Support for custom BOS, EOS, Sep and Pad tokens?

[ghost]

I have a dataset with special tokens, and when fine tuning with the HF trainer I explicitly add those special tokens to the tokenizer. I looked through the documentation but I didn't see anywhere I could add those in the process. I think it would be a pretty useful feature, but I know it can definitely be very complicated to implement. It seems like every single model has decided to have different BOS and EOS tokens and then there's the whole PAD = EOS debacle.