provides higher GLUE score on bert-base-uncased

According to transformers/ at main 路 huggingface/transformers 路 GitHub, I run for bert-base-uncased. I got similar number as in the table. However, many tasks scores are much higher than the numbers reported in the Bert paper
For example, Matthews corr of CoLA is 56.53 on, and is 57.78 for my finetuned results using, but is only 52.1 in the Bert paper Table1.
Could any explain this? Do I miss someting? Thanks!