monolingual-wideNLI / README.md
Gameselo's picture
Update README.md
eda09db verified
|
raw
history blame
2.57 kB
metadata
dataset_info:
  features:
    - name: premise
      dtype: string
    - name: hypothesis
      dtype: string
    - name: label
      dtype:
        class_label:
          names:
            '0': entailment
            '1': neutral
            '2': contradiction
  splits:
    - name: train
      num_bytes: 470925549
      num_examples: 2106339
    - name: test_anli_train_r1
      num_bytes: 6903286
      num_examples: 16946
    - name: test_anli_train_r2
      num_bytes: 18247198
      num_examples: 45460
    - name: test_anli_train_r3
      num_bytes: 38803242
      num_examples: 100459
    - name: test_climate_fever
      num_bytes: 2396137
      num_examples: 7675
    - name: dev
      num_bytes: 47121.524698369816
      num_examples: 120
  download_size: 207805194
  dataset_size: 537322533.5246984
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test_anli_train_r1
        path: data/test_anli_train_r1-*
      - split: test_anli_train_r2
        path: data/test_anli_train_r2-*
      - split: test_anli_train_r3
        path: data/test_anli_train_r3-*
      - split: test_climate_fever
        path: data/test_climate_fever-*
      - split: dev
        path: data/dev-*
task_categories:
  - text-classification
  - zero-shot-classification
language:
  - en
tags:
  - natural-language-inference
  - fact-checking
size_categories:
  - 100M<n<1B

This monolingual (English) NLI dataset is designed for performing Zero-shot classification tasks, on NLI purposes, and is particularly Fact-Checking oriented.

It contains:

  • 8 whole datasets in the training split (SNLI, MNLI, FEVER, QNLI, WNLI, SciTail, RTE, VitaminC);
  • 120 examples for the dev split:
    • 60 from the Climate-FEVER dataset in order to test zero-shot knowledge inference in a specific domain;
    • 60 from the ANLI training splits in order to test pure NLI skills of the model.
  • 2 whole datasets in the test split (3 training splits of the ANLI dataset and the whole Climate-FEVER dataset).

Datasets references: