VBVR-Pro-SFT-Video / README.md
junxiangjoe's picture
Align badges with scorer README
fc4ed0f verified
|
Raw
History Blame Contribute Delete
8.1 kB
metadata
license: cc-by-nc-4.0
language:
  - en
tags:
  - video-generation
  - image-to-video
  - visual-reasoning
  - training
pretty_name: VBVR-Pro-SFT-Video
size_categories:
  - 1M<n<10M
configs:
  - config_name: preview
    default: true
    data_files:
      - split: train
        path: preview/video.parquet

VBVR-Pro-SFT-Video

The video (I2V) supervised-fine-tuning split of VBVR-Pro: 1.24M programmatically generated reasoning instances across 250 parameterized tasks, one tar.gz per task.

At a glance

Property Value
Tasks 250
Instances 1,250,000 (5,000 per task)
Archives 250 tar.gz, one per task
Total size 76.6 GB
Resolution 512 × 512
Video codec H.264, CRF 20

Layout

.
├── tars/
│   ├── G-11_handle_object_reappearance_data-generator.tar.gz
│   └── …                                     # 250 archives, one per task
├── jsonl/
│   ├── G-11_handle_object_reappearance_data-generator.jsonl
│   └── …                                     # 250 files, one per task
└── meta_video_train.json                     # index over the 250 tasks

Each archive holds one task and extracts to:

G-11_handle_object_reappearance_data-generator/                     # task
└── handle_object_reappearance_task/                                # subtask
    ├── handle_object_reappearance_00000000/                        # sample
    │   ├── first_frame.png          # conditioning image  (512 × 512)
    │   ├── metadata.json            # task parameters, ground truth, scoring contract
    │   └── video/
    │       ├── prompt.txt           # instruction
    │       ├── ground_truth.mp4     # reference video     (16 fps, H.264)
    │       └── final_frame.png      # last frame of the reference video
    ├── handle_object_reappearance_00000001/                        # same files
    ├── …
    └── handle_object_reappearance_00004999/                        # 5,000 samples per task

Sample ids run 0000000000004999.

All 250 archives share this shape and no two overlap, so they can be extracted into one directory.

Index files

meta_video_train.json maps every task to its jsonl:

{
  "G-11_handle_object_reappearance_data-generator": {
    "root": ".",
    "annotation": "jsonl/G-11_handle_object_reappearance_data-generator.jsonl",
    "length": 5000,
    "task": "Video-SFT"
  }
}

root is the directory the archives were extracted into; length is that task's row count. Each jsonl row points at one sample, with every path relative to root:

Field Meaning
sample_id e.g. handle_object_reappearance_00000000
first_frame conditioning image
metadata that sample's metadata.json
clip_path reference video
final_frame last frame of the reference video
prompt instruction

Usage

huggingface-cli download Video-Reason/VBVR-Pro-SFT-Image --repo-type dataset --local-dir data/VBVR-Pro-SFT-Image
huggingface-cli download Video-Reason/VBVR-Pro-SFT-Video --repo-type dataset --local-dir data/VBVR-Pro-SFT-Video

The VBVR-Pro training code unpacks the archives and emits the per-trainer manifests in one step:

python training/prepare_data.py \
  --image-archives data/VBVR-Pro-SFT-Image \
  --video-archives data/VBVR-Pro-SFT-Video \
  --output-dir data/prepared

To use the data directly instead, extract every archive into one directory and point root at it:

mkdir -p data/extracted
for f in data/VBVR-Pro-SFT-Video/tars/*.tar.gz; do tar xzf "$f" -C data/extracted; done

License

VBVR-Pro source code, scripts, configuration files and task-specific scoring software — including everything in this repository — are licensed under the Apache License 2.0. VBVR-Pro data and benchmark materials are separately licensed under CC BY-NC 4.0. Model weights and third-party materials remain subject to their applicable model-card and upstream terms. See LICENSE.md for details.

Citation

@misc{xu2026vbvrproscalableverifiablesuite,
      title={VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning},
      author={Junxiang Xu and Ruisi Wang and Fanyi Pu and Maijunxian Wang and Ran Ji and Tongxi Zhou and Chenyang Gu and Jing Zuo and Hongcan Xiao and Yimeng Geng and Wanqi Yin and Wei Chen and Oscar Qian and Zhengan Yan and Ziqi Huang and Haiwen Diao and Liang Pan and Bo Li and Xiangyu Fan and Dezhi Luo and Fengyuan Yu and Zehong Zhao and Qingying Gao and Tinghui Zhu and Yilan Zhang and Jingqi Tong and Pinyuan Feng and Zhengze Jiang and Letian Wang and Ziyu Guo and Renrui Zhang and Jieneng Chen and Sonia Joseph and Constantin Venhoff and Saman Motamed and Mengyue Yang and Chandra Sripada and Alan Yuille and Philip Torr and Lvmin Zhang and Vikash Kumar and Daniel Khashabi and Nikolaus Kriegeskorte and Rapha\"{e}l Milli\`{e}re and Vincent C. M\"{u}ller and Anyi Rao and Quan Wang and Ziwei Liu and Dahua Lin and Lei Yang and Hokin Deng and Zhongang Cai},
      year={2026},
      eprint={2608.26105},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.26105},
}