potsawee/video-batch-20260316

Video Batch 20260316

This dataset contains a collection of video files and their associated metadata, organized in WebDataset format (tar archives) to support large file sizes and efficient streaming.

Dataset Structure

The dataset consists of multiple .tar shards in the data/ directory. Each .tar file contains pairs of files for every video:

1. {video_id}.mp4: The raw MP4 video file.
2. {video_id}.json: A sidecar JSON file containing the video's metadata.

Metadata Schema

The {video_id}.json files contain the following fields:

• video_id (str): The unique identifier for the video (usually YouTube ID).
• title (str): The scraped title of the video.
• category (str): The video category.
• source_term (str): The search term loosely associated with this video.
• query (str): The exact query used to find the video.
• url (str): The original URL of the video (e.g., YouTube URL).
• uploader (str): The uploader/channel name.
• channel_id (str): The unique ID of the channel.
• upload_date (str): The date the video was uploaded (YYYYMMDD).
• duration (int): The duration of the video in seconds.
• view_count (int): Number of views at the time of scraping.
• file_size_bytes (int): Actual size of the .mp4 file in bytes.

How to Load the Dataset

Because this dataset is stored in WebDataset format, you can stream it efficiently without having to download hundreds of gigabytes at once.

Using HuggingFace datasets (Streaming)

from datasets import load_dataset

# Load the dataset in streaming mode (recommended for large video datasets)
dataset = load_dataset(
    "potsawee/video-batch-20260316", 
    split="train", 
    streaming=True
)

for sample in dataset:
    # 'sample' is a dictionary containing the contents of the tar file for one video
    video_id = sample['__key__']
    
    # Access the metadata dictionary parsed from the .json file
    metadata = sample['json']
    print(f"Video Title: {metadata['title']}")
    
    # Access the video bytes (or path if caching) implicitly from the .mp4 file
    # Depending on how the decoding is set up, you may need to read the raw bytes:
    video_bytes = sample['mp4']
    
    print(f"Processed video: {video_id} \\n")
    break

Alternatively, you can load it using the webdataset library directly if you are using PyTorch and want an efficient DataLoader.

Using webdataset library

import webdataset as wds

# Read directly from HuggingFace using huggingface-cli or direct URLs
url = "https://huggingface.co/datasets/potsawee/video-batch-20260316/resolve/main/data/shard-{000000..000850}.tar"

dataset = wds.WebDataset(url).decode().to_tuple("mp4", "json")

for video_bytes, metadata in dataset:
    print(metadata['title'])
    break

Downloading Individual Tar Files

Because each shard is a standard, uncompressed .tar archive, you don't have to use a dataset library. You can simply download any shard-XXXXXX.tar file directly from the Hugging Face repository and extract it using standard tools.

Using Command Line:

# Download a specific shard
wget https://huggingface.co/datasets/potsawee/video-batch-20260316/resolve/main/data/shard-000005.tar

# Extract the mp4 and json files
tar -xvf shard-000005.tar

Using Python:

import tarfile
from huggingface_hub import hf_hub_download

# Download a specific shard to local cache
tar_path = hf_hub_download(
    repo_id="potsawee/video-batch-20260316", 
    filename="data/shard-000005.tar",
    repo_type="dataset"
)

# Extract it
with tarfile.open(tar_path, "r") as tar:
    tar.extractall("./extracted_videos")

Creating Shards

These shards were generated by aggregating downloaded mp4 files and jsonl logs, partitioning them into chunks of ~500MB each, and compressing them sequentially into uncompressed .tar archives. This guarantees that PyArrow's 2GB column limit is never encountered, and enables massively parallel streaming.