Model's input_ids and labels

[XQZZK]

Hey! This is a very exciting piece of work. But I have a question about the implementation of the code. In train_gptj_toolformer.py, Why are the inputs to the model the same as the labels that need to be predicted ??? Such a objective seems to be a bit different from the objective——a standard language modeling objective, proposed in the original paper. If you're free, Would you mind elaborating on the issue a bit more? I want to make sure I understand it thoroughly.