Fill-Mask • 0.1B • Updated • 3.22M • • 362
BERT release
updated
Regroups the original BERT models released by the Google team. Except for the models marked otherwise, the checkpoints support English.
Note Base BERT model, smaller variant. Trained on the "cased" dataset, meaning that it wasn't lowercase and all accents were kept. 12-layer, 768-hidden, 12-heads , 110M parameters
Fill-Mask • 0.1B • Updated • 59.9M • • 2.69kNote Base BERT model, smaller variant. Trained on the "uncased" dataset, meaning that it was lowercase and all accents were removed. 12-layer, 768-hidden, 12-heads , 110M parameters
Fill-Mask • 0.3B • Updated • 101k • • 38Note Large BERT model, larger variant. Trained on the "cased" dataset, meaning that it wasn't lowercase and all accents were kept. 24-layer, 1024-hidden, 16-heads, 340M parameters
Fill-Mask • 0.3B • Updated • 2.95M • • 147Note Large BERT model, larger variant. Trained on the "uncased" dataset, meaning that it was lowercase and all accents were removed. 24-layer, 1024-hidden, 16-heads, 340M parameters
Fill-Mask • 0.2B • Updated • 3.66M • • 593Note Base BERT model, smaller variant. The list of supported languages is available here: https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages 104 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
Fill-Mask • 0.1B • Updated • 529k • • 1.44kNote Base BERT model, smaller variant. Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters
Fill-Mask • 0.3B • Updated • 3.89k • • 24Note Large BERT model, larger variant. Trained on the "cased" dataset, meaning that it wasn't lowercase and all accents were kept. Whole word masking indicates a different preprocessing where entire words are masked rather than subwords. The BERT team reports better metrics with the wwm models. 24-layer, 1024-hidden, 16-heads, 340M parameters
Fill-Mask • 0.3B • Updated • 12.1k • • 23Note Large BERT model, larger variant. Trained on the "uncased" dataset, meaning that it was lowercase and all accents were removed. Whole word masking indicates a different preprocessing where entire words are masked rather than subwords. The BERT team reports better metrics with the wwm models. 24-layer, 1024-hidden, 16-heads, 340M parameters