SpanBERT / BERT
Created on October 13|Last edited on October 14
Comment
SpanBERT
9
BERT
9
Trim - SpanBERT LR/ELR Sweep
It seems that 5e-5 is a good learning rate to use for the encoder.
Trim BERT LR/ELR Sweep
We have to do this because we are using trim==100 so we kinda need to know the best performing run here as well.
SpanBERT seems to be much better than BERT. So we're going to do some full dataset experiments with it starting now.
Add a comment