CATIE-AQ/frenchQA

Dataset information

Dataset concatenating QA datasets with context available in French and open-source.
In addition, an augmented version of these datasets has been added (same context but different questions to create data in SQuAD 2.0 format).
In total, there are 221,348 training data, 910 validation data and 6,376 test data.
In practice, due to the restrictive license for the FQUAD 1.0 dataset, we can only share 200,617 rows of the 221,348 training data and 3,188 rows of the 6,376 test data.
So to obtain the complete dataset, the user will have to concatenate this dataset with the fquad dataset available here.
Our methodology is described in a blog post available in English or French.

Usage

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

dataset
DatasetDict({
    train: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers'],
        num_rows: 200617
    })
    validation: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers'],
        num_rows: 910
    })
    test: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers'],
        num_rows: 3188
    })
})

Dataset

Dataset details

Dataset	Format	Train split	Dev split	Test split	Available in frenchQA
piaf	SQuAD 1.0	9 224 Q & A	X	X	Yes
piaf_v2	SQuAD 2.0	9 224 Q & A	X	X	Yes
fquad	SQuAD 1.0	20 731 Q & A	3 188 Q & A (is not used for training, but as a test dataset)	2 189 Q & A (not freely available)	No due to the license
fquad_v2	SQuAD 2.0	20 731 Q & A	3 188 Q & A (is not used for training, but as a test dataset)	X	Yes
lincoln/newsquadfr	SQuAD 1.0	1 650 Q & A	455 Q & A	X	Yes
lincoln/newsquadfr_v2	SQuAD 2.0	1 650 Q & A	455 Q & A	X	Yes
pragnakalp/squad_v2_french_translated	SQuAD 2.0	79 069 Q & A	X	X	Yes
pragnakalp/squad_v2_french_translated_v2	SQuAD 2.0	79 069 Q & A	X	X	Yes

Split

• train corresponds to the concatenation of the training dataset from pragnakalp/squad_v2_english_translated + lincoln/newsquadfr + PIAFv1.2 + the augmented version of each dataset in SQuADv2 format (no shuffle has been performed)
• validation corresponds to the concatenation of the newsquadfr validation dataset + this same dataset expanded in SQuAD v2 format (= newsquadfr_v2) (no shuffle performed)
• test corresponds to the concatenation of the fquad dataset SQuAD v1 in SQuAD v2 format (here we can only share the SQuAD v2 format)

Question type statistics

The question type distribution is as follows:

Type of question	Frequency in percent
What (que)	55.02
Who (qui)	15.96
How much (combien)	7.92
When (quand)	6.90
Where (où)	3.15
How (comment)	3.76
What (quoi)	2.60
Why (pourquoi)	1.25
Other	3.44

The number of questions containing a negation, e.g. "What was the name of Chopin's first music teacher who was not an amateur musician?", is estimated at 3.55% of the total questions.

For information, the distribution of the complete dataset (containing FQUAD 1.0 and FQUAD 1.0 data in SQUAD 2.0 format) is as follows:

Type of question	Frequency in percent
What (que)	55.12
Who (qui)	16.24
How much (combien)	7.56
When (quand)	6.85
Where (où)	3.98
How (comment)	3.76
What (quoi)	2.94
Why (pourquoi)	1.41
Other	2.14

The number of questions containing a negation, e.g. "What was the name of Chopin's first music teacher who was not an amateur musician?", is estimated at 3.07% of the total questions.

Citation

@misc {frenchQA2023,  
    author       = { {ALBAR, Boris and BEDU, Pierre and BOURDOIS, Loïck} },  
    organization  = { {Centre Aquitain des Technologies de l'Information et Electroniques} },  
    title        = { frenchQA (Revision 6249cd5) },  
    year         = 2023,  
    url          = { https://huggingface.co/CATIE-AQ/frenchQA },  
    doi          = { 10.57967/hf/0862 },  
    publisher    = { Hugging Face }  
}

License

cc-by-4.0