Reproducing the result for STS-B

[cooelf]

Hi, thanks for your codes! I am struggling to reproduce the results for STS-B but the results are far from the reported one in the BERT paper with lots of runs.
The following is my best result (still low) along with the corresponding hyper-parameters.

pearson:  [dev] 90.42%    [test] 83.0%
Spearman: [dev] 90.04%    [test] 81.2%
learning rate: 2e-5
number of epochs: 5
max_seq_len: 128

Have you reproduced the results and can you share some hints or settings?