--- license: other license_name: per-source pretty_name: "Spatial MMCoT v2" language: - en task_categories: - visual-question-answering configs: - config_name: jigsaw default: true data_files: - split: train path: jigsaw/preview/train/*.parquet - split: validation path: jigsaw/preview/validation/*.parquet - config_name: multiview_counting data_files: - split: train path: multiview_counting/preview/train/*.parquet - split: validation path: multiview_counting/preview/validation/*.parquet - config_name: perspective_taking data_files: - split: train path: perspective_taking/preview/train/*.parquet - split: validation path: perspective_taking/preview/validation/*.parquet - config_name: cross_view data_files: - split: train path: cross_view/preview/train/*.parquet - split: validation path: cross_view/preview/validation/*.parquet - config_name: state_tracking data_files: - split: train path: state_tracking/preview/train/*.parquet - split: validation path: state_tracking/preview/validation/*.parquet - config_name: navigation data_files: - split: train path: navigation/preview/train/*.parquet - split: validation path: navigation/preview/validation/*.parquet --- # Spatial MMCoT v2 Interleaved image-text chain-of-thought for spatial reasoning. Each row is a question with one or more input images, a written plan, one or more **helper pictures** that a model is trained to draw, a reading of each picture, and the answer. 63,871 rows (train 62,136 / validation 1,735) from 8 sources, grouped by the kind of task into the 6 subsets below -- choose one in the viewer. The validation split is held out by scene. The questions, options, answers and helper pictures come from the upstream sources (see Licences); the `messytable_relabel` questions and answers were derived again from the original MessyTable annotations. Every thought was written by a model: Qwen3.8-27B in thinking mode (temperature 1.0, top-p 0.95, top-k 20) was shown the row's input images, its helper pictures in order, the question, the options and the answer, and wrote the plan before each helper (what is still unknown, and what the picture should show to settle it) and the reading after the last one. The text is therefore written toward the stored answer, and it can misdescribe a picture (see the known limits below). ## Subsets | subset | the questions | the helper picture | rows from (tasks) | train | validation | |---|---|---|---|---:|---:| | `jigsaw` | Jigsaw reassembly: which candidate pieces fill the grey holes of a photo, or in which order shuffled strips of a photo belong | the completed / reassembled photo | `zebra_jigsaw`, `thinkmorph_jigsaw` | 16,185 | 458 | | `multiview_counting` | Multi-view counting: how many objects of a kind are on a cluttered table seen in 2-7 photos, or which kind is the most numerous | the real overhead photo of the same table | `messytable_relabel` | 6,334 | 107 | | `perspective_taking` | Perspective taking: after walking to a marked X and turning 90 degrees, is an object closer / further, on the left / right | the rendered view from the X after the turn | `ipt_pet` | 8,393 | 211 | | `cross_view` | Cross-view reasoning: objects seen from two spots of one room: which one shows in both, which is nearest to a named object, how many | a wide panorama of the room (it does not wrap around) | `vdrop` | 3,990 | 166 | | `state_tracking` | Multi-step state tracking: where a bouncing ball exits, what a scene of solids looks like after a sequence of edits, how many remain | the trajectory or scene after each step, one picture per step | `motif` (ball_tracking_naive, manipulation_naive), `zebra_multihop` | 14,632 | 445 | | `navigation` | Navigation and route planning: the move sequence across a frozen lake avoiding holes, or the number of turns on the route through a maze | the grid with the route drawn | `thinkmorph_spatial_nav`, `motif` (maze_naive) | 12,602 | 348 | ## What a row contains (the viewer) Columns from left to right: `source` and `task` (the upstream source and task), `question` and `options` as the model receives them (the system prompt is in the training files), `answer`, `input_image_0..` (the inputs), then the chain in the order the model writes it -- `thought_0` (the plan), `target_image_0` (the helper picture), `thought_1`, `target_image_1`, ... -- and the last `thought_` column (the reading of the last helper and the conclusion). Image and thought slots past the row's own are null. Then `row_uid`, `category`, `num_input_images` and `num_steps` (K, the number of helpers), `split`, `license` (the row's licence), `source_scene_corpus`, `est_tokens` and `filter_flags` (the flags below). ## How v2 differs from v1 v2 is v1 ([yrlyrl/spatial-mmcot](https://huggingface.co/datasets/yrlyrl/spatial-mmcot), one repository per source, same images and text) after an audit of the training mixture (a stratified review of 270 rows and mechanical checks over every row): - `ipt_mvc` is removed (counts that miss visible paintings, hidden type conventions, options that give the answer, a top-down map that cannot show what is counted). - the cross-view