CATIE-AQ/frenchNLI

Dataset information

Dataset concatenating NLI datasets available in French and open-source.
There are a total of 569,895 rows, of which 555,798 are for training, 3,780 for validation and 10,317 for testing.

Usage

from datasets import load_dataset
dataset = load_dataset("CATIE-AQ/frenchNLI")

Dataset

Details of rows

Dataset Original	Splits	Note
xnli	392,574 train / 2,490 validation / 5,010 test	We only keep the French split
MoritzLaurer/multilingual-NLI-26lang-2mil7 (fr_anli)	24,973 train	Data from machine translation
MoritzLaurer/multilingual-NLI-26lang-2mil7 (fr_fever)	24,940 train	Data from machine translation
MoritzLaurer/multilingual-NLI-26lang-2mil7 (fr_ling)	5,000 train	Data from machine translation
MoritzLaurer/multilingual-NLI-26lang-2mil7 (fr_mnli)	24,999 train	Data from machine translation
MoritzLaurer/multilingual-NLI-26lang-2mil7 (fr_wanli)	24,996 train	Data from machine translation
Lajavaness/SICK-fr	4,404 train / 495 validation / 4,511 test	Data from machine translation
ssilwal/CASS-civile-nli	53,140 train
maximoss/rte3-french	800 validation / 800 test
maximoss/daccord-contradictions	1,034 train

Removing duplicate data and leaks

The sum of the values of the datasets listed here gives the following result:

DatasetDict({
    train: Dataset({
        features: ['premise', 'hypothesis', 'label', 'dataset'],
        num_rows: 556060
    })
    validation: Dataset({
        features: ['premise', 'hypothesis', 'label', 'dataset'],
        num_rows: 3785
    })
    test: Dataset({
        features: ['premise', 'hypothesis', 'label', 'dataset'],
        num_rows: 10321
    })
})

However, a data item in training split A may not be in A's test split, but may be present in B's test set, creating a leak when we create the A+B dataset.
The same logic applies to duplicate data. So we need to make sure we remove them.
After our clean-up, we finally have the following numbers:

DatasetDict({
    train: Dataset({
        features: ['premise', 'hypothesis', 'label', 'dataset'],
        num_rows: 555798
    })
    validation: Dataset({
        features: ['premise', 'hypothesis', 'label', 'dataset'],
        num_rows: 3780
    })
    test: Dataset({
        features: ['premise', 'hypothesis', 'label', 'dataset'],
        num_rows: 10317
    })
})

Columns

dataset_train = dataset['train'].to_pandas()
dataset_train.head()

premise 	                                             hypothesis 	                                        label 	dataset
0 	L'écrémage conceptuel de la crème a deux dimen... 	Le produit et la géographie sont ce qui fait t...   1 	    xnli
1 	Tu sais pendant la saison et je suppose qu'à t... 	Vous perdez les choses au niveau suivant si le...   0 	    xnli
2 	Un de nos numéros vous fera suivre vos instruc... 	Un membre de mon équipe exécutera vos ordres a...   0 	    xnli
3 	Qu'est-ce que tu en sais ? Tout ceci est à nou... 	Cette information leur appartient.                  0 	    xnli
4 	Ouais je te dis ce que si tu vas prix certaine... 	Les chaussures de tennis ont une gamme de prix.     1 	    xnli

• the premise column contains the premise text
• the hypothesis column contains the hypothesis text
• the label column contains the label (0 → implication, 1 → neutral, 2 → contradiction)
• the dataset column identifies the row's original dataset (if you wish to apply filters to it)

Split

• train corresponds to the concatenation of xnli + fr_anli + fr_fever + fr_ling + fr_mnli + fr_wanli + SICK-fr + CASS-civile-nli
• validation corresponds to the concatenation of xnli + SICK-fr
• test corresponds to the concatenation of xnli + SICK-fr

Citations

Lajavaness/SICK-fr

Dataset by Lajavaness (2023)
https://huggingface.co/datasets/Lajavaness/SICK-fr by Lajavaness (2023)

maximoss/daccord-contradictions

Maximos Skandalis, Richard Moot, and Simon Robillard. 2023. DACCORD : un jeu de données pour la Détection Automatique d’énonCés COntRaDictoires en français. In Actes de CORIA-TALN 2023. Actes de la 30e Conférence sur le Traitement Automatique des Langues Naturelles (TALN), volume 1 : travaux de recherche originaux -- articles longs, pages 285–297, Paris, France. ATALA.

maximoss/rte3-french

Dataset by Maximos Skandalis (2023)
https://huggingface.co/datasets/maximoss/rte3-french (2023)

MoritzLaurer/multilingual-NLI-26lang-2mil7

@article{laurer_less_2022,
    title = {Less {Annotating}, {More} {Classifying} – {Addressing} the {Data} {Scarcity} {Issue} of {Supervised} {Machine} {Learning} with {Deep} {Transfer} {Learning} and {BERT} - {NLI}},  
    url = {https://osf.io/74b8k},  
    language = {en-us},  
    urldate = {2022-07-28},  
    journal = {Preprint},  
    author = {Laurer, Moritz and Atteveldt, Wouter van and Casas, Andreu Salleras and Welbers, Kasper},  
    month = jun,  
    year = {2022},  
    note = {Publisher: Open Science Framework},

ssilwal/CASS-civile-nli

Dataset by Silviya Silwal (2023)
https://huggingface.co/datasets/ssilwal/CASS-civile-nli by Silviya Silwal (2023)

XNLI

@InProceedings{conneau2018xnli,  
  author = {Conneau, Alexis and Rinott, Ruty and Lample, Guillaume and Williams, Adina and Bowman, Samuel R. and Schwenk, Holger and Stoyanov, Veselin},  
  title = {XNLI: Evaluating Cross-lingual Sentence Representations},  
  booktitle = {Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},  
  year = {2018},  
  publisher = {Association for Computational Linguistics},  
  location = {Brussels, Belgium},}

FrenchNLI

@misc {frenchSTS_2025,   
    author       = { {BOURDOIS, Loïck} },  
    organization  = { {Centre Aquitain des Technologies de l'Information et Electroniques} },  
    title        = { frenchNLI (Revision 475d4f9) },  
    year         = 2025,  
    url          = { https://huggingface.co/datasets/CATIE-AQ/frenchNLI },  
    doi          = { 10.57967/hf/5943 },  
    publisher    = { Hugging Face } }

License

cc-by-4.0