yrlyrl/spatial-mmcot

Spatial MMCoT v1

Spatial reasoning data in one interleaved format: input image(s) and a question, then rounds of plan → target image, then a read-back and the answer. How much of that the text carries differs by source (supervision_kind in each source's meta, and each card says which):

• full_interleaved (motif, thinkmorph_jigsaw, thinkmorph_spatial_nav, zebra_jigsaw, zebra_multihop): the trace interleaves text and target images (drawn or rendered states; the source cards say which), and the read-back comes after the image it reads.
• visual_aux (ipt_mvc, ipt_pet): in r3 the text came from a text-only chain of thought that never saw the generated image (meta keeps that value); the r4 text was rewritten by a model shown every image, the targets included.
• visual_only (vdrop): no reasoning text: every thought is empty.

This release: r4, the training subset

Every thought was rewritten by Qwen3.8-27B (thinking mode, prompt v4.1) from the row's own pictures, question, options and answer; the images, questions, options and answers are those of the previous release (r3), checked row by row. This is the subset used for training: 740 rows label_doubt, 77 rows over_budget, 1 rows R.no_rewrite are removed (each source's reports/subset_dropped.jsonl names them), and two r3 sources, cova and zebra_tetris, are left out because the self-check found their long chains drift (2.1 faithfulness errors per row on both). Their repositories keep the r3 rows and are not linked below. Each source card has an r4 section with its own counts and self-check.

Also part of this release, outside the table below:

• yrlyrl/spatial-mmcot-messytable_relabel: messytable_relabel: real MessyTable photographs, counting and most-numerous questions asked again from MessyTable's own per-instance labels, the overhead camera's photo as the helper image; built by its own chain (tag mt1) and published in its own repository.
• yrlyrl/lvr-data-mmcot_mix_v1: the training mixture mmcot_mix_v1: every real-photo training row (zebra_jigsaw, thinkmorph_jigsaw's ADE20K / SUN RGB-D rows, messytable_relabel) and as many synthetic rows, split equally over ipt_mvc, ipt_pet, motif, thinkmorph_spatial_nav, vdrop and zebra_multihop, shuffled into one loader-schema dataset (recipe and per-row list in its manifest).

Known issues

Every source card has a Known issues section: rows with a measured per-row problem, listed in the source's reports/known_issues/ so they can be filtered out (no row was removed for them), and measured caveats that cannot be listed row by row. Read it before training on a source.

Rows listed per issue (a row can be listed under more than one issue):

• ipt_mvc: options_give_answer 1,483
• ipt_pet: contradicting_answers_same_images 15; plan_turn_rule_backwards 1
• motif: count_options_give_answer 146
• thinkmorph_jigsaw: target_shared_with_other_row 12; text_states_other_arrangement 3; validation_target_near_copy_of_train 1
• thinkmorph_spatial_nav: validation_maze_near_twin_in_train 46; readback_claims_path_fails 30; plan_misplaces_start 4
• zebra_jigsaw: readback_names_no_option 237; text_says_options_not_shown 25; target_shared_with_other_row 16; question_states_answer 1
• zebra_multihop: target_repeats_earlier_target 665; readback_states_no_count 5

Sources and licences

Each source is its own repository, with its own licence; the licence column gives each source's upstream licence, and each source card's Citation section names the upstream work to cite.

source	train	validation	target image slots	distinct target images	licence	repo
ipt_mvc	8,866	208	9,074	4,680	undeclared	yrlyrl/spatial-mmcot-ipt_mvc
ipt_pet	8,409	211	8,620	4,968	undeclared	yrlyrl/spatial-mmcot-ipt_pet
motif	20,298	631	44,243	44,210	mit	yrlyrl/spatial-mmcot-motif
thinkmorph_jigsaw	5,651	143	5,794	5,788	undeclared + ADE20K / SUN RGB-D terms	yrlyrl/spatial-mmcot-thinkmorph_jigsaw
thinkmorph_spatial_nav	5,645	162	5,807	5,807	undeclared	yrlyrl/spatial-mmcot-thinkmorph_spatial_nav
vdrop	6,543	264	6,807	1,152	cc-by-4.0	yrlyrl/spatial-mmcot-vdrop
zebra_jigsaw	10,538	316	10,854	10,846	cc-by-nc-4.0 + ImageNet terms	yrlyrl/spatial-mmcot-zebra_jigsaw
zebra_multihop	2,865	89	10,870	10,139	cc-by-nc-4.0	yrlyrl/spatial-mmcot-zebra_multihop
total	68,815	2,024	102,069	87,590

Distinct target images are counted by content hash within each source.

