BDRC/tibetan-page-orientation-classifier-dataset

Tibetan Page Orientation Dataset

Covers 7 Tibetan script families: Danyig, Druma, Gyuyig, Multi-Scripts, Pedri, Tsugdri, Uchen.

Dataset composition

Each manuscript page appears twice: once as the original scan (non_flipped) and once rotated 180° (flipped). The model's task is to distinguish these two orientations.

Scripts are balanced — each of the 7 script families contributes the same number of pages (downsampled to the smallest family).

Script (script)	train pages	val pages	test pages	total pages	total images
Danyig	488	61	61	610	1,220
Druma	488	61	61	610	1,220
Gyuyig	488	61	61	610	1,220
Multi-Scripts	488	61	61	610	1,220
Pedri	488	61	61	610	1,220
Tsugdri	488	61	61	610	1,220
Uchen	488	61	61	610	1,220

Splits

Manuscript-stratified 80/10/10 split: each manuscript work appears in exactly one of train / val / test (no data leakage across splits).

Split	Pages	Images	Manuscripts
train	3,416	6,832	1,798
validation	427	854	214
test	427	854	226
Total	4,270	8,540

Page-level split manifest: splits/split.json.

Split policy: work_hash_balanced — manuscript leaks: train↔val 0; train↔test 0; val↔test 0.

Parquet schema

Column	Type	Values
id	string	W123-I456__non_flipped or W123-I456__flipped
image_bytes	binary	JPEG page image
class	string	non_flipped (label 0) or flipped (label 1)
script	string	Parent script family (e.g. Uchen, Druma, Multi-Scripts)

Shards: train/train-*.parquet, val/val-*.parquet, test/test-*.parquet. Train rows are shuffled — non_flipped and flipped images from all scripts are interleaved rather than grouped by page or class.

See split_stats.json and split_stats.md for row-level counts.

Load in Python

from datasets import load_dataset

ds = load_dataset("BDRC/dinov3-binary-page-orientation-dataset")
train = ds["train"]       # 6,832 images (3,416 pages × 2)
val   = ds["validation"]  # 854 images
test  = ds["test"]        # 854 images

from io import BytesIO
from PIL import Image

row = train[0]
img = Image.open(BytesIO(row["image_bytes"])).convert("RGB")
print(row["id"])     # e.g. W123-I456__non_flipped
print(row["class"])  # non_flipped or flipped
print(row["script"]) # e.g. Uchen

Train a model

python experiments/flip_classifier/train.py \
  --hf-dataset BDRC/dinov3-binary-page-orientation-dataset

Citation

@misc{bdrcpageorientation,
  title  = {Tibetan Page Orientation Dataset},
  author = {Buddhist Digital Resource Center and OpenPecha},
  year   = {2026},
  url    = {https://huggingface.co/datasets/BDRC/tibetan-page-orientation-classifier-dataset},
  note   = {Images from BDRC}
}

License

Images taken from the open access collection of the Buddhist Digital Resource Center. Not all images are in the public domain, some are from recent publications possibly under copyright. We provide the images under the Fair Use copyright exception, but any reuse of this dataset will have to be based on a copyright analysis. We provide the classification data under the CC0 1.0 Universal (Public Domain Dedication).

Acknowledgements

All images are provided by the Buddhist Digital Resource Center (BDRC). This dataset was developed by Dharmaduta from specifications provided by BDRC for the project "The BDRC Etext Corpus", with funding from the Khyentse Foundation. Buddhist Digital Resource Center (BDRC). Developed by Dharmaduta / OpenPecha.