zhuchen03 / modifying-memory Goto Github PK

Hi @zhuchen03

Thanks for releasing the code. I was running the pre-preprocessing script for zsRE and got 187,509/69,850 questions for train/eval and 147,909/47,156 facts for train/eval. I saw in the paper it was mentioned there are 197,829/59,527 questions for train/eval and 147,905/46,156 facts for train/eval.

I am guessing the number of facts for training was a typo since KILT has 147,909 facts for training. However, I am not sure about the difference in the number of questions for train/eval. I understand that randomness might cause the actual questions to be different but I thought the number of questions should be the same since there is consistent rule to determine how many questions are split into train/eval.