Question 590A public-sector NLP model team generates synthetic citizen requests and must ensure they are not copied from private submissions. Why does the recommended approach address the observed issue? Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. Fine-tuning changes model behavior and style but is not a reliable substitute for retrieving current authoritative records. Overly small chunks can lose context while oversized chunks may dilute relevant passages and consume context budget. ExplanationCorrect Answer A:Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. This explains the relevant intervention: Generate synthetic fine-tuning examples only within approved privacy constraints and validate realism, diversity, and label correctness before training. Confirm by: Compare generated examples against known label rules and check a held-out sample for duplication, leakage, and task fidelity. Other Answers B:Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. The more specific mechanism behind the reported outcome is: Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. C:Fine-tuning changes model behavior and style but is not a reliable substitute for retrieving current authoritative records. The more specific mechanism behind the reported outcome is: Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. D:Overly small chunks can lose context while oversized chunks may dilute relevant passages and consume context budget. The more specific mechanism behind the reported outcome is: Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias.
Question 591A public-sector NLP model team generates synthetic citizen requests and must ensure they are not copied from private submissions. What observable evidence should the engineering team collect? Compare a retrieval-first baseline and a small qualified fine-tuning pilot against separate knowledge-freshness and output-format tests. Compare generated examples against known label rules and check a held-out sample for duplication, leakage, and task fidelity. Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. Compare retrieved-neighbor quality on a held-out domain query set after reindexing all required documents. ExplanationCorrect Answer B:Compare generated examples against known label rules and check a held-out sample for duplication, leakage, and task fidelity. This directly tests whether the intervention works: Generate synthetic fine-tuning examples only within approved privacy constraints and validate realism, diversity, and label correctness before training. The technical reason is: Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. Other Answers A:Fine-tuning changes model behavior and style but is not a reliable substitute for retrieving current authoritative records. The applicable evidence for this requirement is: Compare generated examples against known label rules and check a held-out sample for duplication, leakage, and task fidelity. C:Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. The applicable evidence for this requirement is: Compare generated examples against known label rules and check a held-out sample for duplication, leakage, and task fidelity. D:Vectors from different embedding models or dimensionalities are not safely interchangeable within one unchanged similarity index. The applicable evidence for this requirement is: Compare generated examples against known label rules and check a held-out sample for duplication, leakage, and task fidelity.
Question 592A public-sector NLP model team generates synthetic citizen requests and must ensure they are not copied from private submissions. Which documented safeguard helps sustain the improvement? Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models. Record synthetic-data generation methods and review quality and rights alongside ordinary labeled data. Document the decision criteria and update cadence before collecting sensitive fine-tuning examples. Require staged retrieval configuration rollout with a rollback threshold for groundedness or relevance regressions. ExplanationCorrect Answer B:Record synthetic-data generation methods and review quality and rights alongside ordinary labeled data. This reduces the failure risk because Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. Immediate action: Generate synthetic fine-tuning examples only within approved privacy constraints and validate realism, diversity, and label correctness before training. Other Answers A:Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. The preventive control that addresses this condition is: Record synthetic-data generation methods and review quality and rights alongside ordinary labeled data. C:Fine-tuning changes model behavior and style but is not a reliable substitute for retrieving current authoritative records. The preventive control that addresses this condition is: Record synthetic-data generation methods and review quality and rights alongside ordinary labeled data. D:A/B tests isolate the effect of retrieval configuration changes on relevance, groundedness, cost, and latency. The preventive control that addresses this condition is: Record synthetic-data generation methods and review quality and rights alongside ordinary labeled data.
Question 593An invoice extraction model improves field accuracy after fine-tuning but unexpectedly loses refusal behavior on inappropriate requests. Which supported solution best fits this production constraint? Prepare high-quality, consented, correctly formatted training examples with deduplication, leakage checks, and representative task coverage. Generate synthetic fine-tuning examples only within approved privacy constraints and validate realism, diversity, and label correctness before training. Evaluate a domain-appropriate embedding model and rebuild vector indexes consistently when changing embedding representation. Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. ExplanationCorrect Answer D:Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. This is appropriate because Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. Verify with: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. Other Answers A:Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. The required action here is: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. B:Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. The required action here is: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. C:Vectors from different embedding models or dimensionalities are not safely interchangeable within one unchanged similarity index. The required action here is: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline.
Question 594An invoice extraction model improves field accuracy after fine-tuning but unexpectedly loses refusal behavior on inappropriate requests. Why is the selected configuration more appropriate here? Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. Low relevance thresholds can admit unrelated evidence, while overly strict cutoffs may omit needed sources entirely. Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. ExplanationCorrect Answer D:Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. This explains the relevant intervention: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. Confirm by: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. Other Answers A:Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. The more specific mechanism behind the reported outcome is: Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. B:Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. The more specific mechanism behind the reported outcome is: Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. C:Low relevance thresholds can admit unrelated evidence, while overly strict cutoffs may omit needed sources entirely. The more specific mechanism behind the reported outcome is: Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter.
