Microsoft AI-300 (Operationalizing Machine Learning and Generative AI Solutions) Practice Test 58

Question 570

An application team evaluates competing retrieval improvements before production rollout. An operator is tasked to route a limited share of users to an approved retrieval variant and monitor real-world task success before expansion. Which mechanism provides that capability?

  1. Query-set holdout
  2. Chunking A/B experiment
  3. Hybrid versus vector A/B
  4. Production cohort rollout

Explanation

Correct Answer

D:

Cohort rollout limits exposure to ranking changes. It should have clear assignment, telemetry, and rollback criteria.

Incorrect Answers

A:

A stable query set isolates the effects of indexing and search changes; changing the benchmark during comparison can invalidate conclusions; the outstanding need is to route a limited share of users to an approved retrieval variant and monitor real-world task success before expansion.

B:

Controlled chunking trials show whether boundaries affect retrieval and answer quality; changing the generator simultaneously would confound attribution; this approach leaves the team unable to route a limited share of users to an approved retrieval variant and monitor real-world task success before expansion.

C:

Paired search experiments capture exact-code and paraphrase behavior; a theoretical advantage should be checked on actual queries; the needed capability is to route a limited share of users to an approved retrieval variant and monitor real-world task success before expansion.

Question 571

A generative AI team considers whether to customize a foundation model for a specialized workflow. The immediate design goal is to try clear instructions and representative examples before paying for a model customization process. Which approach addresses that goal?

  1. Model support verification
  2. Cost-benefit estimation
  3. Prompt engineering baseline
  4. Supervised fine-tuning selection

Explanation

Correct Answer

C:

Prompt iteration is often a lower-cost first step for behavior guidance. It should be evaluated on the same cases proposed for fine-tuning.

Incorrect Answers

A:

Fine-tuning availability differs across model families and offerings; a model's inference availability does not guarantee that it can be fine-tuned; the requirement, by contrast, is to try clear instructions and representative examples before paying for a model customization process.

B:

Tuning has lifecycle costs beyond a one-time training job; a small quality increase may not justify added operational complexity; the team would still need to try clear instructions and representative examples before paying for a model customization process.

D:

Supervised tuning adjusts model behavior toward demonstrated examples; it requires suitable training data, supported model offerings, and post-tuning evaluation; the implementation under review needs to try clear instructions and representative examples before paying for a model customization process.

Question 572

A generative AI team considers whether to customize a foundation model for a specialized workflow. The team is required to use retrieval when answers must reflect frequently changing internal policies rather than embedding those facts in model weights. Which option is the best fit?

  1. RAG before factual fine-tuning
  2. Cost-benefit estimation
  3. Task-specific training objective
  4. Model support verification

Explanation

Correct Answer

A:

RAG supplies current approved evidence at inference time. Fine-tuning alone is a poor substitute for rapidly changing authoritative documents.

Incorrect Answers

B:

Tuning has lifecycle costs beyond a one-time training job; a small quality increase may not justify added operational complexity; the requested task remains to use retrieval when answers must reflect frequently changing internal policies rather than embedding those facts in model weights.

C:

A clear objective guides example selection and evaluation; customization can degrade other capabilities if training is misaligned; this alternative does not achieve the requirement to use retrieval when answers must reflect frequently changing internal policies rather than embedding those facts in model weights.

D:

Fine-tuning availability differs across model families and offerings; a model's inference availability does not guarantee that it can be fine-tuned; the actual design criterion is to use retrieval when answers must reflect frequently changing internal policies rather than embedding those facts in model weights.

Question 573

A generative AI team considers whether to customize a foundation model for a specialized workflow. An operator is tasked to customize task behavior using high-quality input-output examples when prompting is insufficient and the model supports tuning. Which mechanism provides that capability?

  1. RAG before factual fine-tuning
  2. Supervised fine-tuning selection
  3. Prompt engineering baseline
  4. Model support verification

Explanation

Correct Answer

B:

Supervised tuning adjusts model behavior toward demonstrated examples. It requires suitable training data, supported model offerings, and post-tuning evaluation.

Incorrect Answers

A:

RAG supplies current approved evidence at inference time; fine-tuning alone is a poor substitute for rapidly changing authoritative documents; the actual design criterion is to customize task behavior using high-quality input-output examples when prompting is insufficient and the model supports tuning.

C:

Prompt iteration is often a lower-cost first step for behavior guidance; it should be evaluated on the same cases proposed for fine-tuning; this approach leaves the team unable to customize task behavior using high-quality input-output examples when prompting is insufficient and the model supports tuning.

