from datasets import load_dataset

ds = load_dataset("expa-ai/BHDD")

# Access a sample
sample = ds["train"][0]
sample["image"]  # PIL Image (28x28 grayscale)
sample["label"]  # int 0–9

With PyTorch:

from datasets import load_dataset
from torchvision import transforms

ds = load_dataset("expa-ai/BHDD")

transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.5,), (0.5,)),
])

def preprocess(example):
    example["image"] = transform(example["image"].convert("L"))
    return example

ds = ds.map(preprocess)
ds.set_format(type="torch", columns=["image", "label"])

Dataset Details

Format: 28×28 grayscale (uint8, 0–255), integer labels 0–9
Script: Myanmar (Burmese), called sar-lone ("round script")
Collection: The Expa.AI Research Team organized a community collection. Contributors wrote digits on A4 paper; the team photographed sheets with phone cameras and extracted digits using an Android app with adaptive thresholding and contour detection
Quality: 20-member annotation team reviewed samples, then two data engineers did a final pass. No duplicates exist within or across splits

Baseline Results

Model	Accuracy	Macro F1	Parameters
MLP (2 hidden layers)	99.40%	0.993	—
CNN (2 conv layers)	99.75%	0.996	421K
Improved CNN (3 conv + BN + augmentation)	99.83%	0.998	431K

Only 47 of 27,561 test samples are misclassified by the best model. Errors cluster around the 0–1 pair (closed vs. open circle).

Citation

@article{aung2025bhdd,
  title   = {{BHDD}: A Burmese Handwritten Digit Dataset},
  author  = {Swan Htet Aung and Hein Htet and Htoo Say Wah Khaing and Thuya Myo Nyunt},
  year    = {2025},
  eprint  = {2603.21966},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url     = {https://arxiv.org/abs/2603.21966}
}