Abilt models typically perform worse than their bases at the same tasks, so while I'd use one to evaluate content knowledge, I'd probably ultimately stick to on

Abilt models typically perform worse than their bases at the same tasks, so while I'd use one to evaluate content knowledge, I'd probably ultimately stick to one from a family I could fool with abstraction or coerce through system prompt.