D:

Fine-tuning availability differs across model families and offerings; a model's inference availability does not guarantee that it can be fine-tuned; the needed capability is to customize task behavior using high-quality input-output examples when prompting is insufficient and the model supports tuning.

Question 574

A generative AI team considers whether to customize a foundation model for a specialized workflow. The immediate design goal is to define the exact behavior or format to improve instead of treating fine-tuning as a universal accuracy booster. Which approach addresses that goal?

  1. Supervised fine-tuning selection
  2. Task-specific training objective
  3. RAG before factual fine-tuning
  4. Model support verification
  5. Prompt engineering baseline
  6. Cost-benefit estimation

Explanation

Correct Answer

B:

A clear objective guides example selection and evaluation. Customization can degrade other capabilities if training is misaligned.

Incorrect Answers

A:

Supervised tuning adjusts model behavior toward demonstrated examples; it requires suitable training data, supported model offerings, and post-tuning evaluation; the needed capability is to define the exact behavior or format to improve instead of treating fine-tuning as a universal accuracy booster.

C:

RAG supplies current approved evidence at inference time; fine-tuning alone is a poor substitute for rapidly changing authoritative documents; the team would still need to define the exact behavior or format to improve instead of treating fine-tuning as a universal accuracy booster.

D:

Fine-tuning availability differs across model families and offerings; a model's inference availability does not guarantee that it can be fine-tuned; the specified outcome instead requires the team to define the exact behavior or format to improve instead of treating fine-tuning as a universal accuracy booster.

E:

Prompt iteration is often a lower-cost first step for behavior guidance; it should be evaluated on the same cases proposed for fine-tuning; the implementation under review needs to define the exact behavior or format to improve instead of treating fine-tuning as a universal accuracy booster.

F:

Tuning has lifecycle costs beyond a one-time training job; a small quality increase may not justify added operational complexity; the requested task remains to define the exact behavior or format to improve instead of treating fine-tuning as a universal accuracy booster.

Question 575

A generative AI team considers whether to customize a foundation model for a specialized workflow. The team is required to compare preparation, training, hosting, and maintenance expense with measured gains against the baseline. Which option is the best fit?

  1. Model support verification
  2. Prompt engineering baseline
  3. Cost-benefit estimation
  4. RAG before factual fine-tuning

Explanation

Correct Answer

C:

Tuning has lifecycle costs beyond a one-time training job. A small quality increase may not justify added operational complexity.

Incorrect Answers

A:

Fine-tuning availability differs across model families and offerings; a model's inference availability does not guarantee that it can be fine-tuned; the specified outcome instead requires the team to compare preparation, training, hosting, and maintenance expense with measured gains against the baseline.

B:

Prompt iteration is often a lower-cost first step for behavior guidance; it should be evaluated on the same cases proposed for fine-tuning; the implementation under review needs to compare preparation, training, hosting, and maintenance expense with measured gains against the baseline.

D:

RAG supplies current approved evidence at inference time; fine-tuning alone is a poor substitute for rapidly changing authoritative documents; this alternative does not achieve the requirement to compare preparation, training, hosting, and maintenance expense with measured gains against the baseline.

Question 576

An agent answers questions about policies that are rewritten every week. A proposal suggests fine-tuning to memorize those policy changes, but the team also plans customization of a base model for stable output behavior and has not checked regional tuning support. Which TWO decisions should guide the design before any tuning job begins? Choose TWO.

  1. Task-specific training objective
  2. Prompt engineering baseline
  3. Model support verification
  4. Cost-benefit estimation
  5. RAG before factual fine-tuning

Explanation

Correct Answers

C:

Fine-tuning availability differs across model families and offerings; a model's inference availability does not guarantee that it can be fine-tuned. Here, feasibility is checked against the supported model and target region.

E:

RAG supplies current approved evidence at inference time; fine-tuning alone is a poor substitute for rapidly changing authoritative documents. Here, frequently edited policy facts stay in retrieval instead of being baked into weights.

Incorrect Answers

A:

A clear objective guides example selection and evaluation; customization can degrade other capabilities if training is misaligned; the observed fault still involves stale policy facts and an unverified tuning-support constraint.

B:

Prompt iteration is often a lower-cost first step for behavior guidance; it should be evaluated on the same cases proposed for fine-tuning; the immediate requirement instead concerns stale policy facts and an unverified tuning-support constraint.

D:

Tuning has lifecycle costs beyond a one-time training job; a small quality increase may not justify added operational complexity; this mechanism leaves unresolved stale policy facts and an unverified tuning-support constraint.

