The fin-tuned checkpoint results are not the same as the the ones in paper

[gmftbyGMFTBY]

Hi, first of all, thanks for your wonderful works.

I use the fine-tuned checkpoint on douban corpus to re-implement the performance in your paper.
But the results I got are lower than that in your paper:

It is much lower than the results in the paper

[gmftbyGMFTBY]

Check the dataloader, missing the [EOS] token. BERT-FP works fine. close it.