REVISOR-25k

A multi-task video understanding dataset for training video LLMs with reinforcement learning (GRPO). The dataset contains ~25k samples spanning Video QA and Temporal Grounding tasks.

Dataset Structure

The dataset is organized into 4 subsets:

Total: 25,359 samples

Data Fields

Usage

from datasets import load_dataset

# Load all subsets
ds = load_dataset("YOUR_USERNAME/REVISOR-25k")

# Load a specific subset
video_qa = load_dataset("YOUR_USERNAME/REVISOR-25k", split="video_r1")
temporal = load_dataset("YOUR_USERNAME/REVISOR-25k", split="time_r1")

Directory Structure

REVISOR-25k/
├── README.md
├── data/
│   ├── video_r1/train.parquet
│   ├── time_r1/train.parquet
│   ├── cg_bench/train.parquet
│   └── rextime/train.parquet
└── videos/
    ├── video_r1/
    ├── time_r1/
    ├── cg_bench/
    └── rextime/

Citation

If you use this dataset, please cite:

@inproceedings{li2026revisor,
  title={Revisor: Beyond textual reflection, towards multimodal introspective reasoning in long-form video understanding},
  author={Li, Jiaze and Yin, Hao and Tan, Wenhui and Chen, Jingyang and Xu, Boshen and Qu, Yuxun and Chen, Yijing and Ju, Jianzhong and Luo, Zhenbo and Luan, Jian},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5059--5069},
  year={2026}
}