Answering questions about images (VQA, GQA).
265K images with 1.1M questions. Balanced dataset to reduce language biases found in v1.