Heretic: Automatic censorship removal for language models

How do you remove censorship that appears due to the biased selection of training data?

in that case you'd need to do actual training/finetuning with a dataset that has information about things that were left out of the original training data.