---
language:
- en
license: apache-2.0
size_categories:
- 10M<n<100M
task_categories:
- image-text-to-text
- video-text-to-text
- visual-question-answering
tags:
- llava
- multimodal
- video
- spatial-reasoning
configs:
- config_name: viewer_caption_30s
  data_files:
  - split: preview
    path: viewer/caption_30s.parquet
- config_name: viewer_caption_60s
  data_files:
  - split: preview
    path: viewer/caption_60s.parquet
- config_name: viewer_caption_180s
  data_files:
  - split: preview
    path: viewer/caption_180s.parquet
- config_name: viewer_caption_gt10min
  data_files:
  - split: preview
    path: viewer/caption_gt10min.parquet
- config_name: viewer_spatial
  data_files:
  - split: preview
    path: viewer/spatial.parquet
---

# LLaVA-OneVision-2-Data

[**Project Page**](https://evolvinglmms-lab.github.io/LLaVA-OneVision-2/) | [**Paper**](https://huggingface.co/papers/2605.25979) | [**GitHub**](https://github.com/EvolvingLMMs-Lab/LLaVA-OneVision-2)

Training data for the LLaVA-OneVision-2 multimodal model family, covering large-scale video and spatial reasoning corpora used in mid-training.

## Dataset Composition

| Subset | Format | Description |
|---|---|---|
| `mid_training_video/60s_rest/` | WebDataset (`.tar`) | 10,809 shards of ~60s video clips |
| `mid_training_video/caption_v0/split_30s.jsonl` | JSONL | Captions for 30-second video clips |
| `mid_training_video/caption_v0/split_60s.jsonl` | JSONL | Captions for 60-second video clips |
| `mid_training_video/caption_v0/split_180s.jsonl` | JSONL | Captions for 180-second video clips |
| `mid_training_video/caption_v0/split_gt10min.jsonl` | JSONL | Captions for >10-minute video clips |
| `spatial/` | WebDataset (`.tar`) | 84 shards of spatial reasoning data (refcoco, visual genome, pointing, 3D, etc.) |
| `mid_training_video/mapping/mapping_{5s,10s,30s,60s,180s,gt10min}.csv` | CSV | Maps each video clip's `dst_path` to its source `youtube_id` and `[start_time, end_time]` window |

## Preview Configs

The `viewer_*` configs above expose small Parquet samples so the Hugging Face Dataset Viewer can render the data directly in the browser:

- **`viewer_caption_30s`** — 5 caption samples from 30-second clips
- **`viewer_caption_60s`** — 5 caption samples from 60-second clips
- **`viewer_caption_180s`** — 3 caption samples from 180-second clips
- **`viewer_caption_gt10min`** — 1 caption sample from >10-minute clips
- **`viewer_spatial`** — 10 spatial-reasoning samples with embedded thumbnail images, mixed across tasks (refcoco, visual genome, pointing, ca1m, osd, crosspoint, erqa, roborefer)

These previews are intended for **schema inspection only**. For training, use the full `mid_training_video/` and `spatial/` shards.

## Sample Usage

You can download the dataset using the `huggingface-cli`:

```bash
huggingface-cli download --repo-type dataset --resume-download \
    --local-dir ./LLaVA-OneVision-2-Data \
    mvp-lab/LLaVA-OneVision-2-Data
```

## Citation

If you find *LLaVA-OneVision-2* useful in your research, please consider citing the following papers:

```bibtex
@inproceedings{LLaVA-OneVision-2,
  title={LLaVA-OneVision-2},
  author={llava-onevision contributors},
  booktitle={arXiv},
  year={2026}
}

@inproceedings{LLaVA-OneVision-1.5,
  title={LLaVA-OneVision-1.5: Fully Open Framework for Democratized Multimodal Training},
  author={An, Xiang and Xie, Yin and Yang, Kaicheng and Zhang, Wenkang and Zhao, Xiuwei and Cheng, Zheng and Wang, Yirui and Xu, Songcen and Chen, Changrui and Wu, Chunsheng and Tan, Huajie and Li, Chunyuan and Yang, Jing and Yu, Jie and Wang, Xiyao and Qin, Bin and Wang, Yumeng and Yan, Zizhen and Feng, Ziyong and Liu, Ziwei and Li, Bo and Deng, Jiankang},
  booktitle={arXiv},
  year={2025}
 }
```