unfundedResearcher/Minecraft-GLB2Schem-RepairPairs-v1

Paired (generated input, ground-truth target) Minecraft schematics for training a model that turns an approximate voxelisation into a real build.

What a sample is

Each sample is three files inside a WebDataset TAR shard:

File	Meaning
<id>.input.schem	GENERATED. Produced by voxelising the source .glb. Approximate and noisy.
<id>.target.schem	GROUND TRUTH. The original schematic, copied byte-for-byte, never modified.
<id>.json	Per-sample metadata (hashes, dimensions, block counts, runtime).

input.schem is machine-generated from the GLB. target.schem is the untouched original. Target files are byte-identical to the source dataset; their SHA-256 is recorded in each sample's JSON and in the manifest.

Source dataset: unfundedResearcher/Minecraft-Fable-SchemGLB-v1

How the input was generated

The .glb files are voxelised at 1 GLB lattice unit = 1 Minecraft block using a single global configuration for every file. No ground-truth schematic was ever consulted to choose a sample's scale, dimensions, rotation, origin, palette or occupancy.

• Block palette and per-block average colours come from ObjToSchematic, commit b611b52c7a87cad99127908598bab019fe9e7c3f (res/palettes/all.ts, res/atlases/vanilla.atlas) - 292 full-cube blocks.
• ObjToSchematic's own headless voxeliser is not used: it normalises every model to a fixed height along a constraint axis, which would destroy world-space proportions. The voxeliser here is a deterministic surface voxeliser. See converter/ and conversion_config.json.
• Colour to block by nearest neighbour in RGB. Dithering disabled. No per-file palette selection.
• Output is Sponge schematic version 2, DataVersion 2975, gzip with mtime=0, so conversion is bit-for-bit reproducible.

Known characteristics - read before training

• The source GLBs are hollow. Interior faces were culled when those GLBs were exported, so surface voxelisation can only recover a shell. interior_fill=true fills fully-enclosed air pockets using only the generated grid. This is the largest systematic difference between input and target: inputs may over- or under-fill interiors relative to ground truth.
• Input and target dimensions often differ. The input grid is cropped to actual occupancy, while ground-truth schematics frequently carry padding air. Both are recorded per sample.
• The source GLBs have an inverted texture V coordinate (Minecraft convention with v=0 at the bottom, versus glTF's v=0 at the top). The converter corrects this on load (uv_v_flip=true). Without the correction every UV samples transparent atlas rows and no blocks are produced at all.

Layout

data/pairs-00000.tar ...        ~250 pairs per shard
manifest.parquet / manifest.csv one row per published sample
conversion_config.json          every converter option
validation_10_report.json       the pre-flight 10-sample gate
validation_10_contact_sheet.png visual check of that gate
failures.jsonl                  every sample that failed, with reason
remote_verification.json        remote listing checked against the local manifest
converter/                      full converter source
requirements.txt

Reproducing

pip install -r requirements.txt
python converter/pipeline.py --root . --work ./run \
    --repo unfundedResearcher/Minecraft-GLB2Schem-RepairPairs-v1 --token $HF_TOKEN --shard-size 250 --workers 2

conversion_config.json hash: 4056132ca74bdba9

Run statistics

{
  "source_pairs_discovered": 11508,
  "source_incomplete_pairs": 0,
  "source_duplicate_ids": 0,
  "successful_conversions": 11505,
  "failed_conversions": 3,
  "shards": 48,
  "shard_size": 250,
  "median_conversion_seconds": 0.249,
  "p95_conversion_seconds": 5.806,
  "mean_conversion_seconds": 1.173,
  "total_conversion_cpu_seconds": 13491.0,
  "converter_commit": "b611b52c7a87cad99127908598bab019fe9e7c3f",
  "conversion_config_hash": "4056132ca74bdba9",
  "destination": "https://huggingface.co/datasets/unfundedResearcher/Minecraft-GLB2Schem-RepairPairs-v1",
  "note": "Metadata regenerated from an independent audit of the published shard contents. Every row corresponds to a sample physically present in the shard named, with input/target SHA-256 verified against the bytes."
}