Extend tokens.txt with new tokens on pretrained model
gorosei-dev opened this issue · 1 comments
gorosei-dev commented
Suppose I want to further train the pretrained model on more data, but the new data contains some new tokens that are not covered in the tokens.txt / bpe.model, and I want the new model to be able to recognize these new tokens, how can I achieve this without retraining from scratch?
JinZr commented
You can reuse all parameters of your pre-trained model except for the output layer part, also remember to modify the lang_dir you are using for the later fine-tuning