---
language:
- en
pretty_name: CoCode-World Data
size_categories:
- 1K<n<10K
tags:
- video
- world-model
- multi-view
- multi-agent
- proxy-rendering
- game-world
configs:
- config_name: train
  data_files:
  - split: train
    path: train/train-*.tar
---

# CoCode-World Data

**Synchronized multi-view video data for multi-player world models**

[Project Page](https://iapple233.github.io/CCW/) · [Code](https://github.com/IApple233/CoCode-World) · [Model](https://huggingface.co/iapple1/CoCode-World-Model)

## Dataset overview

CoCode-World Data contains **5,000 paired training examples** for synchronized two-player video generation. Each example includes a target RGB video, a corresponding 3D proxy video, an English caption, and frame alignment metadata.

The release also provides scene geometry, player actions, camera parameters, world-state observations, and a proxy rendering toolkit for exploring alternative scene representations.

## Dataset structure

| Property | Description |
|---|---|
| Split | Training |
| Examples | 5,000 |
| Camera layout | Two views arranged side by side |
| View order | `agent_000` on the left; `agent_001` on the right |
| Resolution | 2560 × 720; 1280 × 720 per view |
| Duration | 124 frames at 24 FPS, approximately 5.17 seconds |
| Captions | English scene descriptions and per-view timelines |
| Storage | 25 TAR shards, each containing 200 examples |

Each example contains the following files:

```text
<sample_id>.proxy.mp4
<sample_id>.target.mp4
<sample_id>.caption_v1.txt
<sample_id>.alignment.json
<sample_id>.json
```

The sample index is available as `manifests/train.jsonl` and `manifests/train.parquet`. It links each example to its shard, media, captions, and metadata.

## Download and use

```bash
pip install -U huggingface_hub
hf download iapple1/CoCode-World-Data --repo-type dataset \
  --include 'train/*' 'manifests/train.jsonl' \
  --local-dir datasets/CoCode-World-Data
```

The [usage guide](docs/USAGE.md) provides extraction and streaming examples. To prepare the data for CoCode-World training, use the official [code repository](https://github.com/IApple233/CoCode-World):

```bash
python scripts/prepare_training_data.py \
  --dataset-dir datasets/CoCode-World-Data \
  --output-dir datasets/prepared
```

See the [training guide](https://github.com/IApple233/CoCode-World/blob/main/docs/TRAINING.md) for caching and model training.

## Scene metadata and rendering

Additional metadata covers 21 recording episodes and includes scene geometry, player controls, camera trajectories, world-state observations, and frame mappings. The accompanying proxy rendering toolkit supports alternative colors, materials, and geometric representations.

See the [metadata overview](metadata/README.md) and [documentation](docs/) for the available files and reconstruction workflow. Match the two views using the recorded timestamps and alignment records, and preserve their left/right camera assignment.

## Collection and selection

The release is a selected subset of a larger paired-video collection. Examples were selected after automated quality screening, with coverage balanced across recording episodes. The target videos, proxy videos, and English captions are provided as recorded. The release is intended for training; the separate [evaluation examples](https://github.com/IApple233/CoCode-World/tree/main/examples/evaluation) are provided in the code repository.

## Intended use and limitations

The dataset supports research on multi-player world models, multi-view video generation, and scene-conditioned rendering. Automated screening does not guarantee that every example is free of artifacts. Scene metadata describes recorded trajectories and actions; it does not provide a simulator for arbitrary new action sequences.

## Data rights

Game-derived videos and assets remain subject to the rights and terms of their respective owners. Software included with the dataset retains its accompanying license. Users are responsible for ensuring that their intended use complies with the applicable terms.