./get_started.sh
./split_glove.sh
python preprocess.py
python train.py
My implementation achieve 53.47 F1 score, and 41.425 Exact Match on the SQuAD hidden test set. It falls short of the original authors' scores, but still handily beats a heavily feature-engineered logistic regression implmentation.
You can see the run here on CodaLab.