atalaydenknalbant/MathCaptcha10k

Dataset Details

• Dataset Name: MathCaptcha10k
• Curated by: Atalay Denknalbant
• License: Creative Commons Attribution 4.0 International (CC BY 4.0)
• Repository: https://www.kaggle.com/datasets/atalaydenknalbant/mathcaptcha10k

Dataset Description

A corpus of 10 000 synthetic arithmetic‐captcha images rendered at 200×70 px. Each image contains exactly two base-10 numbers (1–2 digits), a single + or – operator, an = sign and a trailing question mark (e.g. 96-41=?). Every example in the train split includes:

image	ocr_text	result
96-41=?	"96-41=?"	55

…where ocr_text is the exact characters in the image, and result is the integer answer.

The test split consists of 11 766 unlabeled captchas in Unlabeled/ folder.

---

Examples of the Captchas

Easy example

Challenging example

Even state-of-the-art vision-language models often mis‐OCR the more distorted variants (see the “challenging” sample above).

---

Uses

• Direct uses:

  • Train and evaluate OCR/vision-language models on simple arithmetic recognition.
  • Benchmark visual math-solving capabilities.

• Out-of-scope uses:

  • Handwritten digit OCR.
  • Complex mathematical notation beyond two-term arithmetic.

---

Dataset Structure

• Splits

  • train (10 000 labeled examples)
  • test (11 766 .png files in Unlabeled/)

• Features

  • image (PNG file)
  • ocr_text (string, e.g. "75-26=?")
  • result (int, e.g. 49)

---

Dataset Creation

Curation Rationale

Synthetic captchas provide a controlled environment for training and benchmarking. Even top tier vision language methods struggle with some distortions motivating manual QA to ensure label accuracy.

Source Data

Programmatically generated using CaptchaMvc.Mvc5’s standard arithmetic template.

Data Collection & Processing

1. Generate 10 000 PNG captchas via CaptchaMvc.Mvc5.
2. Run a VLM-based OCR pipeline, then manually verify and correct every label in a Streamlit QA app.

Annotator:

• Atalay Denknalbant

---

Personal & Sensitive Information

None. Captchas contain no personal data.

---

Bias, Risks & Limitations

• Purely synthetic; may not generalize to natural or handwritten text.
• Limited to two-term, 1–2 digit arithmetic.

---

Recommendations

Combine with broader OCR datasets for real-world text recognition tasks.

---

Citation

@misc{atalay_denknalbant_2025,
  title        = {MathCaptcha10k},
  author       = {Atalay Denknalbant},
  year         = {2025},
  howpublished = {\url{https://www.kaggle.com/ds/7779792}},
  publisher    = {Kaggle},
  DOI          = {10.34740/KAGGLE/DS/7779792}
}

APA

Denknalbant, A. (2025). MathCaptcha10k [Data set]. Kaggle. https://doi.org/10.34740/KAGGLE/DS/7779792

Dataset Card Authors

• Atalay Denknalbant

Dataset Card Contact

• Atalay Denknalbant (questions & feedback)