Hugging Face's logo

Datasets:
clue

Tasks:
Text Classification
Multiple Choice
Sub-tasks:
topic-classification
semantic-similarity-scoring
natural-language-inference
Languages:
Chinese
Multilinguality:
monolingual
Size Categories:
100K<n<1M
Language Creators:
other
Annotations Creators:
other
Source Datasets:
original
ArXiv:
Tags:
coreference-nli
qa-nli
License:
Dataset card Viewer Files Community
8

cluewsc2020 label is wrong

#5
by Doreamonzzz - opened
Discussion
Doreamonzzz

It seems that on line 359 of clue.py, the code should be

"label": datasets.features.ClassLabel(names=["false", "true"]),

instead of 

"label": datasets.features.ClassLabel(names=["true", "false"]),

While on line 178, the label_classes are right:

label_classes=["false", "true"],
albertvillanova
 Datasets Maintainers org

Thanks for reporting, @Doreamonzzz .
We are investigating it...

albertvillanova
 Datasets Maintainers org

After investigation, we have checked that the data is coherent with "label": datasets.features.ClassLabel(names=["true", "false"]) (the order in label_classes=["false", "true"] is ignored)
See: https://huggingface.co/datasets/clue/viewer/cluewsc2020/train

  • When the source data is False, it is labeled as False and assigned the integer 1
  • When the source data is True, it is labeled as True and assigned the integer 0

Maybe we could align both definitions (or better, use only one), so that no other people get confused about a possible issue with the labels.

Doreamonzzz

Thank you, my concern has been solved.

Doreamonzzz changed discussion status to closed
albertvillanova
 Datasets Maintainers org

This PR defines class labels only once for cluewsc2020 config: #6

Sign up or log in to comment