This model uses pretrained distilroberta-base weights, fine tuned on the WNC wiki edits corpus to detect biased language. See https://github.com/rpryzant/neutralizing-bias for more details.