performance F1 lower than 79

[yangjingyi]

Hi,
I use this code and train with spanbert large config, but I only get 78.37 F1 with test dataset.
Do you train with epoch 20 and get F1 80%?

[Fantabulous-J]

I trained the large model with the exact spanbert large config using 5 random seeds. The results range from 79.9 to 80.1. Please make sure you did not change any hyperparameters and do data preprocessing correctly. Also, when extracting the top k spans, please make sure you used the c++ code rather than the topk function in pytorch.