google-research/bert ? reverse-engineered prompt

Build me a TensorFlow BERT project that can both pretrain and fine tune a BERT model on text data.

I want to be able to create pretraining data, train on masked language modeling and next sentence prediction, then reuse the same model for common NLP tasks like text classification, question answering, and feature extraction. It should also include a simple tokenizer and utilities for turning raw text into model inputs.

Please make it work with the provided sample text and include a few basic tests so I can tell the core pieces are wired up correctly. If anything in the docs is unclear, look up the current TensorFlow guidance online and follow the standard BERT workflow.