Datasets:
license: mit
pretty_name: CCORE Three.js 3D Scenes
language:
- en
task_categories:
- text-generation
size_categories:
- 1K<n<10K
tags:
- 3d
- threejs
- code
- reasoning
- conversational
- html
- javascript
configs:
- config_name: default
data_files:
- split: train
path:
- data/train-00000.parquet
- data/train-00001.parquet
- data/train-00002.parquet
- data/train-00003.parquet
- split: validation
path:
- data/validation-00000.parquet
default: true
- config_name: previews
data_files:
- split: train
path:
- previews/train-00000.parquet
- previews/train-00001.parquet
- previews/train-00002.parquet
- previews/train-00003.parquet
- previews/train-00004.parquet
- previews/train-00005.parquet
- previews/train-00006.parquet
- previews/train-00007.parquet
- previews/train-00008.parquet
- previews/train-00009.parquet
- previews/train-00010.parquet
- previews/train-00011.parquet
- previews/train-00012.parquet
- previews/train-00013.parquet
- previews/train-00014.parquet
- previews/train-00015.parquet
- split: validation
path:
- previews/validation-00000.parquet
dataset_info:
- config_name: default
features:
- name: id
dtype: string
- name: title
dtype: string
- name: level
dtype: int32
- name: domain
dtype: string
- name: voxel
dtype: bool
- name: messages
list:
- name: role
dtype: string
- name: content
dtype: string
- name: reasoning_content
dtype: string
splits:
- name: train
num_examples: 3933
num_bytes: 363885742
- name: validation
num_examples: 213
num_bytes: 19822852
download_size: 134444227
dataset_size: 383708594
- config_name: previews
features:
- name: id
dtype: string
- name: title
dtype: string
- name: level
dtype: int32
- name: image
dtype: image
splits:
- name: train
num_examples: 3933
num_bytes: 1294921572
- name: validation
num_examples: 213
num_bytes: 72048979
download_size: 1366958111
dataset_size: 1366970551
CCORE Three.js 3D Scenes
4,146 high-quality 3D scene examples with prompts, full thinking traces, and complete HTML/JavaScript responses.
The scene code and thinking traces were generated with MiniMax M3.1 Flash.
This dataset brings together a broad range of Three.js scenes, from individual objects and detailed interiors to architecture, landscapes, animated mechanisms, and voxel environments. It is intended for training and evaluating models that generate 3D scenes from natural-language instructions.
This dataset was used to train Qwen3.5-4B-3D-Coder.
Scene examples
The images below are rendered references from examples included in this dataset. A matching reference image is available for every example in the optional previews configuration.
 Dragonrider Aerie scene-000013 |
 Temple Garden scene-003268 |
 Bookbinding Bench scene-002556 |
 Canal Lock and Weir scene-000597 |
 Medieval Cathedral and Builders' Quarter — Voxel Study scene-003894 |
 Power-Station Decommissioning Exhibit scene-003933 |
 Optical Telescope Dome Cutaway scene-003559 |
|
Dataset contents
| Split | Scene examples |
|---|---|
| Train | 3,933 |
| Validation | 213 |
| Total | 4,146 |
The original training and validation split is preserved. Each example includes its full prompt, thinking trace, and response, without truncation. Complexity levels range from 1 to 10 and are annotations, rather than model benchmark scores.
Format
The default configuration contains conversational data in Parquet format, ready to load with Hugging Face Datasets.
| Field | Description |
|---|---|
id |
Stable scene identifier, shared with the reference-image configuration. |
title |
Human-readable scene title. |
level |
Annotated scene complexity, from 1 to 10. |
domain |
Scene subject category. |
voxel |
Whether the scene uses a voxel style. |
messages |
Conversation with an optional system message, a user prompt, and an assistant response. |
Within the assistant message, reasoning_content contains the thinking trace and content contains the full response, including the HTML/JavaScript code. Other messages have a null reasoning_content.
The optional previews configuration contains id, title, level, and an image feature for each scene. It uses the same splits and row order as the default configuration. Load it separately when you want to inspect or reuse the visual references.
Usage
from datasets import load_dataset
dataset = load_dataset("ccore/ThreeJS-3D-Scenes")
sample = dataset["train"][0]
messages = sample["messages"]
thinking = messages[-1]["reasoning_content"]
response = messages[-1]["content"]
level = sample["level"]
# Optional rendered references, aligned by id and split.
previews = load_dataset("ccore/ThreeJS-3D-Scenes", "previews")
assert previews["train"][0]["id"] == sample["id"]
image = previews["train"][0]["image"]
The reference images show the included target scenes. They are separate from the text conversations. Some scenes contain animation or interactive controls that a single screenshot cannot capture.
License
This dataset is released under the MIT License, to the extent of the rights held by the dataset publisher.
What’s next
I’m preparing a companion dataset for GLB-3D-Coder, following a similar approach and with roughly the same number of examples as this release. Its focus is on generating reusable 3D assets with GLB export, ready to bring into game editors such as Godot.
Beyond that, I’m planning a larger dataset of 100,000 high-quality examples, with context lengths ranging from 8K to 128K tokens, targeting a training run roughly 15× larger than the current GLB model’s run. The first goal is reliable one-shot asset generation, followed by a version designed to work directly inside an agent harness, generating, testing, and refining assets.
As with my previous releases, I plan to share the datasets and model weights openly. Community support would help fund the data generation, validation, and GPU compute needed to make these next steps possible.
❤️ Support Future Open-Source AI Models
I'm independently developing and releasing specialized AI models and datasets for 3D generation, procedural environments, and creative coding, with the goal of making these tools accessible to everyone.
Training new models, generating high-quality datasets, validating thousands of examples, and running experiments require significant GPU resources, time, and effort.
If you enjoy this project and would like to help me train more models, expand the datasets, and explore new capabilities, any contribution is greatly appreciated!
My goal is to keep sharing models, weights, datasets, and improvements openly with the community.
☕ Support via Ko-fi
Crypto Donations
Bitcoin (BTC) — Bitcoin Network (SegWit)
bc1qr2gnr9g0uzk8rmsq0k6xpjgg4l7r0rfyrq69e6
Ethereum (ETH) — Ethereum Mainnet (ERC20)
0xa17608FCdeAd9e717Cd42e04127175527b4E75CB
Solana (SOL / USDC / USDT) — Solana Network
GhNmYmezKi1uWGKZeVxgddjiRW2rYNHZrf84ZaMcnha
Every contribution helps fund more GPU compute, better datasets, new training runs, and future public releases.
Thank you for supporting independent open-source AI research!