Question 770
A generator labels adversarial prompt-injection attempts as successful compliant answers. What correction is needed?
Explanation
Correct Answer
Synthetic labels must reward correct boundary handling rather than teach the model to follow malicious injected instructions.
Incorrect Answers
Create reviewed negative and boundary cases with correct refusal labels for underrepresented classes. This is useful for the distinct issue: A fine-tuning set has many positive approvals but few legitimate refusals, causing unsafe permissive behavior. The relevant technical rationale or evidence for that issue is: Class coverage should reflect critical decision boundaries instead of duplicating the majority positive response pattern. Here the reported issue is: A generator labels adversarial prompt-injection attempts as successful compliant answers. The relevant response is: Review adversarial examples against safety policy and label safe refusal or containment behavior. The first result cannot verify this different correction: Synthetic labels must reward correct boundary handling rather than teach the model to follow malicious injected instructions.
Review a small seed batch for factual quality and distribution before large-scale generation. It addresses another case: A fine-tuning team wants to scale synthetic generation from 20 to 20,000 cases without checking the first batch. For that case, the distinguishing justification or acceptance check is: Early validation catches defective prompting and labeling rules before they are amplified across thousands of examples. The present case instead states: A generator labels adversarial prompt-injection attempts as successful compliant answers. It calls for Review adversarial examples against safety policy and label safe refusal or containment behavior. Showing success for the other case would not establish: Synthetic labels must reward correct boundary handling rather than teach the model to follow malicious injected instructions.
The alternative, Separate data generation sources from the reserved test set and screen for leakage, belongs to a different technical condition: An LLM creates new training examples by paraphrasing the exact held-out test set. There the source-specific mechanism or proof is: Contaminating holdout items inflates apparent accuracy and invalidates an independent evaluation of generalization. The condition asked about here is: A generator labels adversarial prompt-injection attempts as successful compliant answers. A suitable correction is Review adversarial examples against safety policy and label safe refusal or containment behavior; the targeted justification or result is: Synthetic labels must reward correct boundary handling rather than teach the model to follow malicious injected instructions.