Questions about Phase1 training

[klevin-ken]

The training phase1 meets some errors at

and the error message is the same as clovaai/lffont#33

But I am confused how to let the target-generated characters contain all the components in the style characters. During spliting the seen_unis and unseen_unis, I just random split all the characters into train_char.json and val_char.json like the image below:

I think this step may exist some problems. So how do I split seen_unis and unseen_unis ? Thank you !