mquake-multi-hop-accuracy-metric

Summary

In the MQuAKE benchmark, a multi-hop question set is scored as fully correct even if the model only nails one of the three generated questions, rather than requiring all three to be right. This makes the reported accuracy numbers more generous than a stricter all-or-nothing metric would be, so anyone comparing model performance on this task should keep in mind that the bar for "correct" is relatively low.