README.md

metadata

license: gpl-3.0
language:
  - en
tags:
  - captcha
  - ocr
  - synthetic
  - adversarial
size_categories:
  - 100K<n<1M

4-captcha dataset

Grayscale synthetic four-digit captcha images for robust digit-string recognition and adversarial fine-tuning experiments.

Splits

Split	Images	Labels
clean/train	100,000	4-digit string
clean/val	5,000	4-digit string
clean/test	5,000	4-digit string
adv/{vit,cnn}/train	20,000 each	same as source clean image
adv/{vit,cnn}/val	1,000 each	same as source clean image
adv/{vit,cnn}/test	1,000 each	same as source clean image

Total stored images: 132,000 for a full release with both ViT and CNN adversarial splits.

Digit combinations for clean train are sampled with replacement from an 8,000-combo pool. Clean val and test use disjoint 1,000-combo pools. Adversarial splits are generated per model from trained clean checkpoints with FGSM.

Image format

• Resolution: $320 \times 80$ pixels
• Channels: 1, grayscale PNG
• Label file: labels.csv with columns filename, label

Rendering

Each image contains four digits $d_1 d_2 d_3 d_4$, $d_i \in {0,\ldots,9}$.

Per-digit variation:

• independent TrueType font from a system font pool
• native glyph height about $0.45$–$0.55$ of image height
• rotation uniform in $[-15°, 15°]$
• small horizontal and vertical shift
• optional slight overlap between neighbours

Canvas background is white or light gray.

Global transforms

Applied to the full image after digit placement, with seed 42:

• elastic deformation
• 2–4 dark Bézier curves, thickness 1–3 px
• Gaussian noise with $\sigma \sim \mathrm{Uniform}(5, 20)$ on the 0–255 scale
• Gaussian blur with kernel $3 \times 3$ or $5 \times 5$
• brightness and contrast jitter
• gamma scaling

Adversarial splits

FGSM with

$$x_{adv}=clip(x+\epsilon \cdot sign({\mathrm{\nabla }}_{x}L),0,1)x\_\{adv\}\; =\; clip(x\; +\; \backslash varepsilon\; \backslash cdot\; sign(\backslash nabla\_x\; L),\; 0,\; 1)$$

where $L$ is the sum of four cross-entropy terms over digit positions.

Training adversarial images use $\varepsilon \in {0.015, 0.03}$ in equal proportion. Validation and test adversarial images use the same $\varepsilon$ set per split. Each model has its own adversarial folders under adv/vit/ and adv/cnn/.

Download

The full dataset is shipped as a single gzip-compressed tar archive data.tar.gz. After extraction the layout is

data/
├── clean/
│   ├── train/
│   ├── val/
│   └── test/
└── adv/
    ├── vit/
    │   ├── train/
    │   ├── val/
    │   └── test/
    └── cnn/
        ├── train/
        ├── val/
        └── test/

from huggingface_hub import hf_hub_download
import tarfile

archive = hf_hub_download(
    repo_id="pymlex/4-captcha",
    repo_type="dataset",
    filename="data.tar.gz",
)
with tarfile.open(archive, "r:gz") as tar:
    tar.extractall(path="data")

huggingface-cli download pymlex/4-captcha data.tar.gz --repo-type dataset
tar -xzf data.tar.gz -C data

Preview

The preview/ folder holds four sample images per split with matching labels.csv files. Use it to inspect rendering and adversarial noise before downloading the archive.

Split	Preview path
clean train	preview/clean/train/
clean val	preview/clean/val/
clean test	preview/clean/test/
adv vit	preview/adv/vit/{train,val,test}/
adv cnn	preview/adv/cnn/{train,val,test}/

Code and training

Pipeline source: github.com/pymlex/4-captcha

Model checkpoints and evaluation artefacts: pymlex/4-captcha-solvers

Citation

@misc{zyukov2026_4captcha,
  author       = {Alex Zyukov},
  title        = {4-captcha: Synthetic Captcha Recognition and Adversarial Fine-tuning},
  year         = {2026},
  howpublished = {\url{https://github.com/pymlex/4-captcha}}
}

The dataset is under GPL-3.0 license.