questions that ask a direction -- from the viewpoint of one photo (`perspective_taking`) or between two objects (`relative_direction`) -- are removed (`vdrop_helper_misleads_direction` below): their helper panorama is taken from a third spot, and answering with it made a model worse than answering without it (Qwen3.8-27B with thinking, the same 29 sampled questions with and without the panorama: -21 and -13 points). - rows removed, by reason: | reason | rows | |---|---:| | vdrop_helper_misleads_direction | 2,616 | | motif_answer_in_operations | 741 | | multihop_unstated_removal | 562 | | multihop_text_determined | 280 | | validation_maze_near_twin_in_train | 46 | | vdrop_broken_panorama | 35 | | readback_claims_path_fails | 30 | | contradicting_answers_same_images | 15 | | plan_misplaces_start | 4 | | text_states_other_arrangement | 3 | | plan_turn_rule_backwards | 1 | | validation_target_near_copy_of_train | 1 | | question_states_answer | 1 | - repaired in place: the frozen-lake instruction's `\boxed{{}}` (now `\boxed{}`); ProcTHOR one-word type names in the perspective-taking questions and texts (`diningtable` -> `dining table`); "360-degree panorama" wording in the cross-view texts. - easy or redundant rows are kept and **flagged** in `filter_flags`, so a training run can choose its subset: | flag | meaning | rows | |---|---|---| | `A.vdrop_pair_n=` | rows that share this row's two input views (and so its panorama) | cross_view 4,156 | | `A.nav_moves=` | number of moves in a frozen-lake answer | navigation 5,727 | | `A.manip_count` | a manipulation question that asks for a count (object positions do not matter) | state_tracking 1,782 | | `A.vdrop_wrap_claim` | the text says the panorama wraps around (it does not) | cross_view 464 | | `A.multihop_untouched` | no operation can change the counted class (the answer is the initial count) | state_tracking 574 | | `A.repeat_target` | a helper repeats an earlier helper of the row (a no-op step) | state_tracking 564 | | `A.text_issue=` | a text-only issue found by a rule (the label and pictures are fine) | jigsaw 261, state_tracking 5 | | `A.options_give_answer` | the four consecutive numeric options alone give the answer | state_tracking 146 | | `A.shared_target` | the same photo is the target of another puzzle row | jigsaw 28 | Known limits (measured in the audit, not fixed): about half of the rows have at least one statement in the text that does not match the pictures (wrong object or shape names, miscounts of incidental objects), while answers and helpers are almost always right; the jigsaw decoy pieces are crops of the visible photo; several upstream panoramas are taken from a third spot rather than from either input camera. ## Training files Each subset directory also holds the training format: `/train/*.parquet` and `/validation/*.parquet` (`image_list` list: the inputs then the helpers; `num_input_images`; `instruction_list` with the system prompt, question and options; `output_text_list` with the K+1 segments `..` ... `....`; `row_uid`), `/meta/` (the per-row sidecar, same order) and `/parquet_info.json` (keys `//`). The three formats are aligned file by file and row by row. ```python from datasets import load_dataset rows = load_dataset("yrlyrl/spatial-mmcot-v2", "perspective_taking", split="train") # the viewer's rows from huggingface_hub import snapshot_download # the training format root = snapshot_download("yrlyrl/spatial-mmcot-v2", repo_type="dataset", allow_patterns=["navigation/train/*", "navigation/validation/*", "navigation/parquet_info.json"]) # BAGEL / ThinkMorph trainer: run from `root` with data_dir: navigation/train, parquet_info_path: navigation/parquet_info.json ``` ## Licences Each row keeps the licence of its upstream source (the `license` column): | source | upstream | licence | |---|---|---| | `ipt_pet` | [weikaih/imaginative-perception-token-pet-ipt](https://huggingface.co/datasets/weikaih/imaginative-perception-token-pet-ipt) | undeclared | | `messytable_relabel` | [leo66666/messytable](https://huggingface.co/datasets/leo66666/messytable) | undeclared | | `motif` | [OpenRaiser/MoTiF](https://huggingface.co/datasets/OpenRaiser/MoTiF) | mit | | `thinkmorph_jigsaw` | [ThinkMorph/Jigsaw_Assembly](https://huggingface.co/datasets/ThinkMorph/Jigsaw_Assembly) | undeclared | | `thinkmorph_spatial_nav` | [ThinkMorph/Spatial_Navigation](https://huggingface.co/datasets/ThinkMorph/Spatial_Navigation) | undeclared | | `vdrop` | [QianYangMILA/vdrop-crossview-8k](https://huggingface.co/datasets/QianYangMILA/vdrop-crossview-8k) | cc-by-4.0 | | `zebra_jigsaw` | [multimodal-reasoning-lab/Zebra-CoT](https://huggingface.co/datasets/multimodal-reasoning-lab/Zebra-CoT) | cc-by-nc-4.0 | | `zebra_multihop` | [multimodal-reasoning-lab/Zebra-CoT](https://huggingface.co/datasets/multimodal-reasoning-lab/Zebra-CoT) | cc-by-nc-4.0 | Sources whose upstream declares no licence (`ipt_pet`, `messytable_relabel`, `thinkmorph_jigsaw`, `thinkmorph_spatial_nav`) are shared for research use only; the upstream terms apply, and the copy is removed on the authors' request.