📂 YouTube Embedding Datasets (SeahorseDB)

This repository provides datasets containing CLIP-generated embeddings and metadata for YouTube video scenes. The embeddings are extracted based on timestamps indicating specific scenes within the videos. These datasets are part of the SeahorseDB collection, accessible on Hugging Face.

---

📊 Dataset Overview

Dataset Name	Size	Description
youtube-1M-dataset	1 Million	Contains CLIP embeddings and metadata for 1M YouTube video scenes.
youtube-2M-dataset	2 Million	Contains CLIP embeddings and metadata for 2M YouTube video scenes.
youtube-15M-dataset	15 Million	Contains CLIP embeddings and metadata for 15M YouTube video scenes.

All datasets are stored in Parquet format for efficient storage and processing.

---

🗃️ Dataset Structure

Each dataset includes the following fields:

Field Name	Data Type	Description
feature	sequence	1024-dimensional CLIP-generated embedding vector representing the scene.
title	string	Title of the YouTube video.
channel_name	string	Name of the YouTube channel hosting the video.
video_id	string	Unique identifier for the YouTube video.
timestamp	string	Timestamp indicating the specific scene in the video.
id	int64	Unique identifier for each row.
__index_level_0__	int64	Dummy field

---

📥 How to Access

The datasets can be downloaded directly from Hugging Face:

• SeahorseDB - YouTube Embeddings (1M, 2M, 15M)

You can load the Parquet files using libraries like Pandas or PyArrow.

Example in Python:

import pandas as pd

# Load the 1M dataset
df = pd.read_parquet("path/to/youtube-1M-dataset.parquet")
print(df.head())

# Access embeddings and metadata
embeddings = df['feature'].tolist()  # 1024-dimensional embeddings for scenes
timestamps = df['timestamp'].tolist()
metadata = df['title'].tolist()