---
license: cc-by-nc-4.0
pretty_name: Dream-Cubed Human
tags:
  - minecraft
  - voxel
  - 3d-generation
  - human-authored-data
  - diffusion-models
  - structure-generation
size_categories:
  - 1M<n<10M
viewer: false
---
# Dream-Cubed Human

Dream-Cubed Human is a dataset of voxelized Minecraft structures and terrain represented as 32x32x32 block chunks. It is intended for research on controllable 3D generation, Minecraft-like structure generation, inpainting, outpainting, and synthetic environment generation.

This repository contains raw chunks extracted from six human-authored Minecraft map sources and a processed augmented dataset used for training. The processed `human_augmented_dataset/` also includes natural biome chunks from the companion Dream-Cubed Natural release; the raw files for those natural chunks are hosted in the separate natural dataset repository.

The core dataset is hosted separately and can be found at https://huggingface.co/datasets/dream-cubed/DreamCubedNatural

## Dataset Contents

The release is organized into two main directories:

- `human_map_chunks/`: raw `.npz` archives scraped from six human-authored Minecraft map sources. The subdirectories are `ancient_empire`, `kingdom_of_sarano`, `marethea`, `osirion`, `sora_kingdom`, and `the_5_bridges`.
- `human_augmented_dataset/`: processed, sharded class-conditional dataset containing human-authored chunks augmented with natural biome chunks from the companion natural release.

All six human-authored source maps were provided by Timtenth_Buildings. Creator: https://www.youtube.com/@timtenth_buildings8113. The source labels are ancient_empire, kingdom_of_sarano, marethea, osirion, sora_kingdom, and the_5_bridges. The original maps are not redistributed here.

Permission to use these maps and release voxelized chunks in this format was granted explicitly by the map creator, Timtenth_Buildings. The creator does not permit commercial products based on this data or attempts to reconstruct, clone, or redistribute the original maps from which the chunks were extracted.

## File Format

Raw human-map chunks are stored as NumPy `.npz` archives containing integer Minecraft block-ID tensors. Samples are 32x32x32 voxel chunks, with batching depending on the archive.

The processed dataset contains train/validation split directories with:

- `manifest.json`: split metadata, including paths, shapes, dtypes, sample counts, label counts, and sharding information when present.
- voxel data: memory-mapped `voxels.npy` files or voxel shards, depending on the packaged split.
- label array: integer class labels, with the exact filename recorded in `manifest.json`.
- mapping files: block and label index mappings used by Dream-Cubed training and inference code.

Always read paths from `manifest.json` or `sharded_parent_manifest.json` rather than hardcoding file names.

## Loading Example

The processed release is sharded. Download the desired variant with its manifests,
voxel shards, label arrays, and mapping file. The following paths are relative to
the downloaded dataset repository root:

| Variant | Training split | Validation split | Mapping file (inside variant) |
| --- | --- | --- | --- |
| `human_augmented_dataset` | `natural_human_balanced_dataset_cleaned_cc_dir` | `natural_human_balanced_dataset_cleaned_cc_val_dir` | `natural_human_balanced_dataset_cleaned_cc_oh_mappings.pt` |

The supplied `shard_processed_dataset.py` restores the original contiguous NumPy
arrays without loading the complete dataset into RAM. Install NumPy, then run:

```bash
python -m pip install numpy
python shard_processed_dataset.py restore human_augmented_dataset restored_human_augmented_dataset
```

Choose a new, empty output directory. Restoration retains the downloaded shards
and creates an additional full copy of the voxel arrays, so allow sufficient disk
space for both. It is not needed for the already unsharded representative samples.

```python
import json
from pathlib import Path
import numpy as np

dataset_dir = Path("restored_human_augmented_dataset")
split_dir = dataset_dir / "natural_human_balanced_dataset_cleaned_cc_dir"
manifest = json.loads((split_dir / "manifest.json").read_text())
voxels = np.load(split_dir / manifest["paths"]["voxels"], mmap_mode="r", allow_pickle=False)
labels = np.load(split_dir / manifest["paths"]["biome_labels"], mmap_mode="r", allow_pickle=False)
print(voxels.shape, labels.shape)
print(voxels[0].shape, int(labels[0]))
mapping_path = dataset_dir / "natural_human_balanced_dataset_cleaned_cc_oh_mappings.pt"
print(mapping_path)
```

Voxel values in processed arrays are vocabulary indices. Use the accompanying
mapping file when decoding them, and keep the checkpoint's vocabulary and dataset
mapping paired. The human-augmented labels include natural-biome classes as well
as human-map source classes.

