---
license: apache-2.0
task_categories:
- image-to-text
tags:
- captcha
- ocr
- synthetic
- vision-language
size_categories:
- 10K<n<100K
---

# Synthetic CAPTCHA Dataset

Synthetic distorted-text CAPTCHA images with exact character-string labels, used
to fine-tune [`Hansda/minicpm-v4_6-captcha-lora`](https://huggingface.co/Hansda/minicpm-v4_6-captcha-lora).

## Contents

- **Format:** HuggingFace `imagefolder` (`image`, `text` columns), with
  `metadata.jsonl` per split.
- **Splits:** `train` (10,000) and `test` (2,000) — 12,000 images total.
- **Labels:** 4–7 characters from `ABCDEFGHJKLMNPQRSTUVWXYZabcdefghjkmnpqrstuvwxyz23456789`
  (visually ambiguous `I O 0 1 l` are excluded).

## Generation

Images are procedurally generated (Pillow + PyTorch) with per-character random
font / size / rotation / shear / stretch, wave & bezier warping, background
gradients and noise, random lines, curves and arcs, and varied colours — to
mimic classic distorted-text CAPTCHAs.

## Load

```python
from datasets import load_dataset
ds = load_dataset("Hansda/captcha-dataset")
print(ds["train"][0]["text"], ds["train"][0]["image"].size)
```

## Note

The per-character size is randomized independently, so upper/lower case is not
always visually separable for ambiguous glyphs (`c/C`, `s/S`, `v/V`, …). This
makes strict case-sensitive scoring inherently hard; case-insensitive reading is
the more reliable target.