yrlyrl/lvr-data-mmcot_mix_v1

Spatial MMCoT training mixture mmcot_mix_v1

38,806 training rows: every real-photograph training row of the release and as many synthetic rows, split equally over the synthetic task families -- real 19,403, synthetic 19,403. The rows are those of the per-source repositories below (same row_uid, same bytes; their 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 export's row_uid digest and the counts below.

Recipe

name: mmcot_mix_v1
seed: mmcot-mix-v1
split: train
real_corpora:
- imagenet
- ade20k
- sunrgbd
- messytable
real: all
synthetic:
  sources:
  - ipt_mvc
  - ipt_pet
  - motif
  - thinkmorph_spatial_nav
  - vdrop
  - zebra_multihop
  total: match_real
  allocation: equal

Rows

source	rows	licence of the source
ipt_mvc	3,308	undeclared
ipt_pet	3,308	undeclared
messytable_relabel	6,334	undeclared
motif	3,308	mit
thinkmorph_jigsaw	2,531	undeclared
thinkmorph_spatial_nav	3,307	undeclared
vdrop	3,307	cc-by-4.0
zebra_jigsaw	10,538	cc-by-nc-4.0
zebra_multihop	2,865	cc-by-nc-4.0

side	rows
real	19,403
synthetic	19,403

scene corpus	rows
imagenet	10,538
procedural_grid	8,546
procthor	6,616
messytable	6,334
infinigen	3,307
ade20k	1,780
blender_clevr	934
sunrgbd	751

task	rows
visual_jigsaw	10,538
multiview_count	6,068
perspective_taking	3,902
mvc_count	3,308
Spatial_Navigation	3,307
multihop_counting	2,865
relative_distance	1,385
ball_tracking_naive	1,253
maze_naive	1,121
manipulation_naive	934
relative_direction	691
jigsaw_1x3	529
jigsaw_3x1	522
jigsaw_2x2	513
jigsaw_2x1	484
jigsaw_1x2	483
anchor	351
counting	286
multiview_most_common	266

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_v1/ (the directory must keep the name: the parquet_info.json keys are mmcot_mix_v1/train/<file>) and use data_dir: <DATA_ROOT>/mmcot_mix_v1/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.