The Question Relevance Prediction and Explanation (QRPE) Dataset contains contains 53,911 (I+, Q, P, I−) tuples generated from as many premises. In total, it contains 1530 unique premises and 28,853 unique questions. Among the 53,911 premises, 3876 are second-order, attributed object premises while the remaining 50,035 are first-order object/scene premises. We divide our dataset into two parts – a training set with 35,486 tuples that are generated from the VQA training set and a test set with 18,425 tuples generated from the VQA validation set.
Train set (12.2 MB)
Test set (6.3 MB)
Some qualitative examples from the QRPE dataset: