These numbers are familiar - 30522 is the default for an ELECTRA model. I have a code example which works using TF2, Transformers, and the ktrain library ( https://colab.research.google.com/drive/1mSeeSfVSOT7e-dVhPlmSsQRvpn6xC05w )

Is ktrain throwing out some values from the tokenizer, or is this a difference in TF2 / PyTorch models? I have written down a note to fix this, but please let me know if you can make progress