How can I reproduce the results in SuperGlue paper?

[runzeer]

I used the training parameters according to the supoerglue paper. I used the bert-large-cased model weights and finetuned the BoolQ dataset. The paper said the dev phase results is 77.7. But I can only get 74.2. The training parameters are listed as below:

Batchsize: 16 lr: 2e-5 epoch:10 optimizer:Adam

So if convenient, could you share your training parameters?

[zphang]

The results for SuperGLUE were computed using a legacy version of the jiant repository, found here: https://github.com/nyu-mll/jiant-v1-legacy