we need more transparent bias metrics and better disclosure of training data sources for these models