GitLab

change layernorm code to pytorch's native layer norm

loads the tokenizer for each checkpoint, to solve the reproducability…

Documentation auto-deploy

Add RoBERTa to the comment as it was not explicit that RoBERTa don't use token_type_ids.

Decode now calls private property instead of public method

swap order of optimizer.step() and scheduler.step()

distilbert: fix number of hidden_size

Generative finetuning

DilBERT