MIT Press

Skip to Main Content

Table 6:

The MCQA training datasets used. #Question denotes the number of training samples in our version of each dataset, N the number of choices.

Science Datasets	#Questions	N
OpenBookQA Mihaylov et al., 2018	4,957	4
SciQ Welbl et al., 2017	11,675	4
TextBookQA Kembhavi et al., 2017	7,611	4/5
ARC Dataset++ Clark et al., 2018	4,035	4/5

MCQL Liang et al., 2018	6,318	4

Science Collection (total)	34,596	5

Science Datasets	#Questions	N
OpenBookQA Mihaylov et al., 2018	4,957	4
SciQ Welbl et al., 2017	11,675	4
TextBookQA Kembhavi et al., 2017	7,611	4/5
ARC Dataset++ Clark et al., 2018	4,035	4/5

MCQL Liang et al., 2018	6,318	4

Science Collection (total)	34,596	5