I did research on this topic in a cyber safety research project. We focused on cyber bullying specifically but encountered the issue of non-toxic profanity as w

I did research on this topic in a cyber safety research project. We focused on cyber bullying specifically but encountered the issue of non-toxic profanity as well. We used neural representation learning as well as feature methods and indeed, common profanity words are weak predictors in the better models. Still we found instances of non-toxic profanity being classified as bullying. An immediate solution is to apply multitask methods to your target dataset and include the one proposed in OP. It's always good to have more resources like this, even though SurgeHQ overstates the size of their resources by large margin in their copy. The 1000 post instances of their dataset is far from "the largest": I have several aggression, toxicity and bullying human-annotated datasets right here with over 100k instances.