Datasets:

dleemiller
/

all-nli-distill

language:
  - en
multilinguality:
  - monolingual
size_categories:
  - 1M<n<10M
task_categories:
  - feature-extraction
  - sentence-similarity
pretty_name: AllNLI
tags:
  - sentence-transformers
dataset_info:
  - config_name: pair-class-distill
    features:
      - name: premise
        dtype: string
      - name: hypothesis
        dtype: string
      - name: label
        dtype:
          class_label:
            names:
              '0': contradiction
              '1': entailment
              '2': neutral

Dataset Card for AllNLI

This dataset is a concatenation of the SNLI and MultiNLI datasets.

This is the same dataset as sentence-transformers/all-nli pair-class split; however, the label ids are not identical, and teacher scores have been added from dleemiller/ModernCE-large-nli.

I have also added hashes for score lookup, since a lookup must be added into a custom loss function, if using the sentence transformers CrossEncoder trainer.

The hashes were computed straightforwardly as follows:

df.hash = df.apply(lambda x: hashlib.md5(f"{x.premise}\n{x.hypothesis}".encode()).hexdigest(), axis=1)