VoxHouse3D 1.1.0

4,030 player-built Minecraft structures cut from archived community server worlds, labelled by a single annotator as house (2,007) or other_structure (2,023). Released with the paper Shelter and Access: One-Sided, Falsifiable Interior Measures for Generated Voxel Buildings (NeurIPS 2026 Workshop on Embodied Spatial Reasoning; PDF) to support one-sided, falsifiable evaluation of generated buildings. Code: github.com/KyleLin0927/structgauge. It is not a leaderboard.

Quick start

python3 reproduce/reproduce.py

Standard library only. Recomputes the AUC table below from measures/measures.csv, verifies every file against CHECKSUMS.sha256, and exits non-zero on any disagreement.

pip install numpy 'scipy>=1.12'
python3 code/measure.py

Recomputes shelter and access for 20 builds from the voxel geometry, with the implementation that produced the CSV, and reports the largest difference (--all for every build). To measure your own builds, see code/README.md.

measure	AUC, house vs other_structure
shelter	0.8681
access	0.8424
size-only baseline solid_fraction	0.6550
size-only baseline log_block_count	0.6275

Confidence intervals are in reproduce/expected.json. The paper's isovist baseline is not part of this release.

Files

path	contents
geometry/*.npz	one file per build, 8.0 MB in total (see "Format")
measures/measures.csv	per build: shelter, access, solid_fraction, log_block_count, block_count
metadata/builds.jsonl	per build: public_id, class, world_code, world, world_xyz (minimum corner), dims, license_spdx, geometry
metadata/worlds.csv	world code, world save, build count, release licence
metadata/overlap_exclusion_pairs.json	59 cross-label pairs whose bounding boxes overlap (110 builds); excluding them moves each AUC by less than 0.004
reproduce/	reproduce.py, expected.json, legacy_map.json, decode_legacy.py
code/	the structgauge package, measure.py, MANIFEST.json (per-file md5)
DATASHEET.md	how the data was collected and labelled
ATTRIBUTION.md	source archive and declared licence of each world
LICENSE.txt	licence terms
CHECKSUMS.sha256	SHA-256 of every other file

Per-voxel field caches and rendered views are not included; both can be recomputed from the geometry.

Format

Each .npz holds vocab (sorted block-state strings, "" = empty) and blockstates (uint16 index into vocab, one per cell, indexed [x, y, z] with y up). Arrays are not tight crops: 603 of the 4,030 have no all-empty layer, and the rest carry padding, most often three empty layers on top.

Block-state strings come in two forms, and no build mixes them:

form	example	builds	worlds
modern	minecraft:oak_stairs[facing=north]	2,320	w07-w28, and 32 of the 34 builds of w11
legacy	legacy:53:2	1,710	w01-w06, and 2 builds of w11

Legacy worlds were saved before the block flattening and only ever stored the numeric 1.12 (id, data) pair; this release has no mapping from legacy to modern strings. The measures are defined on the 1.12 pair for both forms: reproduce/legacy_map.json maps every string in the release to its [id, data] pair, and reproduce/decode_legacy.py applies it. For modern strings the pair is a lossy projection (an oak door has 64 states but only 16 data values).

block_count counts solid cells only (torches, plants and panes are excluded), so it differs from the number of non-empty cells; log_block_count is its natural logarithm. solid_fraction is block_count divided by the volume of the tight box around the non-empty cells, not by dims.

Identifiers

VH3D-w07-00b60446
-+-- -+- ----+---
 |    |      +- sha1("{world}|{x},{y},{z}|{dx},{dy},{dz}"), first 8 hex
 |    +-------- world code, see metadata/worlds.csv
 +------------- dataset prefix

{world} is the world field of builds.jsonl, {x},{y},{z} is world_xyz and {dx},{dy},{dz} is dims. The identifier does not depend on the label, so relabelling a build does not change it.

Clustering

Builds from the same world are correlated, so, as in the paper, reproduce.py resamples whole worlds (28 world saves, 26 of them containing houses) and draws both classes from the same resample. The worlds are of very different sizes: the largest, w03, holds 40.1% of the builds, and the three largest hold 79.6% of the houses. Resampling individual builds would give intervals that are too narrow.

Licence, credits and takedown

The data are released under CC BY-NC 4.0 (LICENSE.txt); the scripts in code/ and reproduce/ are Apache-2.0. The structures were built by anonymous players on public servers and reached us through Internet Archive uploads by the server communities. ATTRIBUTION.md lists each world's source and the licence its uploader declared; attribution is required. The files contain geometry only: no sign text, book contents or player names.

To have a structure removed, open a thread on this dataset's Community tab, file an issue at https://github.com/KyleLin0927/structgauge/issues, or write to kylelin0927@gmail.com. It will be removed in the next release.

Citation

See CITATION.cff. Paper: https://github.com/KyleLin0927/structgauge/blob/main/paper/shelter_and_access_esr2026.pdf

Known limitations

• One annotator. An inter-annotator check on a stratified sample of 100 builds by three non-author annotators, blind to labels and measures, gives majority-vote agreement 93/98 (κ = 0.89; paper, Appendix B).
• Clustered sources. See "Clustering".
• 143 builds were withheld for content (player-skin pixel art, copyrighted images, portraits of real people, text signage), all from the other_structure class.
• Near-duplicates and containment overlaps may remain; see DATASHEET.md.