Datasets:
image imagewidth (px) 72 300 | text stringlengths 4 10 | repeat int32 1 13 |
|---|---|---|
111J8 | 1 | |
111XM | 1 | |
113dk | 1 | |
113sB | 1 | |
114ty | 1 | |
115rB | 1 | |
116OC | 1 | |
116gw | 1 | |
11AWu | 1 | |
11AvS | 1 | |
11Cbx | 1 | |
11CgC | 1 | |
11D8l | 1 | |
11EB7 | 1 | |
11FI9 | 1 | |
11GdB | 1 | |
11Iwb | 1 | |
11OTg | 1 | |
11OUZ | 1 | |
11RKd | 1 | |
11U9a | 1 | |
11VFv | 1 | |
11XrW | 1 | |
11Ydw | 1 | |
11YpK | 1 | |
11ZJB | 1 | |
11bav | 1 | |
11ePW | 1 | |
11ePd | 1 | |
11g54 | 1 | |
11gIZ | 1 | |
11kL8 | 1 | |
11nW8 | 1 | |
11pDK | 1 | |
11q3M | 1 | |
11rUA | 1 | |
11rbg | 1 | |
11rqV | 1 | |
11uyC | 1 | |
11xqY | 1 | |
11zhw | 1 | |
11zkc | 1 | |
12Agx | 1 | |
12E8r | 1 | |
12Gyu | 1 | |
12KzM | 1 | |
12Lgp | 1 | |
12LsN | 1 | |
12MEk | 1 | |
12NuE | 1 | |
12OJ9 | 1 | |
12Q9a | 1 | |
12SY7 | 1 | |
12WFp | 1 | |
12cgs | 1 | |
12fk4 | 1 | |
12gWd | 1 | |
12iis | 1 | |
12m3v | 1 | |
12qx3 | 1 | |
12rDE | 1 | |
12tEF | 1 | |
12w4j | 1 | |
12xjG | 1 | |
1319f | 1 | |
133xl | 1 | |
1349p | 1 | |
13EhX | 1 | |
13G2k | 1 | |
13GER | 1 | |
13JOz | 1 | |
13JZC | 1 | |
13Mmv | 1 | |
13N3a | 1 | |
13Nku | 1 | |
13OVS | 1 | |
13PDA | 1 | |
13Tm3 | 1 | |
13UhS | 1 | |
13Vcf | 1 | |
13Vsz | 1 | |
13Wvp | 1 | |
13XHp | 1 | |
13XXx | 1 | |
13ZZt | 1 | |
13aGb | 1 | |
13bd5 | 1 | |
13hR4 | 1 | |
13icl | 1 | |
13nGb | 1 | |
13qTh | 1 | |
13s2B | 1 | |
13yA5 | 1 | |
13yDt | 1 | |
1431K | 1 | |
147sl | 1 | |
1483t | 1 | |
14ADy | 1 | |
14CMy | 1 | |
14FMT | 1 |
End of preview. Expand in Data Studio
KaptchaOCR recognition dataset
CAPTCHA images with text labels, used to train the ft-tiny / ft-small /
ft-medium recognition models.
The images are CAPTCHA samples collected from public sources.
Splits
| Split | Images |
|---|---|
train |
169,311 |
val |
5,000 |
test_atlas |
1,000 |
test_birch |
1,019 |
test_cobalt |
107 |
test_dune |
674 |
test_ember |
1,000 |
test_fjord |
599 |
The training label list has 194,355 rows because some images are oversampled.
Columns
image— CAPTCHA image.text— case-sensitive label from a 73-character set.repeat— number of times the image appears in the oversampled training label list (1for validation and benchmark splits).
Usage
from datasets import load_dataset
ds = load_dataset("zaltahir/kaptchaocr-rec", split="train")
print(ds[0]["text"])
Curation
- exact MD5 deduplication across all sources
- pHash near-duplicate removal (Hamming ≤ 8) inside train/val and between train and test, to prevent leakage
- test sets stratified by label length and held out before pool building
Links
- Downloads last month
- 110