Question 595An invoice extraction model improves field accuracy after fine-tuning but unexpectedly loses refusal behavior on inappropriate requests. Which controlled test would best verify the outcome? Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. Sample the curated dataset for labeling consistency, schema validity, privacy requirements, and overlap with held-out tests. Compare a retrieval-first baseline and a small qualified fine-tuning pilot against separate knowledge-freshness and output-format tests. Split approved test queries between baseline and candidate retrieval variants and compare consistent outcome measures. ExplanationCorrect Answer A:Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. This directly tests whether the intervention works: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. The technical reason is: Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. Other Answers B:Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. The applicable evidence for this requirement is: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. C:Fine-tuning changes model behavior and style but is not a reliable substitute for retrieving current authoritative records. The applicable evidence for this requirement is: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. D:A/B tests isolate the effect of retrieval configuration changes on relevance, groundedness, cost, and latency. The applicable evidence for this requirement is: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics.
Question 596An invoice extraction model improves field accuracy after fine-tuning but unexpectedly loses refusal behavior on inappropriate requests. What governance control should accompany this configuration? Track data rights, labeling guidelines, source versions, and a separate evaluation split for every fine-tuning run. Record synthetic-data generation methods and review quality and rights alongside ordinary labeled data. Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models. Track embedding model versions and coordinate index migrations with rollback-ready deployment configuration. ExplanationCorrect Answer C:Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models. This reduces the failure risk because Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. Immediate action: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. Other Answers A:Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. The preventive control that addresses this condition is: Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models. B:Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. The preventive control that addresses this condition is: Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models. D:Vectors from different embedding models or dimensionalities are not safely interchangeable within one unchanged similarity index. The preventive control that addresses this condition is: Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models.
Question 597A text classification model is customized for a new industry and needs a reversible production rollout with ongoing bias checks. What should the AI operations engineer do first? Prepare high-quality, consented, correctly formatted training examples with deduplication, leakage checks, and representative task coverage. Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. Use retrieval or prompt changes for frequently changing factual knowledge and consider fine-tuning only for a stable, evidence-backed model behavior gap. Combine lexical keyword matches with vector similarity retrieval when queries include both semantic concepts and exact product identifiers. ExplanationCorrect Answer B:Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. This is appropriate because Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. Verify with: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. Other Answers A:Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. The required action here is: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. C:Fine-tuning changes model behavior and style but is not a reliable substitute for retrieving current authoritative records. The required action here is: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. D:Hybrid search improves coverage when semantic similarity alone misses literal codes or keyword search misses paraphrased meaning. The required action here is: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline.
Question 598A text classification model is customized for a new industry and needs a reversible production rollout with ongoing bias checks. Which infrastructure or model-lifecycle behavior is relevant? Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. Fine-tuning changes model behavior and style but is not a reliable substitute for retrieving current authoritative records. Hybrid search improves coverage when semantic similarity alone misses literal codes or keyword search misses paraphrased meaning. Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. ExplanationCorrect Answer D:Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. This explains the relevant intervention: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. Confirm by: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. Other Answers A:Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. The more specific mechanism behind the reported outcome is: Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. B:Fine-tuning changes model behavior and style but is not a reliable substitute for retrieving current authoritative records. The more specific mechanism behind the reported outcome is: Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. C:Hybrid search improves coverage when semantic similarity alone misses literal codes or keyword search misses paraphrased meaning. The more specific mechanism behind the reported outcome is: Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter.
Question 599A text classification model is customized for a new industry and needs a reversible production rollout with ongoing bias checks. Which test differentiates the desired fix from unrelated behavior? Sample the curated dataset for labeling consistency, schema validity, privacy requirements, and overlap with held-out tests. Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. Compare generated examples against known label rules and check a held-out sample for duplication, leakage, and task fidelity. Measure retrieval precision and recall on queries containing both catalog IDs and natural-language problem descriptions. ExplanationCorrect Answer B:Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. This directly tests whether the intervention works: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. The technical reason is: Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. Other Answers A:Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. The applicable evidence for this requirement is: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. C:Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. The applicable evidence for this requirement is: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics. D:Hybrid search improves coverage when semantic similarity alone misses literal codes or keyword search misses paraphrased meaning. The applicable evidence for this requirement is: Run a fixed evaluation and safety suite on candidate and baseline before a small production canary with monitored metrics.
Question 600A text classification model is customized for a new industry and needs a reversible production rollout with ongoing bias checks. Which operating practice is most likely to prevent recurrence? Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models. Track data rights, labeling guidelines, source versions, and a separate evaluation split for every fine-tuning run. Record synthetic-data generation methods and review quality and rights alongside ordinary labeled data. Version chunking settings and reindex using a tested migration plan when document structure changes. ExplanationCorrect Answer A:Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models. This reduces the failure risk because Fine-tuning can alter both desired task behavior and unintended safety characteristics, so versioned evaluation and rollback matter. Immediate action: Deploy a fine-tuned candidate as a versioned deployment and compare its task success, safety, and drift with the approved baseline. Other Answers B:Fine-tuning quality depends strongly on input labels and diversity; inconsistent examples can reinforce the wrong output behavior. The preventive control that addresses this condition is: Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models. C:Synthetic examples can expand coverage but can also propagate hallucinated labels, hidden real-data copies, or distribution bias. The preventive control that addresses this condition is: Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models. D:Overly small chunks can lose context while oversized chunks may dilute relevant passages and consume context budget. The preventive control that addresses this condition is: Require reproducible training metadata, release approvals, and a tested rollback path for fine-tuned models.