simuler/GraftSet-Video

Introduction

GraftSet-Video is the external video exemplar dataset introduced in GraftOOD: Grafting Out-of-Distribution Exemplars into Advantage-Weighted Rectification for Video Reward Optimization. It provides prompt-matched video sequences for augmenting online video reward optimization with direct exemplar supervision.

Dataset overview

Property	Value
Videos	9,588
Unique prompts	2,397
Videos per prompt	4
Source model identifiers	9
Video resolution	832 × 480 pixels (width × height)
Frames per video	81
Frame rate	16 fps
Stored reward	HPSv3 score
Mean / standard deviation of reward	9.735 / 1.413

Source composition

model_name records the video generator identifier.

model_name	Videos
wan2_2	2,164
vidu_q2	1,163
seedance_1_5	1,144
kling_2_5	1,098
kling3	915
seedance2	911
veo3_1	883
hailuo	830
sora2	480
Total	9,588

Repository structure

.
├── README.md
├── metadata.jsonl
└── videos/
    ├── hailuo/
    ├── kling3/
    ├── kling_2_5/
    ├── seedance2/
    ├── seedance_1_5/
    ├── sora2/
    ├── veo3_1/
    ├── vidu_q2/
    └── wan2_2/
        └── 1_wan2_2.mp4

Metadata schema

Field	Type	Description
file_name	string	Video path relative to the directory containing metadata.jsonl, such as videos/wan2_2/1_wan2_2.mp4.
prompt_id	string	Shared identifier for the four videos belonging to one prompt.
prompt	string	Original text prompt paired with the video.
model_name	string	Source generator identifier.
reward	float	Original offline HPSv3 score, Higher values indicate higher scores under this reward model.

Example record, formatted across multiple lines for readability:

{
  "file_name": "videos/wan2_2/1_wan2_2.mp4",
  "prompt_id": "398b3a665bd230e5",
  "prompt": "The camera starts on a vibrant, diamond-shaped kite dancing against a brilliant blue sky. A slow tilt down reveals a young daughter in a sun-hat and a pink dress, her face lit with joy as she expertly handles the kite string in a sprawling green park. Her laughter rings out as the kite soars higher. Later, the scene shifts to her sitting on a picnic blanket with her family. The camera focuses on a close-up of hands cutting a large, juicy watermelon into perfect triangular slices, ready to be shared on this perfect summer day.",
  "model_name": "wan2_2",
  "reward": 9.292619972229003
}

Loading the dataset

Run the following from the dataset repository root. It uses only the Python standard library:

import json
from collections import defaultdict
from pathlib import Path

root = Path.cwd()
with (root / "metadata.jsonl").open(encoding="utf-8") as f:
    records = [json.loads(line) for line in f if line.strip()]

by_prompt = defaultdict(list)
for record in records:
    video_path = root / record["file_name"]
    assert video_path.is_file(), video_path
    by_prompt[record["prompt_id"]].append(record)

assert len(records) == 9588
assert len(by_prompt) == 2397
assert all(len(group) == 4 for group in by_prompt.values())

example = records[0]
print(example["prompt"])
print(root / example["file_name"])
print(example["reward"])

Load with Hugging Face Datasets

Install the library:

pip install -U datasets

From the dataset repository root:

from pathlib import Path
from datasets import Video, load_dataset

dataset = load_dataset(
    "videofolder",
    data_dir=str(Path.cwd()),
    split="train",
    drop_labels=True,
)
dataset = dataset.cast_column("video", Video(decode=False))
print(dataset[0])

VideoFolder associates file_name with a video feature and retains the additional metadata columns. decode=False returns the underlying video reference without decoding frames. For decoded video access, follow the Hugging Face video-loading documentation and install its required decoder dependencies.

Once this repository is published on the Hub, replace YOUR_NAMESPACE/GraftSet-Video below with its actual repository ID:

from datasets import Video, load_dataset

dataset = load_dataset(
    "YOUR_NAMESPACE/GraftSet-Video",
    split="train",
    drop_labels=True,
)
dataset = dataset.cast_column("video", Video(decode=False))

Paper and citation

If you find this project useful, please consider citing:

@article{fu2026graftood,
  title={GraftOOD: Grafting Out-of-Distribution Exemplars into Advantage-Weighted Rectification for Video Reward Optimization},
  author={Fu, Zheming and He, Ruizhe and Ning, Wenbin and Ma, Xiaoxiao and Yang, Xuan and Mu, Lianrui and Huang, Haoyang and Liu, Chang and Wang, Lei and Fu, Siming},
  journal={arXiv preprint arXiv:2610.13470},
  year={2026}
}
@article{fu2026videoawr,
  title={VideoAWR: Reinforcing Video Flow Models with Decoupled-Timestep Advantage-Weighted Rectification},
  author={Fu, Zheming and Ning, Wenbin and He, Ruizhe and Shang, Wei and Ma, Xiaoxiao and Mu, Lianrui and Liu, Chang and Wang, Lei and Fu, Siming},
  journal={arXiv preprint arXiv:2610.13338},
  year={2026}
}
@article{fu2026flowawr,
  title={FlowAWR: Online Adaptive Flow Reinforcement via Advantage-Weighted Rectification},
  author={Fu, Zheming and He, Ruizhe and Shang, Wei and Ma, Xiaoxiao and Wang, Lei and Liu, Chang and Fu, Siming},
  journal={arXiv preprint arXiv:2606.30376},
  year={2026}
}

Acknowledgments

We acknowledge the authors and maintainers of the source video reward-model resources, the video generation systems represented in this release, and HPSv3. Full bibliographic references for these resources are provided in the accompanying paper.