trainer.evaluate returns 'epoch' from training

[PhilipMay]

I am training a BERT model: trainer.train()

Then I call evaluate_result = trainer.evaluate(labeled_dataset_test)

The value of evaluate_result looks like this:

{'eval_loss': 0.5908029079437256, 
'eval_acc': 0.8282828282828283, 
'eval_bac': 0.8243021346469622, 
'eval_mcc': 0.7422526698197041, 
'eval_f1_macro': 0.826792009400705, 
'epoch': 3.0, 
'total_flos': 1373653507542624}

IMO the dict should not contain 'epoch': 3.0,. That is the number of epochs from training. It has nothing to do with evaluation...

[sgugger]

You can easily ignore that value though.

The problem is that you won't have it at each eval during the training loop if we don't include it. There could be something smarter there, but it would take time for something that is just purely cosmetic.

[PhilipMay]

I added a small PR for improved documentation about this: #8273

[PhilipMay]

Closing this since PR was merged.