Task1 is not learned well

[jojonki]

This is a really nice repository and I really appreciate it!

Although I have tried task1, the accuracy didn't increase well. The accuracy is around 21.4% after 60 epochs. How did you obtain almost 100% accuracy in your environment?

python train.py --train data/train-task-1.tsv --dev data/dev-task-1-500.tsv --vocab data/vocab-task-1.tsv --save_dir checkpoints/task-1/model
...
[INFO] 2018-06-05T15:41:36-0400: Evaluation: (0, 0, 0) (train.py:105)
[DEBUG] 2018-06-05T15:41:36-0400: Saving checkpoint (train.py:108)
[INFO] 2018-06-05T15:41:55-0400: Epoch: 61; Train loss: 0.03631829832971794; Dev loss: 0; (train.py:101)
[INFO] 2018-06-05T15:42:17-0400: Epoch: 62; Train loss: 0.035734415055687974; Dev loss: 0; (train.py:101)
0it [00:00, ?it/s]
[INFO] 2018-06-05T15:42:17-0400: Evaluation: (0, 0, 0) (train.py:105)
[DEBUG] 2018-06-05T15:42:17-0400: Saving checkpoint (train.py:108)

python test.py --test data/test-task-1.tsv --candidates data/candidates.tsv \
  --vocab data/vocab-task-1.tsv --checkpoint_dir checkpoints/task-1 --emb_dim 32 
(1274, 4662, 0.21462264150943397)
// (pos, neg, pos/(pos+neg))