[Retracted] Analyzing the Effect of Masking Length Distribution of MLM: An Evaluation Framework and Case Study on Chinese MRC Datasets

Table 7

The number of questions of each MRC dataset.

Dataset	Question #	Train question #	Dev question #	Test question #	Percentage of the train set
SQuAD2.0	151,054	130,319	11,873	8862	86.27%
SQuAD1.1	107,702	87,599	10,570	9533	81.33%
TQA	26,260	15,154	5309	5797	57.71%
MovieQA	21,406	14,166	2844	4396	66.18%
MCScript	13,939	9731	1411	2797	69.81%
DREAM	10,197	6116	2040	2041	59.98%
ARC-E	5197	2251	570	2376	43.31%
WikiQA	3047	2118	296	633	69.51%
ARC-C	2590	1119	299	1172	43.20%
ProPara	488	391	54	43	80.12%
Short span	46,473	31,390	6774	8309	67.54%
Long span	20,241	12,053	2376	5812	59.55%
Short cloze	58,500	40,500	9000	9000	69.23%
Long cloze	58,500	40,500	9000	9000	69.23%