ZahidYasinMittha/American-Sign-Language-Dataset

American Sign Language (ASL) Dataset

Description:
This dataset contains 108,618 videos representing 2,208 ASL words, with each word having a minimum of 30 videos. The videos were scraped, collected from multiple sources, and preprocessed to ensure consistency, quality, and usability for machine learning and gesture recognition tasks. Each video is ≤10 MB, optimized for storage and model training.
The dataset can be used for ASL gesture recognition, video-based ML tasks, and model training for research and educational purposes.

Folder Structure

E:\Dataset\American-Sign-Language-Dataset ├── part1\ # First subset of videos (≤10,000 files) ├── part2\ # Second subset of videos (≤10,000 files) ├── part3\ # Third subset of videos (≤10,000 files) ├── README.md # This file └── dataset.csv # CSV file mapping words to video paths

• Each subfolder contains ≤10,000 video files to comply with Hugging Face repository limits.
• Videos are organized into multiple parts to prevent exceeding per-folder file limits.
• dataset.csv maps each ASL word to its respective video path.
• README.md provides detailed dataset information and usage instructions.

CSV File (dataset.csv)

The CSV contains the following columns:

Column	Description
word	The ASL word represented in the video
video_path	Relative path to the video file in the repo

Example:

word	video_path
HELLO	part1/12345-HELLO.mp4
THANKS	part2/67890-THANKS.mp4
YES	part3/54321-YES.mp4

• All paths are relative to the root dataset folder.

Dataset Statistics

• Total videos: 108,618
• Total words: 2,208
• Minimum videos per word: 30
• Maximum files per folder: ≤10,000 (split into 3 parts)

Usage Example

import pandas as pd

Load CSV

df = pd.read_csv("dataset.csv")

Example: filter videos for the word "HELLO"

hello_videos = df[df['word'] == 'HELLO']['video_path'].tolist() print(hello_videos)

Key Usage Notes

• Recommended to split dataset into subfolders due to repository file limits.
• Dataset combines multiple sources and has been carefully cleaned to remove duplicates and corrupted videos.
• Ensure that all video paths in dataset.csv match the actual subfolder structure when using programmatically.
• Large uploads were handled using Hugging Face's upload-large-folder utility.

Citation

If using this dataset in research:

American Sign Language Dataset, collected and preprocessed by Zahid Yasin Mittha, 2025.

License

This dataset is released under the MIT License.
You are free to use, modify, and distribute it for research, educational, or commercial purposes, with proper attribution.