Thank you very much for the nice dataset !

I have a question about the number of unique answers in the GQA dataset.
When computing the number of unique answers I get:

• 1845 answers for the training split (based on combining each 'answer' of the 10 training files)
• 1852 answers for train + valid
• 1853 for train + valid + test_dev splits.

In the paper, you mention that there are 1878, is this discrepancy caused by some answers only being present in the test split ?

Have a great day :)

Yana