Datasets:

Languages:
English
Multilinguality:
monolingual
Size Categories:
1K<n<10K
Language Creators:
expert-generated
Annotations Creators:
expert-generated
Source Datasets:
original
Tags:
License:

the label is inconsistent

#1
by Tsingfeng - opened

The label is webpage is inconsistent with the label in downloaded dataset.
For example, for the case "How many points make up a perfect fivepin bowling score ?", the label from webpage is "5 (NUM)", but the label from "load_dataset('trec')" is "4", stand for "Location". Obviously, the case is about "number" instead of "location".

Then the following label map is incorrect:
'ABBR' (0): Abbreviation.
'ENTY' (1): Entity.
'DESC' (2): Description and abstract concept.
'HUM' (3): Human being.
'LOC' (4): Location.
'NUM' (5): Numeric value.

Datasets Maintainers org

Hi @Tsingfeng ,

Please note that the labels in trec were fixed on 22 Aug: https://huggingface.co/datasets/trec/commit/1f97567bdd2adedefe8abdaa9bd6ee0e6725b458

That PR:

  • replaces the fine labels, so that there are 50 instead of 47
  • once more labels are added, all they (fine and coarse) have been re-ordered, so that they align with the order in: https://cogcomp.seas.upenn.edu/Data/QA/QC/definition.html
  • the feature names have been fixed: fine_label instead of label-fine
    • to sneak-case (underscores instead of hyphens)
    • words have been reordered

More info:

Therefore, the mapping in the dataset card is the right one.

I guess you are getting the old version of trec when doing load_dataset("trec")... To get the new version, you should update your datasets version:

pip install -U datasets

Sign up or log in to comment