I think it's due to a "collision" between the heuristics that define splits and/or classes based on the directories. There is a drop_labels=True option in the datasets library, if it helps.
Ideally, in this case, we should have two splits (train and test), and no additional label column.
I think the issue also exists with image datasets.
In example dataset https://huggingface.co/datasets/datasets-examples/doc-audio-4, we have an "unexpected" label column with only
nullvalues.I think it's due to a "collision" between the heuristics that define splits and/or classes based on the directories. There is a
drop_labels=Trueoption in the datasets library, if it helps.Ideally, in this case, we should have two splits (train and test), and no additional
labelcolumn.I think the issue also exists with image datasets.