Fine tune Bert Model
Hello everyone,
I've been using BERT for token classification and fine tuned it to address my problem, which was basically adding some more tags to the already existing list of tags of the model.
The fine tuning worked well and the model is performing pretty good, however, after investigation, I've noticed obvious misses. I want now to fine tune again the version of the BERT model that I already have on a smaller but more targeted dataset specific to the misses that I have noticed.
When going through the exact same process as the first fine-tuning, the model lost all the good prediction that it used to do and turned into rubbish results.
Does anyone has an idea on how to fine tune the BERT model on a smaller dataset ?
Thank you for your help !