A user reports that the model refuses ‘how do I kill a...

To address the false refusal, first confirm the issue is reproducible on a held-out eval set. Next, inspect the safety training data to see if the vocabulary cluster around ‘kill’ is over-represented in harmful-refuse pairs. Fix the issue by adding borderline-positive examples that use the same technical vocabulary in legitimate contexts. Finally, re-evaluate to ensure refusal recall remains intact and add the specific case to a false-positive evaluation set to prevent future regressions.