yrlyrl/lvr-data-mmcot_mix_v3

Spatial MMCoT training mixture mmcot_mix_v3

26,176 training rows -- real 13,088, synthetic 13,088. The rows are those of yrlyrl/spatial-mmcot-v2 at revision f856cdda (same row_uid, same bytes; its train split only), chosen deterministically and shuffled once, so every 128-row group mixes sources. rows.tsv lists every row in file order with its source, task, scene corpus and side; manifest.json has the recipe and its sha1, each input's row_uid digest and the counts below.

Real and synthetic rows 1:1, both sides drawn only from the task kinds whose helper images measurably helped a model answer. In the helper-gain run, Qwen3.8-27B answered each sampled question twice, once with the input images only and once with the row's ground-truth helper images added, without the reasoning text. Thinking-mode accuracy, inputs only -> with helpers (sampled rows):

side	group	accuracy	rows sampled
real	zebra_jigsaw jigsaw (ImageNet)	48.8% -> 62.8% (+14)	43
real	thinkmorph_jigsaw jigsaw (ADE20K, SUN RGB-D)	71.4% -> 97.1% (+26)	35 (32 of them photos)
real	messytable_relabel most common object	85.7% -> 100% (+14)	7
synthetic	motif_maze maze	34.4% -> 75.0% (+41)	32
synthetic	vdrop shared-object anchor	50.0% -> 80.0% (+30)	10
synthetic	thinkmorph_spatial_nav frozen-lake navigation	58.3% -> 81.2% (+23)	48
synthetic	motif_ball_tracking ball tracking	82.9% -> 100% (+17; answering directly 20.0% -> 97.1%)	35

This is mmcot_mix_v2 without messytable multi-view counting (6,068 real rows; -1 point), which is why it is smaller: the real side is 13,088 rows and the synthetic side as many, equal per group and capped by what a group has. Same seed as mix_v2, so every row here is also a row of mix_v2. Mind the balance: jigsaw reassembly is 98% of the real side and 49% of all rows, and multi-view counting is down to the 266 most-common rows. Rows flagged with a text-only known issue (A.text_issue) or with reasoning that calls the panorama wrap-around (A.vdrop_wrap_claim, which is false) are excluded on both sides.

Recipe

name: mmcot_mix_v3
seed: mmcot-mix-v2
split: train
source_from: meta
release: yrlyrl/spatial-mmcot-v2@f856cdda
real_corpora:
- imagenet
- ade20k
- sunrgbd
- messytable
real:
- zebra_jigsaw:visual_jigsaw
- thinkmorph_jigsaw
- messytable_relabel:multiview_most_common
exclude_flags:
- A.text_issue
- A.vdrop_wrap_claim
synthetic:
  groups:
  - motif_maze:maze_naive
  - vdrop:anchor
  - thinkmorph_spatial_nav:Spatial_Navigation
  - motif_ball_tracking:ball_tracking_naive
  total: match_real
  allocation: equal

Rows

source	rows	licence of the source
messytable_relabel	266	undeclared
motif_ball_tracking	4,151	mit
motif_maze	4,151	mit
thinkmorph_jigsaw	2,530	undeclared
thinkmorph_spatial_nav	4,150	undeclared
vdrop	636	cc-by-4.0
zebra_jigsaw	10,292	cc-by-nc-4.0

side	rows
synthetic	13,088
real	13,088

scene corpus	rows
procedural_grid	12,452
imagenet	10,292
ade20k	1,780
sunrgbd	750
infinigen	636
messytable	266

task	rows
visual_jigsaw	10,292
maze_naive	4,151
ball_tracking_naive	4,151
Spatial_Navigation	4,150
anchor	636
jigsaw_1x3	529
jigsaw_3x1	522
jigsaw_2x2	513
jigsaw_2x1	484
jigsaw_1x2	482
multiview_most_common	266

Excluded before either side was drawn (rows of the split carrying the flag, by source:task; a task the recipe does not draw from loses nothing it would have used):

flag	source:task	rows
A.text_issue	zebra_jigsaw:visual_jigsaw	245
A.text_issue	zebra_multihop:multihop_counting	5
A.vdrop_wrap_claim	vdrop:anchor	20
A.vdrop_wrap_claim	vdrop:counting	62
A.vdrop_wrap_claim	vdrop:relative_distance	371

real group	rows (all of the group)
zebra_jigsaw:visual_jigsaw	10,292
thinkmorph_jigsaw	2,530
messytable_relabel:multiview_most_common	266

synthetic group	rows it had	rows taken
motif_ball_tracking:ball_tracking_naive	7,687	4,151
motif_maze:maze_naive	6,991	4,151
thinkmorph_spatial_nav:Spatial_Navigation	5,611	4,150
vdrop:anchor	636	636

Format and loading

The loader schema of the release (image_list list: the inputs, then the target images in order; num_input_images; instruction_list with the system prompt, question and options; output_text_list with the K+1 segments <think>..</think><image_start> ... <image_end><think>..</think><answer>..</answer>; row_uid).

• BAGEL / ThinkMorph trainer: download into <DATA_ROOT>/mmcot_mix_v3/ (the directory must keep the name: the parquet_info.json keys are mmcot_mix_v3/train/<file>) and use data_dir: <DATA_ROOT>/mmcot_mix_v3/train.
• LatentUM: train/pretrain_latentum.py --data <dir>/train --data-format mmcot (data/latentum_mmcot.py).

Each source keeps its own licence and terms (table above; the source cards give the upstream terms in full). The licence of a row is the licence of its source; check every source you use.