Could you share the dataset? about ioi HOT 3 CLOSED

Hi, I have updated the readme and fixed the bug in "main.py" . You can prepare the data according to the instructions. Thanks~

from ioi.

Thanks. Now I used your data, although needed to train word2vec myself. However, the result was worse than before (after ~12h of training):

Dev results..........
('pred_scores: ', 500000)
Step: 38000 	| loss: 4.260 	| acc: 0.803 	|  2019-09-20 03:35:54
recall_2_1:  0.830
recall_at_1: 0.541
recall_at_2: 0.683
recall_at_5: 0.867

Maybe the default parameters aren't quite optimal. I'll do some additional testing.

from ioi.

Hi, I provided the link of word2vec file in "download.sh". You may adjust the LR strategy, which has a great impact on the performance of IOI. In our experiment, we set decay_steps = 5000 when batch_size is 20.

from ioi.