McGill-NLP/GlobalNLI

Dataset Card for global_nli

Dataset Summary

Global NLI is a new text-based benchmark based on the aggregation of existing NLI datasets that are publicly available.

Languages

There are 59 languages available :

{ 'amh': 'Amharic', 'ara': 'Arabic', 'asm': 'Assamese', 'aym': 'Aymara', 'ben': 'Bengali', 'bul': 'Bulgarian', 'bzd': 'Bribri', 'cat': 'Catalan', 'cni': 'Asháninka', 'deu': 'German', 'ell': 'Greek', 'eng': 'English', 'ewe': 'Ewe', 'fra': 'French', 'grn': 'Guarani', 'guj': 'Gujarati', 'hau': 'Hausa', 'hch': 'Wixarika', 'hin': 'Hindi', 'ibo': 'Igbo', 'ind': 'Indonesian', 'jpn': 'Japanese', 'kan': 'Kannada', 'kin': 'Kinyarwanda', 'kor': 'Korean', 'lin': 'Lingala', 'lug': 'Luganda', 'mal': 'Malayalam', 'mar': 'Marathi', 'mya': 'Burmese', 'nah': 'Nahuatl', 'ori': 'Odia (Oriya)', 'orm': 'Oromo', 'oto': 'Otomi', 'pan': 'Punjabi', 'pat': 'Jamaican Patois', 'pol': 'Polish', 'por': 'Portuguese', 'quy': 'Quechua', 'ron': 'Romanian', 'rus': 'Russian', 'shp': 'Shipibo-Conibo', 'sna': 'chiShona', 'sot': 'Sesotho', 'spa': 'Spanish', 'swa': 'Swahili', 'tam': 'Tamil', 'tar': 'Rarámuri', 'tel': 'Telugu', 'tha': 'Thai', 'tur': 'Turkish', 'twi': 'Twi', 'urd': 'Urdu', 'vie': 'Vietnamese', 'wol': 'Wolof', 'xho': 'isiXhosa', 'yor': 'Yoruba', 'zho': 'Chinese', 'zul': 'isiZulu' }

Dataset Structure

Data Instances

The examples look like this for English:

from datasets import load_dataset
data = load_dataset('McGill-NLP/GlobalNLI', 'eng') 
# Please, specify the language code
# A data point example is below:
{
'premise': 'The doors were locked when we went in.',
'hypothesis': 'All of the doors were open.',
'label': 0
}

Data Fields

• premise: a multilingual string variable,
• hypothesis: a multilingual string variable,
• label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

Data Splits

All languages have two splits, dev and test set.

The splits have the following sizes :

Language	validation	test
English	30	600