.
Question 326
A data scientist wants to fine-tune an FM using only the most recent 20 examples from one customer segment. A colleague suggests: Use the 20 biased examples as the sole evidence of universal model quality. Which finding BEST explains why the suggestion should be rejected?
- Compression can reduce capability and must be tested for target use cases
- A narrow sample cannot represent all relevant user populations
- In-context examples are not large-scale weight-training pretraining
- The shortcut is safe because its terminology is familiar to the team
Explanation
Correct Answer
B:A narrow sample cannot represent all relevant user populations. This is the direct defect in the proposed action: Use the 20 biased examples as the sole evidence of universal model quality.
Incorrect Answers
A:This finding is associated with a different problem: An operations team considers using a teacher model to train a smaller student to lower inference expense.
C:This concern belongs to another situation: A company asks whether a general language model can be pretrained on millions of diverse documents before specializing it.
D:Familiar wording is not proof that an action is appropriate. The actual defect is: A narrow sample cannot represent all relevant user populations.
Question 327
A data scientist wants to fine-tune an FM using only the most recent 20 examples from one customer segment. After implementing Build a representative curated training and validation dataset, which observation BEST verifies the intended result?
- The team schedules a future test of whether held-out evaluations include different segments and representative failure cases, but has no result
- Held-out evaluations include different segments and representative failure cases
- The team approves Build a representative curated training and validation dataset but does not run an operational test
- The distilled model reaches acceptable quality on held-out tasks at lower cost
Explanation
Correct Answer
B:The observed result, Held-out evaluations include different segments and representative failure cases, provides direct evidence. A tiny skewed sample risks overfitting and weak generalization.
Incorrect Answers
A:Verify the actual outcome: Held-out evaluations include different segments and representative failure cases.
C:The required observable condition is: Held-out evaluations include different segments and representative failure cases.
D:This observation would be evidence for another case: An operations team considers using a teacher model to train a smaller student to lower inference expense, not the current scenario.
Question 328
The initial scenario was: A data scientist wants to fine-tune an FM using only the most recent 20 examples from one customer segment. The original control is already in place. Now a new segment repeatedly fails the deployed model's responses. Which ADDITIONAL action BEST addresses the new condition?
- Held-out evaluations include different segments and representative failure cases
- Investigate coverage gaps and collect governed examples for that segment
- Use the 20 biased examples as the sole evidence of universal model quality
- Build a representative curated training and validation dataset
Explanation
Correct Answer
B:A new segment repeatedly fails the deployed model's responses. The appropriate added action is: Investigate coverage gaps and collect governed examples for that segment.
Incorrect Answers
A:This describes evidence for the original approach, not a new action: Held-out evaluations include different segments and representative failure cases.
C:This original shortcut remains unsuitable: A narrow sample cannot represent all relevant user populations.
D:This was the original control and is already implemented. A tiny skewed sample risks overfitting and weak generalization.
Question 329
An operations team considers using a teacher model to train a smaller student to lower inference expense. Which option BEST meets the stated AI/ML or AWS requirement?
- Reassess the stronger teacher model or more suitable alternatives
- Describe large-scale pretraining as learning broad patterns from extensive data
- Treat the student's smaller parameter count as proof no quality evaluation is needed
- Describe model distillation as teacher-to-student behavior transfer
Explanation
Correct Answer
D:Distillation targets lower serving cost with some retained teacher capability. In this situation, choose: Describe model distillation as teacher-to-student behavior transfer.
Incorrect Answers
A:This is appropriate only after a distinct change: The application now requires complex legal reasoning beyond the pilot. It is not the best initial response to an operations team considers using a teacher model to train a smaller student to lower inference expense.
B:This can be valid for a different task: A company asks whether a general language model can be pretrained on millions of diverse documents before specializing it. The current request is different and needs the selected approach.
C:Compression can reduce capability and must be tested for target use cases. The proposal, Treat the student's smaller parameter count as proof no quality evaluation is needed, does not satisfy this need.
Question 330
An operations team considers using a teacher model to train a smaller student to lower inference expense. A colleague suggests: Treat the student's smaller parameter count as proof no quality evaluation is needed. Which finding BEST explains why the suggestion should be rejected?
- Compression can reduce capability and must be tested for target use cases
- Uncurated unrelated data is unlikely to teach the required instructions
- The shortcut is safe because its terminology is familiar to the team
- In-context examples are not large-scale weight-training pretraining
Explanation
Correct Answer
A:Compression can reduce capability and must be tested for target use cases. This is the direct defect in the proposed action: Treat the student's smaller parameter count as proof no quality evaluation is needed.
Incorrect Answers
B:This concern belongs to another situation: An enterprise has curated customer-service dialogues and wants a pretrained model to follow its response protocol.
C:Familiar wording is not proof that an action is appropriate. The actual defect is: Compression can reduce capability and must be tested for target use cases.
D:This finding is associated with a different problem: A company asks whether a general language model can be pretrained on millions of diverse documents before specializing it.