Not sure what you mean? It’s more sensitive against disabled people than the other, more rare strings the author chose like “non disabled people”, which would be consistent with my theory that there is more hate against disabled people in the training set and so the model is more sensitive against that being in the sequence of tokens