kozhikov/captcha2-resolver-onnx
Image-to-Text • Updated
Datasets:
Request access to this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains real captcha images. Access is granted manually for research and educational purposes (OCR, label-noise and annotation-quality studies). By requesting access you agree to: not use the data or models trained on it to bypass access controls of any live system; not redistribute the images; and cite this dataset in derived work.
Log in or Sign Up to review the conditions and access this dataset content.
captcha2 — text captcha dataset with full label provenance
1999 grayscale-ish JPEG captchas (≈300×50 px) of 6 lowercase alphanumeric symbols from a
~20-symbol alphabet (2345678abcdefgmnpwxy; visually ambiguous symbols such as 0/o, 1/l/i
are not used), collected from a public web page in October 2026. The source is intentionally not
disclosed.
The interesting part is not the images but how the labels were made and fixed — every label carries its history.
| split | images | label |
|---|---|---|
train |
904 | verified |
validation |
100 | verified (frozen; all reported metrics use it) |
unlabeled |
995 | none verified; llm_label (Claude) only, so expect ~1 in 5 to be wrong |
Columns
| column | meaning |
|---|---|
image |
the captcha (original JPEG bytes) |
filename |
YYYYmmdd_HHMMSS.jpg, collection time |
label |
final, verified label (empty for unlabeled) |
rucaptcha_raw |
answer returned by a paid human solver (RuCaptcha ImageToTextTask) |
rucaptcha_label |
normalized: Latin lowercased; empty if the answer was Cyrillic |
llm_label, llm_confidence |
label annotated by Claude, used as a second opinion (high/low); weak: 78.3% exact |
corrected |
true if label differs from rucaptcha_label |
correction_stage |
length_fix / llm_review / alphabet_fix — see provenance/corrections.csv |
How labels were produced
- Paid human labels. Images were sent to RuCaptcha with
minLength=maxLength=6. ~7% of the returned answers were typed in Cyrillic look-alikes and were discarded; the remaining answers are the starting labels. - Second opinion. Every image was also labeled by Claude (an LLM reading the image), with a
self-reported
high/lowconfidence. These labels are not reliable on their own: only 78.3% of them exactly match the verified labels (95.8% per symbol), versus 88.4% for the paid human labels. They were used only to find disagreements worth a second look. - Corrections: 116 of 1004 human labels changed (11.6%), via 118 logged correction rows (before/after + reason):
length_fix— human answer had 5 or 7 symbols;llm_review— human and LLM labels disagreed; resolved by looking at the image;alphabet_fix— symbol outside the alphabet (z→2,v→y, …).
- Sanity checks: exact length 6, no duplicates, image exists, flags for runs of 3+ identical symbols (kept — they are legitimate).
Known residual noise: 3 labels still contain q or s, which are outside the core alphabet and
are probably errors. They were kept to match the trained model exactly.
Intended use
OCR fine-tuning on small data, label-noise / annotation-quality research, teaching material. Not for circumventing captchas on live systems.
Pipeline code (collection logic, labeling, training, ONNX export): see the GitHub repository
captcha2-resolver.
- Downloads last month
- -
Models trained or fine-tuned on kozhikov/captcha2-dataset