training BERT for relation classfication using "matching the blanks" recommanded method

overview of repo files:

• datautils.py data loading primitives
• simplifying_wrapper.py simple wrapping methods to full finetuning code
• ftbert.ipynb a notebook implementing a dataloader for our spike jsons and then training roberta with it
• rc_transformer.py a pytorch model - matching the blanks-like RC with a transformer
• transformers_rc_finetune.py - all the training boilerplate together with some sane training settings