Question about wiki qa dataset

[RenShuhuai-Andy]

I make some analysis on wiki qa dataset：

• training set：
  Left num: 2118； Right num: 18841；Relation num: 20360；positive example (with label 1) num: 1040（5.1%）
• dev set：
  Left num: 296；Right num: 2708；Relation num: 2733；positive example num: 140（5.12%）
• test set：
  Left num: 633；Right num: 5961；Relation num: 6165；positive example num: 293（4.75%）

I wonder if this is the official way to combine question and answer, because the proportion of positive examples in three set is only 5%, which means if a model outputs 0 forever, it can achieve 95% accuracy? And the best performence of BERT on this dataset is just 95%. The proportion of positive and negative examples is too imbalance?