funa21/tiktok-data

Dataset Card for TikTok Harmful Video Dataset (Vietnamese)

Dataset Details

Dataset Description

This dataset contains TikTok videos collected for research on harmful content detection in Vietnamese. Each sample is stored as a folder with:

• the original video file (video.mp4)
• a metadata file (metadata.json)

The dataset is designed for multimodal learning (video + audio + text from metadata).

• Curated by: Student research project (IE212 – Big Data, UIT, VNU-HCM)
• Language(s): Vietnamese
• License: CC BY-NC 4.0 (non-commercial research/education use)

Dataset Sources

• Source platform: TikTok
• Collection method: Automated crawling (keyword/hashtag-based), then manual filtering

Uses

Direct Use

You can use this dataset for:

• Video classification (e.g., Safe vs Not Safe)
• Multimodal research (video/audio/text)
• Feature extraction pipelines (frames, audio waveform, captions)

Out-of-Scope Use

This dataset is not intended for:

• Commercial use
• Decisions impacting individuals (e.g., banning accounts, legal enforcement)
• Identifying or profiling TikTok users

Dataset Structure

Data is stored using a simple folder-per-video layout:

{video_id}/
├── video.mp4
└── metadata.json

Files

• video.mp4: raw TikTok video
• metadata.json: JSON describing the video (caption/hashtags/stats/etc., depending on what your crawler saved)

Note: This dataset does not necessarily include official train/val/test splits.

How to Use (Python)

Below are simple examples to load and iterate the dataset from a local folder clone.

1. List samples and read metadata

import json
from pathlib import Path

dataset_dir = Path("PATH_TO_DATASET_ROOT")  # e.g. "./tiktok_dataset"

video_folders = [p for p in dataset_dir.iterdir() if p.is_dir()]
print("Total samples:", len(video_folders))

# Read first sample
sample_dir = video_folders[0]
video_path = sample_dir / "video.mp4"
meta_path = sample_dir / "metadata.json"

with meta_path.open("r", encoding="utf-8") as f:
    meta = json.load(f)

print("Sample video_id:", sample_dir.name)
print("Video path:", video_path)
print("Metadata keys:", list(meta.keys()))

2. Build a simple manifest (CSV/JSONL) for training

This is useful if you want a single file listing all samples.

import json
import csv
from pathlib import Path

dataset_dir = Path("PATH_TO_DATASET_ROOT")
out_csv = Path("manifest.csv")

rows = []
for d in dataset_dir.iterdir():
    if not d.is_dir():
        continue
    video_path = d / "video.mp4"
    meta_path = d / "metadata.json"
    if not video_path.exists() or not meta_path.exists():
        continue

    meta = json.loads(meta_path.read_text(encoding="utf-8"))
    caption = meta.get("caption") or meta.get("desc") or ""

    # If you have labels inside metadata.json, try:
    # label = meta.get("label")  # e.g. "safe" / "not_safe"
    # Otherwise set it to empty and label later.
    label = meta.get("label", "")

    rows.append({
        "video_id": d.name,
        "video_path": str(video_path),
        "caption": caption,
        "label": label,
    })

with out_csv.open("w", newline="", encoding="utf-8") as f:
    writer = csv.DictWriter(f, fieldnames=["video_id", "video_path", "caption", "label"])
    writer.writeheader()
    writer.writerows(rows)

print("Wrote:", out_csv, "rows =", len(rows))

3. Extract audio from MP4 (optional)

If you want audio for ASR or audio embeddings, you can extract WAV using ffmpeg.

import subprocess
from pathlib import Path

video_path = Path("PATH_TO_A_VIDEO.mp4")
out_wav = video_path.with_suffix(".wav")

cmd = [
    "ffmpeg", "-y",
    "-i", str(video_path),
    "-ac", "1",          # mono
    "-ar", "16000",      # 16kHz
    str(out_wav)
]
subprocess.run(cmd, check=True)
print("Saved:", out_wav)

4) Read frames with OpenCV (optional)

import cv2

video_path = "PATH_TO_A_VIDEO.mp4"
cap = cv2.VideoCapture(video_path)

frames = []
max_frames = 16

while len(frames) < max_frames:
    ok, frame = cap.read()
    if not ok:
        break
    frames.append(frame)

cap.release()
print("Extracted frames:", len(frames))

Dataset Creation

Curation Rationale

The dataset was created to support Vietnamese-focused research on harmful content detection for short-form videos. It supports multimodal modeling by combining raw video with metadata text.

Data Collection and Processing

• Videos were collected using keyword/hashtag-based crawling.
• Broken/duplicate items were filtered out.
• Each sample is stored as {video_id}/video.mp4 + {video_id}/metadata.json.

Annotations (if applicable)

If your dataset includes labels:

• Classes: Safe / Not Safe
• Method: Manual labeling using project guidelines

(If you do not publish labels, you can remove this part.)

Personal and Sensitive Information

Since the source is social media videos, the dataset may contain personal or sensitive content present in the original videos. No extra personal data was added by the dataset curators. Use the dataset only for non-commercial research and follow ethical data handling practices.

Bias, Risks, and Limitations

• The dataset may reflect TikTok platform bias (recommendation, trends, sampling).
• Harmful content definitions may be subjective and context-dependent.
• The dataset may not represent all demographics or topics equally.

Recommendations

• Use this dataset for research/education only.
• Do not use it as the sole basis for real-world moderation decisions.
• Report limitations and potential bias when publishing results.

Dataset Card Contact

Please open an issue in the dataset repository for questions or problems.