---
title: Giga-World-1-Toydata
canonical_url: "https://www.modelscope.cn/datasets/GigaAI/Giga-World-1-Toydata"
md_url: "https://www.modelscope.cn/datasets/GigaAI/Giga-World-1-Toydata.md"
repository: GigaAI/Giga-World-1-Toydata
last_updated: 2026-07-30
license: apache-2.0
storage_size: "1.9 GB"
downloads: 1783
stars: 0
---
# Giga-World-1-Toydata
> Giga-World-1-Toydata - GigaAI 在 ModelScope 开源的数据集。Giga-World-1 Example Data
GigaAI/Giga-World-1-Toydata 是 ModelScope 魔搭社区上的数据集，存储大小 1.9 GB，采用 apache-2.0 许可。
- **Repository**: GigaAI/Giga-World-1-Toydata
- **License**: apache-2.0
- **Storage size**: 1.9 GB
- **Downloads**: 1783
- **Stars**: 0
- **Last updated**: 2026-07-30
Source: https://www.modelscope.cn/datasets/GigaAI/Giga-World-1-Toydata
---
# Giga-World-1 Example Data
This repository contains example assets and toy datasets for the paper [GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation](https://huggingface.co/papers/2607.02642).
- **Project Page:** https://open-gigaai.github.io/giga-world-1/
- **GitHub Repository:** https://github.com/open-gigaai/giga-world-1
This directory provides example assets and toy datasets for testing the GigaWorld-1 inference, data pipeline, and model training workflow.
## Directory Structure
```text
example/
├── infer_assest/ # Inference / rollout assets
│ ├── control_video.mp4
│ └── input_image.png
├── toy_datapipeline_dataset/ # Raw LeRobot-format toy dataset
│ ├── gt/
│ ├── depth/
│ ├── plucker/
│ ├── sketch/
│ └── labels/
└── toy_train_dataset/ # Model training data
├── nano/
│ ├── dataset_cache.pkl
│ └── episode_*.pt
└── pro/
├── dataset_cache.pkl
└── episode_*.pt
```
## Contents
### `infer_assest/` — Inference Assets
Assets used for quick testing of the image-to-video inference / rollout pipeline.
- `input_image.png`: input image (used as the first frame) for image-to-video generation.
- `control_video.mp4`: control video used by the example inference pipelines.
### `toy_datapipeline_dataset/` — Raw LeRobot Dataset
A compact robot video dataset shipped in the **raw LeRobot data format**, which can be visualized directly with the [LeRobot](https://github.com/huggingface/lerobot) visualization tools.
Dataset structure:
```text
example/toy_datapipeline_dataset/
├── gt/ # RGB videos (ground truth)
│ ├── cam_high/ # head view
│ ├── cam_left_wrist/ # left wrist view
│ └── cam_right_wrist/ # right wrist view
├── depth/ # Depth Anything V2 outputs
│ ├── cam_high/
│ ├── cam_left_wrist/
│ └── cam_right_wrist/
├── plucker/ # Plücker coordinate control signals (left/right per view)
│ ├── episode_000001_left_direction.mp4
│ ├── episode_000001_left_moment.mp4
│ ├── episode_000001_right_direction.mp4
│ └── episode_000001_right_moment.mp4
├── sketch/ # sketch control signals
│ └── cam_high/
└── labels/
├── data.pkl # per-episode metadata
└── config.json
```
Each record in `labels/data.pkl` follows this structure:
```python
{
"action": List[List[float]], # end-effector / joint actions
"data_index": int,
"episode_name": str,
"cam_high_video_path": str,
"cam_left_wrist_video_path": str,
"cam_right_wrist_video_path": str,
"cam_high_depth_path": str,
"cam_left_wrist_depth_path": str,
"cam_right_wrist_depth_path": str,
"qpos": List[List[float]], # current joint angles
"video_height": int,
"video_width": int,
"video_length": int,
"short-prompt": { # from meta/episodes.jsonl
"task1": {
"start_idx": "0",
"end_idx": "299",
"description": "put banana into basket"
}
},
"long-prompt": { # generated by Qwen3-VL on cam_high
"long prompt 1": {
"start_idx": "0",
"end_idx": "299",
"caption": "The robot arm reaches toward ..."
}
}
}
```
### `toy_train_dataset/` — Model Training Data
Model training data for the GigaWorld-1 training workflow. This dataset is prepared for directly validating the model training pipeline on a small-scale example.
Dataset structure:
```text
example/toy_train_dataset/
├── nano/ # nano-scale toy training split
│ ├── dataset_cache.pkl # cached dataset index / metadata for fast loading
│ ├── episode_000000_4834c0369d_s000000_e000129_0-129_121_480_1920.pt
│ ├── episode_000000_4834c0369d_s000129_e000258_0-129_121_480_1920.pt
│ └── ...
└── pro/ # pro-scale toy training split
├── dataset_cache.pkl # cached dataset index / metadata for fast loading
├── episode_000000_453e4c570c_s000000_e000129_0-129_121_480_1920.pt
├── episode_000001_7473d389a6_s000000_e000129_0-129_121_480_1920.pt
├── episode_000002_8436313f94_s000000_e000129_0-129_121_480_1920.pt
└── ...
```
- `nano/`: a smaller toy training split for quick debugging and smoke tests.
- `pro/`: a larger toy training split for validating the full training data loader.
- `dataset_cache.pkl`: cached metadata / dataset index used by the training pipeline.
- `episode_*.pt`: preprocessed training samples. The filename records the source episode id, segment range, frame range, and spatial resolution.
## Usage
Use these files with the main GigaWorld-1 repository scripts.
Example paths:
```bash
EXAMPLE_ROOT=/path/to/example
INFER_ASSETS=$EXAMPLE_ROOT/infer_assest
TOY_PIPELINE_DATASET=$EXAMPLE_ROOT/toy_datapipeline_dataset
TOY_TRAIN_DATASET=$EXAMPLE_ROOT/toy_train_dataset
```
Refer to the main project README for detailed commands covering data preparation, inference, visualization, and training.
## Acknowledgements
We sincerely thank the open-source community and the projects that make this work possible.
Thanks also to many other open-source contributors for their tools, models, and community support.
## Citation
```bibtex
@article{gigaworld2025,
title = {GigaWorld-1: A Roadmap to World Models for Robot Policy Evaluation},
author = {{GigaAI}},
journal = {arXiv preprint},
year = {2025},
eprint = {2511.19861},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}
```
## License
This example data is released under the Apache License 2.0 unless otherwise specified.