license: mit
Multilingual e-SNLI (MLe-SNLI)
In this repo, we provide the training, validation, and testing sets for Multilingual e-SNLI (MLe-SNLI). For more details, find our report here.
Dataset details
MLe-SNLI contains 500K training (train) samples of premise-hypothesis pairs along with their associated label and explanation. We take 100K training samples from the original e-SNLI (Camburu et al., 2018) dataset and translate them into 4 other languages (Spanish, German, Dutch, and French). We do the same for all 9824 testing (test) and validation (dev) samples, giving us 49120 samples for both test and dev splits.
| Column | Description |
|---|---|
premise |
Natural language premise sentence |
hypothesis |
Natural language hypothesis sentence |
label |
From entailment, contradiction, or neutral |
explanation_1 |
Natural language justification for label |
language |
From English (en), Spanish (es), German (de), Dutch (nl), French (fr) |
WARNING: the translation quality of MLe-SNLI may be compromised for some natural language samples because of quality issues in the original e-SNLI dataset that were not addressed in our work. Use it at your own discretion.
Download Instructions
To access MLe-SNLI, you can use the HuggingFace Datasets API to load the dataset:
from datasets import load_dataset
mle_snli = load_dataset("rish16/MLe-SNLI") # loads a DatasetDict object
train_data = mle_snli['train'] # 500K samples (100K per lang)
dev_data = mle_snli['dev'] # 49120 samples (9824 per lang)
test_data = mle_snli['test'] # 49120 samples (9824 per lang)
print (mle_snli)
"""
DatasetDict({
train: Dataset({
features: ['premise', 'hypothesis', 'label', 'explanation_1', 'language'],
num_rows: 500000
})
test: Dataset({
features: ['premise', 'hypothesis', 'label', 'explanation_1', 'language'],
num_rows: 49120
})
validation: Dataset({
features: ['premise', 'hypothesis', 'label', 'explanation_1', 'language'],
num_rows: 49210
})
})
"""