Cannot reproduce the best results from the provided pretrained model

[WinterShiver]

I have successfully loaded the pretrained weights from e2e-coref.tgz and get 72.39 on the test set. According to your NAACL18 paper paper, this corresponds to "e2e + ELMo + hyperparameter tuning"? I am wondering that how to reproduce the results of "+ c2f inference" and "+ second-order inference"? Could you please provide the corresponding configures of these two settings? Waiting for your apply.