remiai3/synthetic-captchas-library

🌍 Synthetic Multilingual CAPTCHA Library

Repository: remiai3/synthetic-captchas-library
A multilingual dataset of synthetic 4-character CAPTCHA images designed for OCR, multilingual vision models, and script recognition research. This dataset spans 44 world writing systems and is especially useful for low-resource script OCR training.

📌 Dataset Summary

Each script includes 100,000 unique CAPTCHA images.
The dataset is provided in two parallel formats:

• CSV version (standard ML workflows)
• Parquet version (faster loading for large-scale pipelines)

⚠️ The images inside both folders are identical — only the label file format differs. So per language:

• 100,000 unique images
• 100,000 duplicate copies (same images, different label format)

🗂 Dataset Structure

tiny-captcha-library/ │ ├── 4char/ │ ├── / │ │ ├── images.zip │ │ └── labels.csv │ └── ... │ ├── 4char_parquet/ │ ├── / │ │ ├── images.zip │ │ └── labels.parquet │ └── ... │ ├── Amharic.png ├── Armenian.png └── ...

Folder Details

Folder	Contains	Purpose
4char/	Images + labels.csv	Standard format
4char_parquet/	Same images + labels.parquet	Efficient ML pipelines
Each language folder includes:

• images.zip → CAPTCHA images
• Label file mapping image filename → correct 4-character text

🌐 Supported Scripts

Amharic, Armenian, Arabic, Bengali, Cherokee, Chinese, Georgian, Greek, Gujarati, Hebrew, Hindi, Hanunoo, Japanese, Kaithi, Kannada, Khmer, Lao, Latin, Lisu, Malayalam, Miao, Modi, Myanmar, Odia, Osage, Russian, Sharada, Siddham, Sinhala, Soyombo, Tai Tham, Tai Viet, Takri, Thaana, Tirhuta, Tamil, Telugu, Thai, Ukrainian, Tifinagh, Korean, Tibetan, Ethiopian, Adlam

🖼 Sample CAPTCHA Styles

Amharic	Arabic	Armenian	Bengali	Cherokee	Chinese
Georgian	Greek	Gujarati	Hanunoo	Hebrew	Hindi
Japanese	Kaithi	Kannada	Khmer	Lao	Latin
Lisu	Malayalam	Miao	Modi	Myanmar	Odia
Osage	Russian	Sharada	Siddham	Sinhala	Soyombo
Tai Tham	Tai Viet	Takri	Tamil	Telugu	Thaana
Thai	Tirhuta	Ukrainian	Tifinagh	Korean	Tibetan
Ethiopian	Adlam

🏷 Label Format

CSV

filename	label
img_00001.png	text
img_00002.png	text

Parquet

Same structure as CSV but stored in columnar format.

📊 Dataset Statistics

Metric	Value
Total Scripts	44
Unique Images per Script	100,000
Total Unique Images	4,400,000
Images Including Duplicates	8,800,000
Characters per CAPTCHA	4

🎯 Intended Use

This dataset is designed for:

• Multilingual OCR training
• Vision-language model pretraining
• Script recognition research
• Robust text recognition under distortion
  🚫 Not intended for bypassing real-world CAPTCHA security systems.

⚙️ Example Loading

import pandas as pd

df = pd.read_csv("labels.csv")
print(df.head())