GQA generalization splits used in "Visually Grounded VQA by Lattice-based Retrieval", Reich et al.. Splits are based on GQA's balanced data set, downloadable here.
Each file in the .zip contains a list of question IDs for its named subset. Question IDs are either from GQA balanced train set or val set (no mixing).
| Split | Train | Val |
|---|---|---|
| "Objects" | 763k | 23k |
| "LingVariants | 801k | 20k |
| "Answers" | 862k | 11k |
| "LowResource" | 311k | 132k 1 |
Footnotes
-
The regular/original GQA balanced val set, not included for download here, grab from original GQA dataset release. ↩