Firstly, a Linear Model is fitted on the whole dataset ... The module depends on NumPy, SciPy and Scikit-Learn (>=0.24.2). Python 3.6 or above is supported.

On TPU mode, loading checkpoints from the file system doesn't supported by the bert and bucket should be used. Fine tuning mongolian BERT on TPU, You need own bucket in order to finetune on TPU Fine ...