## Data Collection

The raw human chunks were extracted from selected regions of six human-authored Minecraft map sources using a Minecraft world-selection and chunk-extraction pipeline. The release contains voxelized block-ID tensors rather than complete original worlds.

The processed augmented dataset also includes procedurally generated natural chunks from Minecraft Java Edition v1.12.2. Those natural chunks were collected from locally run Minecraft servers using a custom gRPC extraction API based on Evocraft, with cave and village-specific collection procedures described in the companion natural dataset.

No Minecraft textures, sounds, code, chat logs, account identifiers, player behavior traces, or complete original map files are included in this repository.

## Limitations and Biases

The human-authored portion is limited to six curated map sources and is not representative of all Minecraft building styles, creator communities, or game versions. Region selection emphasizes built structures, so empty terrain, surrounding context, interiors, and chunk boundaries may be unevenly represented.

The augmented processed dataset also inherits the limitations of the natural release: natural chunks come from Minecraft Java Edition v1.12.2 only, later Minecraft biomes and blocks are absent, Mesa and Mushroom biome families are excluded, and biome labels are chunk-level rather than voxel-level semantic labels.

All samples are independent 32x32x32 volumes, so long-range continuity across chunk boundaries is limited. Block compression and preprocessing may simplify rare, decorative, or low-signal blocks. Source labels identify dataset split or source class rather than structure type, quality, style, function, or semantic content.

This dataset is not intended for creator identification, authorship attribution, fairness auditing, preference modeling, commercial products, or attempts to recreate the original source maps.

## License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0).

### Important: Minecraft EULA And Map Creator Rights

The natural biome data in the processed dataset was extracted from procedurally generated Minecraft Java Edition v1.12.2 worlds. Minecraft is © Mojang AB / Microsoft, and use of Minecraft and content derived from it is governed by the **Minecraft End User License Agreement** (https://www.minecraft.net/en-us/eula) and Mojang's **Usage Guidelines** (https://www.minecraft.net/en-us/usage-guidelines).

This dataset contains only block-ID tensors and labels; no Minecraft textures, sounds, code, or other Mojang-owned assets are included or redistributed. **Downstream users are independently responsible for complying with the Minecraft EULA**, including its restrictions on commercial use of Minecraft-derived content.

The human-authored map data is separate creative content and remains the property of the original map creator, Timtenth_Buildings. Permission was granted to use those maps and release voxelized chunks in this format for noncommercial research, but not to build commercial products from this data or to reconstruct, clone, or redistribute the original maps. The CC BY-NC 4.0 license above applies to our compilation and annotation of this dataset. Nothing in our license overrides or grants rights inconsistent with either the Minecraft EULA or the original map creator's rights.

"Minecraft" is a trademark of Mojang AB. This dataset is not affiliated with, endorsed by, or sponsored by Mojang or Microsoft.

## Authors and citation

Tim Merino, Sam Earle, Ryunosuke Iwai, Julian Togelius, Edoardo Cetin. Dream-Cubed: Controllable Generative Modeling in Minecraft by Training on Billions of Cubes. NeurIPS 2026, Evaluations and Datasets Track.

Affiliations: New York University (Tim Merino, Sam Earle, Julian Togelius) and Sakana AI (Ryunosuke Iwai, Edoardo Cetin).

Code: https://github.com/TimMerino1710/DreamCubed

Research correspondence: Tim Merino, tm3477@nyu.edu.