When to stop XLNet finetuning?

[reobroqn]

Hello!
Glad that I find your code. It helps me a lot.
But there is one thing I am not clear yet is when to stop XLNet finetuning.
When I try with BERT they have an evaluation phase to give the accuracy but with XLNet it doesn't.
Why did you stop at 60 train steps? Or it's just pure experiments, try if it works then try another train steps? I couldn't get this point.
Would be appreciate if you can share me something about this.
Thanks for reading!