fail to reproduce the base model result of the TriviaQA Dataset with scripts/trivia.py.

hi,
The f1score in the paper is 75.2 but I only got 74.7 by using trivia.py.
And I have a few questions.

1. The triviaQA attention window hyperparameter is not given in the paper, and in the longformer-base-406/config.json, the attention window is all set to 256, Is this correct?

2. The epoch in the paper is set to 5, but the cheatsheet.txt is set to 4.

3. find some bug in trivia.py.
https://github.com/allenai/longformer/blob/0674e0e7bf10007e0dafd7fb65befe96c11bfcb6/scripts/triviaqa.py#L678
num_devices = 1 or len(args.gpus).
In python, 1 or 8 will be equal to 1.

Thanks.