Describe the bug
According the doc here: https://huggingface.co/docs/hub/datasets-file-names-and-splits#custom-split-name
it should infer the split name from the subdir of data or the beg of the name of the files in data.
When doing this manually through web upload it does not work. it uses "train" as a unique split.
example: https://huggingface.co/datasets/eole-nlp/estimator_chatml
Steps to reproduce the bug
follow the link above
Expected behavior
there should be two splits "mlqe" and "1720_da"
Environment info
website