MassCount-CF

Synthetic counterfactual counting corpus for identifying an additive neural-mass cardinality coordinate in vision-language models (NMCA). Companion dataset to "Counting Requires Mass: An Algebraic and Causal Account of Numerosity in Vision-Language Models."

Corpus version: masscount-cf-1.0.0 Master scenes: 99,989 Delivered images (this upload): 326,623 Object instances: 28,617,018

Structure

Scene graphs are the durable artifact; pixels are regenerable from them at any resolution via the generator (generate.py in the companion repo). Each row in scenes/*.parquet is one scene with per-state geometric descriptors (hull area, coverage, nearest-neighbour statistics, Ripley's K); each row in objects/*.parquet is one placed object with normalized [0,1]^2 coordinates.

versions/masscount-cf-1.0.0/
  scenes/shard_*.parquet      one row per scene, with descriptors + split
  objects/shard_*.parquet     one row per placed object
  graph_report.json           integrity report: dedup, split-leak, dupe counts
  manifest.json                the exact build manifest for this version
  images/{resolution}/{shard}/*.webp   (present only if uploaded with images)

Families

family	scenes
chain	44,400
nuisance	19,915
union_part	7,500
instance_rep	4,800
heldout_natural_bg	4,000
heldout_distractor_only	4,000
heldout_cutout	4,000
union	3,750
tile	3,600
dewind	2,820
tile_base	1,200
chain_origin	4

Splits

Grouped by parent_chain_id, never by image, following the category-disjoint protocol of FSC-147 (Ranjan et al. 2021, CVPR). reference is a small set of shared calibration-origin scenes (X_0 per chain band) and is not a model evaluation split.

split	scenes
test	62,499
train	31,154
val	6,332
reference	4

Integrity

This version passed, at upload time:

• unique scene_id: True
• unique content_hash (excluding the reference split): True
• cross-split content duplicates: 0 (must be 0)
• chains split across train/val/test: 0 (must be 0)

Citation

If you use this corpus, please cite the accompanying paper.

License

CC-BY-4.0. All content is synthetically generated; no real images or copyrighted material are included, with the exception of any user-supplied cutout asset library, which is the uploader's responsibility to license correctly.