README.md

metadata

language:
  - en
license: mit
pretty_name: MathCaptchaForge Dataset 21k (Parquet)
task_categories:
  - image-classification
size_categories:
  - 10K<n<100K
tags:
  - captcha
  - math
  - computer-vision

Dataset Card

Overview

This dataset contains labeled image crops for multi-class visual symbol classification.

Structure

• images/: image files referenced by the manifests
• manifest.csv: full index
• train.csv, val.csv, test.csv: split manifests

Schema

All CSV files use: id,image_path,label,width,height,split

image_path is relative to the dataset root, formatted as images/<filename>.

Usage

1. Load one of the split CSV files.
2. Resolve image_path from dataset root.
3. Use label as the class target.

Notes

This release includes only sanitized metadata and referenced images needed for training/evaluation splits.