I'm not sure there's any benchmark score that'd make me use a model that suddenly starts talking about racist conspiracy theories unprompted. Doubly so for anything intended for production use.