helloworldzzr/Video-Moderation-4225

Video Moderation 4225

This is the fully materialized cleaned dataset used by March-77/video-moderation-vlm to train a Qwen3-VL-2B binary content-moderation adapter.

Sensitive-content warning: the media includes sexual, nudity, violence, disturbing imagery, dangerous behavior, and other harmful-content examples. Use only in a controlled environment for lawful content-safety research.

The project maintainer states that permission was obtained from the original authors to redistribute the media in this cleaned release. No authorization letter or license identifier is fabricated in this repository. Original source terms and restrictions remain applicable; see TERMS.md.

Dataset summary

Sources

SexTok and SafeSora media were not used in the released model and are not included in this dataset. SafeWatch-Bench was not available during preparation.

Labels and prompt

• —A: ALLOW — no obvious policy violation or harm.
• —B: BLOCK — includes pornographic nudity, sexual suggestion or sexualized display, graphic violence, dangerous behavior, prohibited items, and other mapped harmful content.

Every row includes the same Chinese policy prompt used by the classifier. The prompt is an instruction, not a video caption. The model uses actual video frames or image pixels. risk_tags are retained for audit and sliced evaluation but are not inputs to the binary training loss.

File layout

.
├─ train.json
├─ val.json
├─ test.json
├─ train_video_only.json
├─ val_video_only.json
├─ test_video_only.json
├─ dataset_summary.json
├─ audit/
│  ├─ dataset_manifest.jsonl
│  ├─ ugvd_label_map.csv
│  ├─ vhd_filename_map.csv
│  └─ harmful_contents_selection.csv
└─ media/
   ├─ ugvd/*.mp4
   ├─ vhd11k/*.mp4
   └─ harmful_contents/{train,val,test}/*

Paths in JSON rows are relative to media/, for example:

{
  "id": "vhd11k_0581",
  "video": "vhd11k/vhd_0581.mp4",
  "prompt": "请判断该内容是否需要安全拦截……",
  "label": "B",
  "source_dataset": "VHD11K",
  "risk_tags": ["violence_weapons"]
}

Cleaning and split procedure

• —Seed: 42.
• —UGVD: deterministic iterative split balancing A/B and overlapping risk tags.
• —VHD11K: deterministic iterative split balancing A/B and real/synthetic origin.
• —Harmful-Contents: preserve the original source split.
• —VHD11K ZIP filenames were repaired and normalized to short Windows-safe names.
• —UGVD orphan item 2001 was excluded because labels cover IDs 1–2000.
• —Harmful-Contents rows with unreliable/unknown sexy or nudity labels were excluded.
• —Preparation checks ensure no ID overlap across splits and ensure every media path exists and is non-empty.
• —A full preflight scan decoded/checked 3,000 videos and 1,225 images without reporting media errors.

The exact preparation code and audit logic are in `scripts/prepare_data.py`.

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="helloworldzzr/Video-Moderation-4225",
    repo_type="dataset",
    local_dir="datasets/prepared",
)

For the complete Windows training workflow:

git clone --recurse-submodules https://github.com/March-77/video-moderation-vlm.git
Set-Location .\video-moderation-vlm
powershell -ExecutionPolicy Bypass -File .\reproduce.ps1

Known limitations

• —UGVD is generated video and differs from real short-video platforms.
• —Image and video slices come from different source datasets, creating domain confounding in modality comparisons.
• —Binary labels collapse many safety categories and policy severities.
• —Some risk tags overlap and are not exhaustive.
• —No audio, subtitle, uploader context, or platform metadata is included.
• —The release should not be treated as a universal or culturally complete definition of harmful content.
• —Human review remains necessary for high-impact moderation decisions.

Redistribution and attribution

See TERMS.md. Cite the original UGVD, VHD11K, and Harmful-Contents sources when using this release, and clearly state any further transformations.