hunarbatra/4DReasoner-SFT-Data
收藏Hugging Face2026-03-22 更新2026-03-29 收录
下载链接：
https://hf-mirror.com/datasets/hunarbatra/4DReasoner-SFT-Data
下载链接
链接失效反馈官方服务：
资源简介：
---
license: apache-2.0
task_categories:
- video-text-to-text
- visual-question-answering
tags:
- 4d-reasoning
- spatial-reasoning
- video-understanding
- sft
- sharegpt
size_categories:
- 10K<n<100K
---
# 4DReasoner SFT Data
Supervised fine-tuning dataset for 4D (spatial + temporal) video reasoning.
## Overview
| Source | Samples |
|--------|---------|
| spacer | 91,315 |
| dsr | 55,182 |
| cosmos | 2,378 |
| **Total** | **148,875** |
## Format
Each sample is in **ShareGPT** format compatible with [LlamaFactory](https://github.com/hiyouga/LLaMA-Factory):
```json
{
"messages": [
{"role": "user", "content": "<video>\nQuestion ..."},
{"role": "assistant", "content": "Answer ..."}
],
"videos": ["video_filename.mp4"],
"metadata": {...}
}
```
## Sources
- **SpaceR (SR-91k)**: Spatial reasoning over 3D scene videos (ScanNet).
- **DSR**: Dynamic spatial reasoning over real-world videos (Koala36M).
- **Cosmos**: Physical commonsense reasoning with video clips.
## Usage with LlamaFactory
Copy `dataset_info.json` into your LlamaFactory `data/` directory and reference
`4dreasoner_sft` as the dataset name in your training config.
提供机构：
hunarbatra
