aipanjab/handwritten-digits

Handwritten Panjabi Digits

4,417 handwritten Gurmukhi digits (੦ ੧ ੨ ੩ ੪ ੫ ੬ ੭ ੮ ੯), drawn by people on AI Panjab. It is the dataset behind AI Panjab's digit recognition model, aipanjab/digit-recognition, which you can try in its Space.

Data

column	type	example
image	28 x 28 RGB image	a drawn digit, dark on a light background
label	one of 10 classes	2 = ੨

The label is a class number; features["label"].int2str turns it into the digit (0 = ੦ … 9 = ੯).

split	images
train	3,533
test	884

Images per digit:

digit	੦	੧	੨	੩	੪	੫	੬	੭	੮	੯
train	435	350	410	277	343	264	301	440	401	312
test	112	101	111	91	81	79	70	87	79	73

Use it

from datasets import load_dataset

ds = load_dataset("aipanjab/handwritten-digits")
row = ds["train"][0]
row["image"]                                              # a PIL image, 28 x 28
ds["train"].features["label"].int2str(row["label"])       # "੦" ... "੯"

raw-data.zip has the same images as PNG files, in folders by split and digit (raw-data/<split>/<digit>/<n>.png).

How it was made

People drew one digit at a time with a black pen on AI Panjab's drawing pad (about 350 x 350 pixels); each drawing was scaled down to 28 x 28.

Limits

• Dark digits on a light background only; one centred digit per image, filling most of it.
• The classes are not perfectly balanced (੫ has the fewest training images, ੭ the most).
• Drawn with a mouse, finger or stylus on a screen, not written on paper: scanned handwriting looks different.

Licence

CC BY 4.0: you may use it, including commercially, if you credit AI Panjab.

Citation

@misc{aipanjab_handwritten_digits,
  title  = {Handwritten Panjabi Digits},
  author = {{AI Panjab}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/aipanjab/handwritten-digits}
}