Universal Latent Visual Reasoning training data, cleaned. 8 categories (subsets); each has train + validation splits. Every sample: input image + question -> assistant produces <abs_vis_token> + intermediate visual step(s) + \boxed{answer}.

subset train validation

text_cot 333,911 3,533

bbox_highlight 229,237 2,558

bbox_crop 229,237 2,558

depth 40,000 25

edge 40,000 14

segmentation 40,000 326

helper_interleaved 340,210 3,544

scene_graph 40… See the full description on the dataset page: https://huggingface.co/datasets/RuoliuYang/ULVR_v2_clean.