We evaluate the predictions on Precision, Recall, and F1 -- just as what is done in the eval.py script released. On the Kaggle leaderboard, F1 is reported.