We verified LM-written data with human evaluators, who agreed with the data’s labels and rated the examples favorably on both diversity and relevance to the tested behavior. We’ve released our evaluations at

→ View original post on X — @anthropicai,