The accuracies of zero-shot tasks can not be reproduced

[NewDriverLee]

The perplexity results can be exactly reproduced by my experiments. However, the accuracies for zero-shot tasks differs from your reported results, sometimes better sometimes worse. A more interesting thing is that when I use a new server to run this code, the accuracies changed. Do you have any idea about this phenomenon?