Open HAWLYQ opened 5 years ago
Yes, it would be great to have that code.
I have the same issue, that my code only achieves a performance of about F1: 90.0 (test set).
I haven't tried on the test set but I've also only gotten around 93 F1 on the dev set while the advertised performance is higher than 96 F1. For a start, reporting the hyperparameters that this experiment was run with would be helpful.
Hi, did you finally reproduce the results (with document context)?
Could you share your fine_tune codes for NER CoNLL2003 with bert_base(cased) model? I met some problems to redo the reported experiment result ( f1 - 92. 4) .