README.md

---
license: apache-2.0
task_categories:
  - video-text-to-text
  - robotics
language:
  - en
tags:
  - video
  - reasoning
  - perception
  - embodied
  - simulation
  - webdataset
size_categories:
  - 100K<n<1M
---

# VideoReason Training Dataset

A large-scale video reasoning training dataset spanning perception, simulation, and embodied tasks.

## Dataset Overview

| Subset | Samples | Description |
|--------|---------|-------------|
| **perception** | 177,407 | Visual perception tasks |
| **simulation** | 105,818 | 3D scene navigation with camera motion sequences in simulated environments |
| **embodied** | 188,845 | Robotic manipulation tasks |
| **Total** | **472,070** | |

Each sample consists of a video (`.mp4`) paired with a text prompt describing the task.

### Perception Breakdown

| Task | Samples | Description |
|------|---------|-------------|
| Segmentation | 40,000 | Instance segmentation with color-fill visualization |
| Denoising | 32,093 | Image denoising from noisy inputs |
| Low-light Enhancement | 32,103 | Enhancing images captured in low-light conditions |
| Super-resolution | 32,090 | Single-image super-resolution |
| Edge Detection | 30,000 | Boundary and edge extraction |
| Keypoint Detection | 11,121 | Detecting structural keypoints |

### Embodied Breakdown

| Source | Samples | Description |
|--------|---------|-------------|
| DROID | 94,237 | Real-world robotic manipulation from the [DROID dataset](https://droid-dataset.github.io/) |
| RoboTwin | 94,608 | Simulated bimanual robotic manipulation from [RoboTwin](https://robotwin-benchmark.github.io/early-version/) |

## Repository Structure

```
README.md
{subset}/prompts.jsonl          # Metadata: one JSON object per line
{subset}/shards/shard-NNNNNN.tar  # WebDataset tar shards (~1GB each)
```

### prompts.jsonl format

```json
{"video_path": "video/seg/example.mp4", "prompt": "Identify a dining table in this image..."}
```

- `video_path`: original relative path of the video within the subset
- `prompt`: text prompt associated with the video

### Tar shard contents

Each shard is a standard `.tar` archive containing paired files:

```
000000.mp4    # video file
000000.json   # {"video_path": "...", "prompt": "..."}
000001.mp4
000001.json
...
```

The 6-digit key is the global sample index within the subset (matching the line number in `prompts.jsonl`, 0-indexed).

## Download

### From ModelScope

```bash
pip install modelscope

# Download entire dataset
modelscope download --dataset ZaneQiu/XVReason --local_dir videoreason-training

# Download a specific subset
modelscope download --dataset ZaneQiu/XVReason --local_dir videoreason-training --include "perception/*"
```

### From HuggingFace

```bash
pip install huggingface_hub

huggingface-cli download Zane-QIU/videoreason-training --repo-type dataset --local-dir videoreason-training
```

## Extracting Tar Shards

### Option 1: Extract all shards to a directory

```bash
# Extract all shards for a subset
mkdir -p extracted/perception
for f in videoreason-training/perception/shards/shard-*.tar; do
    tar xf "$f" -C extracted/perception
done
```

### Option 2: Use WebDataset for streaming (recommended for training)

```python
import webdataset as wds

dataset = wds.WebDataset("videoreason-training/perception/shards/shard-{000000..000028}.tar")
for sample in dataset:
    video_bytes = sample[".mp4"]
    metadata = json.loads(sample[".json"])
    print(metadata["prompt"])
```

### Option 3: Extract a single shard

```bash
tar xf perception/shards/shard-000000.tar -C output_dir/
```

## License

Please refer to the original dataset licenses for each subset:
- **DROID**: [DROID Dataset](https://droid-dataset.github.io/)
- **RoboTwin**: [RoboTwin](https://robotwin-benchmark.github.io/early-version/)