BDRC/TDLA-Training-Dataset

TDLA Training Dataset

YOLO-format object-detection dataset for Tibetan Document Layout Analysis (TDLA). The dataset contains bounding-box annotations for four layout classes found in Tibetan document page images. It is split into training, validation, and test sets. The train/val split uses iterative multi-label stratification, while the test set is a hand-picked benchmarking set of the most unique page layouts.

Overview

Property	Value
Total annotations	14705
Number of classes	4
Image format	JPEG (.jpg)
Label format	YOLO (.txt)
Splits	train / val / test
Train/Val stratification	Iterative multi-label stratification (seed 42)

Image Source

All images in this dataset are sourced from the Buddhist Digital Resource Center (BDRC) digital library.

Classes

ID	Name	Annotations	% of total annotations
0	header	4550	30.9%
1	Text area	5844	39.7%
2	footnote	456	3.1%
3	footer	3854	26.2%

Annotation Process

Annotations were created on the Ultralytics HUB platform using the following two-stage workflow:

1. Annotation -- Annotators drew bounding boxes for each of the four layout classes (header, Text area, footnote, footer) on every page image.
2. Quality Control -- A dedicated reviewer inspected each annotated image, verifying label correctness, box tightness, and class assignment before the annotation was accepted into the dataset.

Split Methodology

Train / Val

The training and validation sets were split at an 80/20 ratio using iterative multi-label stratification (seed = 42). This approach treats each image as a multi-label sample (an image may contain several classes simultaneously) and iteratively assigns images to splits so that per-class proportions stay as close to the target ratio as possible. The result is a near-uniform 80/20 distribution for every class, as shown in the tables below.

Test (Benchmarking Set)

The test set was curated independently from the train/val split. Pages exhibiting the most unique and diverse layouts were manually selected from the source collection to maximize layout variety. Each selected page was then manually annotated following the same annotation guidelines used for the rest of the dataset. This hand-picked set serves as the benchmarking dataset — a fixed, high-quality reference for evaluating model performance on challenging and atypical page layouts.

Split Statistics

Split	Images
train	2692
val	103
test	313

Annotation Distribution per Split

Class	train	val	test	Total
header	3424	856	270	4550
Text area	4425	1107	312	5844
footnote	299	75	82	456
footer	2912	728	214	3854

Note: A single image can contain multiple annotations of the same class, so annotation counts may exceed image counts.

Directory Structure

TDLA_Training_dataset/
├── images/
│   ├── train/
│   ├── val/
│   └── test/
├── labels/
│   ├── train/
│   ├── val/
│   └── test/
├── train.txt
├── val.txt
├── test.txt
├── data.yaml
└── README.md

Usage

Point your YOLO training config to data.yaml in this directory:

yolo detect train data=TDLA_Training_dataset/data.yaml

The train.txt, val.txt, and test.txt files list relative image paths for each split.

Label Format

Each .txt label file uses the standard YOLO format — one row per bounding box:

<class_id> <x_center> <y_center> <width> <height>

All coordinates are normalized to [0, 1] relative to image dimensions.

License

This dataset is released under the CC0 1.0 Universal (Public Domain Dedication). You are free to copy, modify, and distribute the data, even for commercial purposes, without asking permission.

Acknowledgements

This dataset was developed by Dharmaduta from specifications provided by the Buddhist Digital Resource Center (BDRC) for the BDRC Etext Corpus, with funding from the Khyentse Foundation.