BPE based tokenizer used for the MEHDIE project and the training of a bilingual BERT model.
Vocab size of 52000.