The Dataset Viewer has been disabled on this dataset.

Dream-Cubed Natural

Dream-Cubed Natural is a dataset of procedurally generated Minecraft Java Edition v1.12.2 terrain represented as 32x32x32 voxel chunks. It is intended for research on controllable 3D generation, biome-conditioned voxel modeling, inpainting, outpainting, and synthetic environment generation.

This repository contains only the natural/procedural portion of Dream-Cubed. It includes raw chunks extracted from Minecraft worlds as well as processed class-conditional datasets used for training.

The human map dataset is hosted separately, and can be found at https://huggingface.co/datasets/dream-cubed/DreamCubedHuman

Dataset Contents

The release is organized into four main directories:

  • natural_biome_chunks/: raw NumPy chunk files before training preprocessing. Files are grouped by biome and contain one chunk or batches of up to about 1000 chunks.
  • balanced_dataset/: processed train/validation memmap dataset with approximately balanced biome classes.
  • boosted_dataset/: processed train/validation memmap dataset with village chunks represented at 2x the non-village class target.
  • natural_occ_dataset/: processed train/validation memmap dataset sampled to approximate natural Minecraft biome occurrence rates.

The raw natural data covers these biome labels:

beaches, birch_forest, cave, desert, extreme_hills, forest, ice, jungle, ocean, plains, river, savanna, swampland, taiga, and village.

File Format

Raw chunks are stored as NumPy arrays of integer Minecraft block IDs. A raw file may have shape [32, 32, 32] for a single chunk or [N, 32, 32, 32] for a batch of chunks.

Processed datasets contain:

  • mapping.pt: block and biome index mappings.
  • <dataset>_cc_dir/: training split.
  • <dataset>_val_cc_dir/: validation split.
  • manifest.json: split metadata, including paths, shapes, dtypes, sample counts, and biome counts.
  • voxels_shards/*.npy: sharded voxel arrays; each has shape [N_shard, 32, 32, 32]. Restoration produces a single voxels.npy.
  • label array: integer biome class labels, with the exact filename recorded in manifest.json.

Always read paths from manifest.json, since file names may differ slightly between packaged variants. Processed voxel arrays are sharded as memory-mappable NumPy files to meet hosting file-size limits.

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)
balanced_dataset 32_biome_parts_dir_balanced_cc_dir 32_biome_parts_dir_balanced_cc_val_dir 32_biome_parts_dir_balanced_cc_oh_mappings.pt
boosted_dataset village_boosted_dataset_cc_dir village_boosted_dataset_cc_val_dir village_boosted_dataset_cc_oh_mappings.pt
natural_occ_dataset natural_occurrence_dataset_2_cc_dir natural_occurrence_dataset_2_cc_val_dir natural_occurrence_dataset_2_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:

python -m pip install numpy
python shard_processed_dataset.py restore balanced_dataset restored_balanced_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.

import json
from pathlib import Path
import numpy as np

dataset_dir = Path("restored_balanced_dataset")
split_dir = dataset_dir / "32_biome_parts_dir_balanced_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 / "32_biome_parts_dir_balanced_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

Chunks were collected from locally run Minecraft Java Edition v1.12.2 servers using a custom gRPC extraction API based on Evocraft (https://github.com/real-itu/Evocraft-py). Natural overworld chunks were sampled by traversing generated worlds, cave chunks were identified through downward scanning and air-content criteria, and village chunks were collected around procedurally generated NPC villages located with PyUbiomes/Cubiomes (https://github.com/lilyyy411/Pyubiomes, https://github.com/Cubitect/cubiomes).

No player-created Minecraft maps, player identifiers, chat logs, or gameplay traces are included in this natural release.

Limitations and Biases

The dataset reflects Minecraft Java Edition v1.12.2 procedural generation only. Later Minecraft biomes, blocks, structures, and world-generation behavior are not represented. Mesa and Mushroom biome families are excluded due to low occurrence in the collection pipeline.

Chunks are independent 32x32x32 volumes, so long-range continuity across chunk boundaries is limited. Labels are assigned at the chunk level rather than the voxel level. The processed balanced and village-boosted variants intentionally alter the natural biome distribution, and models trained on those variants may overproduce rare biomes or villages relative to standard Minecraft generation.

License

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

Important: Minecraft EULA

The data in this repository was extracted from procedurally generated Minecraft worlds (version 1.12.2). 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 biome labels — no Minecraft textures, sounds, code, or other Mojang-owned assets are included or redistributed. The CC BY-NC 4.0 license above applies to our compilation and annotation of this data. Downstream users are independently responsible for complying with the Minecraft EULA, including its restrictions on commercial use of Minecraft-derived content. Nothing in our license overrides or grants rights inconsistent with the Minecraft EULA.

"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.

Downloads last month
604

Collection including dream-cubed/DreamCubedNatural