This repository contains pytorch implementation of BERT and DCMN, and the finetuning runners for both models on RACE. This repository also contains reasoning skills proposed for RACE and annotation of 300 randomly sampled test questions.
- pytorch_pretrained_bert contains the Huggingface implementation of BERT. BERT followed by DCMN is added to the implementation.
- run_bert_on_race.py and run_dcmn_on_race.py are the finetune runners.
- run.sh runs the finetune runners.
- race_skills.csv contains a list of reasoning skills required by questions in RACE.
- middle_categorization.csv and high_categorization.csv each contains annotation of 150 questions, and are respectively for middle and high school multiple choice questions.
- Stats.ipynb is a notebook for calculating model test accuracy on questions that require different skills.