datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
ai2thor-path-tracing-qa-train-mmcot-20k-v5cocscbsdgchcai2thor-path-tracing-qa-train-2point-balanced8-mmcot-16kspatial-mmcot-cova
Spatial MMCoT v1 · cova
CoVA-SFT (tsunghanwu/cova), 4 of its spatial configs: 3d_cube, arc, maze, spatialeval. The fifth, seating_arrangement, is excluded: a brute-force solve of its puzzles disagrees with many of its upstream labels, and many of its read-backs name a different person from the answer they state. The upstream problems are posed in text and every image is drawn by the trace, so upstream has no input image. This format needs at least one, so the first drawn image… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-cova.spatial-mmcot-ipt_mvc
Spatial MMCoT v1 · ipt_mvc
Multi-view counting in ProcTHOR houses (IPT MVC-ipt). Four or five ego frames along a trajectory are the input; the target is the top-down map of the explored area. The reasoning text comes from the text chain-of-thought donor named below, which saw (at most four of) the frames but never the map: its plan lists the new objects frame by frame, usually without giving the total, so on most rows the answer can be added up from the plan before the map is… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-ipt_mvc.spatial-mmcot-ipt_pet
Spatial MMCoT v1 · ipt_pet
Perspective taking in ProcTHOR houses (IPT PET-ipt). The input is the current view with an X marked on the floor; the target is the view after moving to the X and turning 90 degrees; the question asks whether a named object gets closer or further, or lies left or right. The reasoning text comes from the text chain-of-thought donor named below, joined by row order (its question and options equal ours on every upstream row, apart from one blank line… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-ipt_pet.spatial-mmcot-ipt_pt
Spatial MMCoT v1 · ipt_pt
Path tracing in ProcTHOR houses (IPT PT-ipt). The input is a top-down map with the route (points 1 and 2, midpoint M1) drawn on it, plus the forward views from points 1 and 2. The target is the view from waypoint M1. The question asks which of four objects you can see on your left or right at M1. The upstream think text is one of 16 fixed phrases (8 for the left side, 8 for the right) such as 'Looking right from midpoint M1:', so the reasoning is taken… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-ipt_pt.spatial-mmcot-messytable
Spatial MMCoT v1 · messytable
Multi-camera tabletop counting on real photographs from MessyTable (Z. Cai, J. Zhang, D. Ren, C. Yu, H. Zhao, S. Yi, C. K. Yeo and C. C. Loy, "MessyTable: Instance Association in Multiple Camera Views", ECCV 2020; https://github.com/caizhongang/MessyTable). The questions, counts and reasoning text come from leo66666/messytable, which pairs those photographs with a counting chain of thought. Several angled views of one cluttered table are the input;… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-messytable.spatial-mmcot-motif
Spatial MMCoT v1 · motif
MoTiF (OpenRaiser/MoTiF), the naive arm of its procedurally generated multi-step tasks: ball_tracking_naive, manipulation_naive, maze_naive (manipulation shows MoTiF's own CLEVR-style 3D renders of solids, not images from the CLEVR dataset). Each step has its own target image. The reflexion arm, which pairs each problem with a deliberately corrupted frame, is not included. sokoban_naive is excluded (7,997 upstream rows, S0.task_excluded): its eight… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-motif.spatial-mmcot-thinkmorph_jigsaw
Spatial MMCoT v1 · thinkmorph_jigsaw
ThinkMorph (arXiv:2510.27492) Jigsaw_Assembly: a picture cut into numbered parts, shown in an order that may be shuffled; decide the arrangement, draw the assembled picture, read it back, answer. According to the ThinkMorph paper (arXiv:2510.27492, appendix on data generation), GPT-4.1 wrote the plan and the read-back from the question and the ground-truth answer, and was told not to reveal the answer. The pictures come from 3 corpora: 3,203… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-thinkmorph_jigsaw.spatial-mmcot-thinkmorph_spatial_nav
Spatial MMCoT v1 · thinkmorph_spatial_nav
ThinkMorph (arXiv:2510.27492) Spatial_Navigation: FrozenLake navigation on 3x3 to 6x6 grids. The one input image is the grid, and the question text is the same on every row (upstream's wording, unchanged: it asks for the moves in \boxed{{}}, an unformatted template left in upstream; the read-back boxes the moves while <answer> holds them bare), so the maze exists only in the image. The first thought describes the grid (start, goal… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-thinkmorph_spatial_nav.spatial-mmcot-vdrop
Spatial MMCoT v1 · vdrop
VDrop cross-view (QianYangMILA/vdrop-crossview-8k, arXiv:2605.27310), Infinigen Indoors rooms. Two egocentric views of one room with partly overlapping fields of view are the input (input_image_0 = cam0, input_image_1 = cam1); the target is upstream's ground-truth panorama render, a wide panorama of the room that spans both views. The questions are multiple choice and the trained answer is the option letter. The upstream release ships no reasoning text… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-vdrop.spatial-mmcot-zebra_jigsaw
Spatial MMCoT v1 · zebra_jigsaw
Zebra-CoT visual jigsaw, type-2 (single-image) rows only, on ImageNet images (mostly photographs; some are web graphics such as banners and flyer templates). Each row has one input image: the source image with its missing piece(s) greyed out, above a panel of four candidate piece sets labelled A-D. The options exist only in that image, so the answer is the letter. The target is the complete source image. Type-1 rows are excluded: three quarters of… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-zebra_jigsaw.spatial-mmcot-zebra_multihop
Spatial MMCoT v1 · zebra_multihop
Zebra-CoT multi-hop object counting over rendered 3D scenes (primitive objects on textured ground under a sky). Many traces open with a viewpoint change, so the first target is often a novel view. The final thought re-examines the last target image, but on some rows it only says it counts the objects and never states the number; the count is then only in <answer>. Released rows carry 2 to 5 target images. 2,057 traces with more target images… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-zebra_multihop.spatial-mmcot-zebra_tetris
Spatial MMCoT v1 · zebra_tetris
Zebra-CoT Tetris: polyomino puzzles of three kinds: apply a sequence of transformations to a shape, tile a shape with a set of pieces, and fill the grid outside a shape. The options are drawn in the input image, so the answer is a letter. Each step has its own target image: transformation puzzles start with a redraw of the start shape, then one image per transformation; tiling puzzles start with the isolated shape, fill-the-complement puzzles with… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-zebra_tetris.