Flipped labels in implicit_ETHICS.json #48
Description
Hi Authors,
I found this problem when exploring the Machine Ethics datasets: The labels of rows in implicit_ETHICS.json seem to be incorrect for "topic"="justice". I further checked the original ETHICS dataset and found out that the label they used in each subset is not constant. For example in cm_train.csv label 0 means "not wrong" and 1 means "wrong", but in justice_train.csv it seems to be vice verse. Thus, half of the labels in the merged dataset implicit_ETHICS.json are somehow flipped.
Please have a look on this issue and correct it if it is really the case. Thanks.