Question 577

A team wants to adapt a large model without updating every pretrained parameter. The immediate design goal is to train low-rank adapter weights for eligible layers instead of modifying all original model parameters. Which approach addresses that goal?

  1. LoRA adaptation
  2. Adapter version governance
  3. Adapter compatibility check
  4. Adapter rank selection

Explanation

Correct Answer

A:

LoRA reduces the number of trainable parameters for supported architectures. It can lower memory needs while preserving a frozen base model.

Incorrect Answers

B:

Adapter releases are incomplete without a compatible base reference; tracking only the adapter filename can lead to silent version mismatches; the needed capability is to train low-rank adapter weights for eligible layers instead of modifying all original model parameters.

C:

Adapters depend on compatible underlying weights and modules; attaching one to an unrelated base model can fail or degrade behavior; the requirement, by contrast, is to train low-rank adapter weights for eligible layers instead of modifying all original model parameters.

D:

The rank influences adaptation capacity and resource use; a larger rank is not automatically more accurate or better generalized; the team would still need to train low-rank adapter weights for eligible layers instead of modifying all original model parameters.

Question 578

A team wants to adapt a large model without updating every pretrained parameter. The team is required to tune the low-rank capacity to balance specialized behavior with parameter count and training cost. Which option is the best fit?

  1. Frozen base weights
  2. Adapter rank selection
  3. LoRA adaptation
  4. Adapter compatibility check

Explanation

Correct Answer

B:

The rank influences adaptation capacity and resource use. A larger rank is not automatically more accurate or better generalized.

Incorrect Answers

A:

Freezing the base reduces trainable parameter count and often memory use; it differs from full-model fine-tuning where all selected weights change; the team would still need to tune the low-rank capacity to balance specialized behavior with parameter count and training cost.

C:

LoRA reduces the number of trainable parameters for supported architectures; it can lower memory needs while preserving a frozen base model; the implementation under review needs to tune the low-rank capacity to balance specialized behavior with parameter count and training cost.

D:

Adapters depend on compatible underlying weights and modules; attaching one to an unrelated base model can fail or degrade behavior; the requested task remains to tune the low-rank capacity to balance specialized behavior with parameter count and training cost.

Question 579

A team wants to adapt a large model without updating every pretrained parameter. An operator is tasked to keep original model parameters fixed when using an adapter-based training method designed for parameter efficiency. Which mechanism provides that capability?

  1. Adapter rank selection
  2. Adapter compatibility check
  3. Full versus PEFT comparison
  4. Frozen base weights

Explanation

Correct Answer

D:

Freezing the base reduces trainable parameter count and often memory use. It differs from full-model fine-tuning where all selected weights change.

Incorrect Answers

A:

The rank influences adaptation capacity and resource use; a larger rank is not automatically more accurate or better generalized; the requested task remains to keep original model parameters fixed when using an adapter-based training method designed for parameter efficiency.

B:

Adapters depend on compatible underlying weights and modules; attaching one to an unrelated base model can fail or degrade behavior; this alternative does not achieve the requirement to keep original model parameters fixed when using an adapter-based training method designed for parameter efficiency.

C:

Parameter-efficient tuning trades customization flexibility for efficiency; empirical task results should decide whether the savings justify the method; the actual design criterion is to keep original model parameters fixed when using an adapter-based training method designed for parameter efficiency.

Question 580

A team wants to adapt a large model without updating every pretrained parameter. The immediate design goal is to match a trained adapter to its intended base model and architecture during deployment. Which approach addresses that goal?

  1. Adapter rank selection
  2. Adapter version governance
  3. Frozen base weights
  4. Adapter compatibility check

Explanation

Correct Answer

D:

Adapters depend on compatible underlying weights and modules. Attaching one to an unrelated base model can fail or degrade behavior.

Incorrect Answers

A:

The rank influences adaptation capacity and resource use; a larger rank is not automatically more accurate or better generalized; the outstanding need is to match a trained adapter to its intended base model and architecture during deployment.

B:

Adapter releases are incomplete without a compatible base reference; tracking only the adapter filename can lead to silent version mismatches; this approach leaves the team unable to match a trained adapter to its intended base model and architecture during deployment.

C:

Freezing the base reduces trainable parameter count and often memory use; it differs from full-model fine-tuning where all selected weights change; the needed capability is to match a trained adapter to its intended base model and architecture during deployment.

Operationalizing Machine Learning and Generative AI Solutions (AI-300) Practice Tests