license: mit
pretty_name: Minecraft Chunks
tags:
- minecraft
- terrain-generation
- voxel
- procedural-generation
- 3d
- games
task_categories:
- other
size_categories:
- 100M<n<1B
configs:
- config_name: default
data_files:
- split: train
path: chunks/*.csv
Minecraft Chunks
2,040 full Minecraft chunks (~200 million voxels) exported block-by-block from vanilla 1.21 worlds, collected to train STEVE-1, a style-conditioned terrain-generation VAE.
Each CSV file is one complete chunk: a 16 x 384 x 16 voxel column from bedrock (Y = -64) to build limit (Y = 319), one row per voxel, 98,304 rows per file.
Schema
| Column | Type | Description |
|---|---|---|
x, z |
int | Block position within the chunk (0-15) |
y |
int | World height (-64 to 319) |
ChunkBiome |
str | Biome at the chunk's center column (one value per chunk) |
Biome |
str | Biome at this exact block |
Block_ID |
str | Bukkit material name (STONE, GRASS_BLOCK, AIR, ...) |
Is_Surface |
bool | Whether this block is the highest non-air block of its column |
Light_Level |
float | Light level at this block |
Block_to_Left / _Right / _Below / _Above / _in_Front / _Behind |
str | Material names of the 6 face-neighbours |
A note on the context columns. Is_Surface, Light_Level, and the six neighbour columns describe the terrain itself. They are included for analysis and visualisation, but if you train a terrain generator, do not condition on them -- at generation time they do not exist yet (that is data leakage, and it is exactly the mistake that motivated rebuilding this project's pipeline). A generator should condition only on position, biome, and noise. See the training pipeline for the leakage-free reference implementation, including a unit test that enforces it.
Biome distribution
Chunk counts by ChunkBiome (2,040 total):
| Biome | Chunks | Biome | Chunks | |
|---|---|---|---|---|
| FOREST | 278 | SUNFLOWER_PLAINS | 55 | |
| PLAINS | 271 | LUKEWARM_OCEAN | 54 | |
| RIVER | 208 | DESERT | 54 | |
| DARK_FOREST | 194 | BIRCH_FOREST | 45 | |
| MANGROVE_SWAMP | 154 | BAMBOO_JUNGLE | 38 | |
| OLD_GROWTH_BIRCH_FOREST | 147 | COLD_OCEAN | 30 | |
| ERODED_BADLANDS | 112 | SAVANNA | 30 | |
| BADLANDS | 99 | OCEAN | 11 | |
| JUNGLE | 95 | SWAMP | 5 | |
| SPARSE_JUNGLE | 85 | FLOWER_FOREST | 2 | |
| BEACH | 72 | TAIGA | 1 |
The distribution is imbalanced (worlds are mostly forest and plains); the reference pipeline compensates with a biome-balanced sampler at training time.
How it was collected
A custom PaperMC plugin reads chunks block-by-block on a live server and writes one CSV per chunk:
/grabchunkdata-- the current chunk/grabchunkarea <radius>-- every chunk in a square radius (bulk collection)/grabbiome <biome> <count>-- spiral-search targeted collection of rare biomes/grabblock <block> <count>-- targeted collection of chunks containing a block
Filenames follow <prefix>_<chunkX>_<chunkZ>.csv (targeted collection inserts the biome or material name into the prefix).
Usage
With the reference pipeline (recommended -- this is the format its GT_DATA_DIR expects):
from huggingface_hub import snapshot_download
path = snapshot_download("ghosteau/minecraft-chunks", repo_type="dataset")
# then: set GT_DATA_DIR to <path>/chunks and run the training notebook
As a plain table:
from datasets import load_dataset
ds = load_dataset("ghosteau/minecraft-chunks") # ~200M rows
One chunk as a 3D array:
import numpy as np
import pandas as pd
df = pd.read_csv("chunks/cluster1_-100_-1.csv")
grid = np.full((16, 384, 16), "", dtype=object)
grid[df["x"], df["y"] + 64, df["z"]] = df["Block_ID"] # [X, Y, Z]
Related
- Model trained on this data: ghosteau/STEVE-1
- Training pipeline, plugin source, and fine-tuning notebook: github.com/ghosteau/generative-terrain
License and attribution
MIT. The data consists of block coordinates, material identifiers, and biome labels exported from procedurally generated Minecraft worlds. Not affiliated with or endorsed by Mojang or Microsoft; "Minecraft" is a trademark of Mojang Synergies AB.