Benchmark overlap

• ipt_mvc: the IPT MVC AI2-THOR evaluation set draws its options the same way as this source (four consecutive counts, the correct one never the largest), so a model trained here can gain on it from the options alone; see the source card.
• ipt_pet shares no ProcTHOR house with the IPT PET-AI2THOR evaluation set (weikaih/imaginative-perception-token-pet-eval-ai2thor), but every one of that set's questions uses a template found in this source's training rows, and its renders come from the same ProcTHOR renderer: that benchmark is in-domain, not out-of-distribution, for a model trained on it.
• motif: VisWorld-Eval ballgame uses the same task text as ball_tracking_naive, so treat it as in-domain; VisWorld-Eval multihop asks the same kind of question as manipulation_naive about CLEVR-style scenes (no shared images found by S11), so treat it as near-domain; sokoban_naive is excluded from this release (its card says why).
• thinkmorph_jigsaw: 2 puzzles whose assembled photograph is a What's Up image were removed (S11); upstream ThinkMorph Jigsaw_Assembly still contains them.
• thinkmorph_spatial_nav: mazes identical to VSP-maze evaluation items were removed (S11), but VSP-maze is in-domain (the ThinkMorph paper uses it as its in-domain benchmark for this task): it uses the same FrozenLake renderer and question text, and most of its 3x3 items have a training maze that differs only in where the start is. Only its 7x7 and 8x8 items are held out, and they are larger than any training grid (3x3 to 6x6).

Files

parquet_info_all.json merges the released sources' indexes for a mixed training run. Its keys are <source>/<split>/<file>, and the loader matches them against os.path.join(data_dir, file) exactly. Download each source into <root>/<source> (the source name, not the repository name) and either run from <root> with data_dir: <source>/train, or rewrite every key k to os.path.join(root, k) and use data_dir: os.path.join(root, '<source>', 'train'). Do not strip the <source>/ prefix: the sources' file names collide. A source whose keys do not match adds no rows to the mixed group, and there is no error.

num_used_data counts chunk files, not rows: the loader repeats a source's file list up to that many entries, so a source read with the same value as the others but twice as many files has each of its rows read half as often. Give every source the same multiple of its training chunk-file count, 128 per file as below: every file is then listed 128 times, every row group is read, and each source's share is proportional to its row groups (all row groups hold up to 128 rows, so up to each file's last, partial group, to its rows).

source	training chunk files	row groups	num_used_data
ipt_mvc	2	71	256
ipt_pet	1	66	128
motif	1	159	128
thinkmorph_jigsaw	1	45	128
thinkmorph_spatial_nav	1	45	128
vdrop	1	52	128
zebra_jigsaw	1	83	128
zebra_multihop	1	23	128

The rows of ipt_mvc (4 to 5 inputs), vdrop (2 inputs) have more than one input image; every other source has one. A run that includes them needs a loader that reads num_input_images (the UnifiedEditIterableDataset of the IPT release, https://github.com/weikaih04/Imaginative-Perception-Token, data/interleave_datasets/edit_dataset.py, or ThinkMorph's with the two-line change shown in those cards). The stock ThinkMorph loader takes image_list[0] as the only input and trains the other inputs as generated images, without an error: on a row with two inputs and one target it trains the second input after the plan and the real target after the answer, and on a row with three or more inputs and one target it never trains the real target.

The zebra_* thoughts keep upstream's 'THOUGHT n:' labels inside <think>; the other sources have none (the zebra cards show how to strip them).

reports/ holds the cross-source stage reports: export_summary.json, s10_summary.json, s12_summary.json, s13_summary.json, s10_label_conflicts.jsonl. The frozen S12 validation keys are in splits/ (val_keys.tsv: <source>, <corpus>::<key> for keyed policies, where <corpus> is meta.source_scene_corpus and <key> is meta.scene_id (meta.geometry_uid for zebra_jigsaw and thinkmorph_spatial_nav), with two exceptions: ipt_mvc and ipt_pet write the ProcTHOR house as scene-<N>, which meta.scene_id spells house_<N> / ProcTHOR_House_<N>, and in cova, motif, zebra_tetris and zebra_multihop rows that share an input image carry the smallest key of their group; val_members.tsv: <source>, first-image pHash for cluster policies; key_digest.tsv: per-source key digests). They are split keys, not data rows; the rows are in the per-source repositories.