The page images in this dataset are scans of Tibetan texts from the BDRC digital library and are provided on a FAIR-USE basis for research. No copyright license is granted. By requesting access you acknowledge that you are solely responsible for performing your own copyright / rights analysis before any use, and that the Buddhist Digital Resource Center (BDRC) accepts no liability for any misuse of this material.

TiBLAD — Tibetan Book Layout Analysis Dataset

YOLO-format object-detection dataset for TiBLA (Tibetan Book Layout Analysis). It contains bounding-box annotations for four layout classes (header, text-area, footer, footnote) on scanned modern Tibetan book pages, split into training, validation, and test sets.

Models, code & paper

Primary model: BDRC/TiBLA-RTDETR (RT-DETR-l, AGPL-3.0)
Permissive alternatives: BDRC/TiBLA-PP-DocLayout-L (Apache-2.0), BDRC/TiBLA-RFDETR (Apache-2.0)
Code: github.com/buda-base/tibla
Paper: buda-base/papers (papers/2026-tibetan-book-layout) — arXiv link forthcoming

Overview

Property	Value
Total annotations	25460
Total images	8325
Number of classes	4
Image format	JPEG (.jpg)
Label format	YOLO (.txt)
Splits	train / val / test
Split unit	series; also volume-clean
Unique volumes	4,533
Unique series	3,807
Archive scans	6,832
BDRC archive, PDF-derived (not augmented)	296
PDF-extracted, then augmented	1,197 (train only)

Image Source

All images are sourced from the Buddhist Digital Resource Center (BDRC) digital library. At a high level there are two sources: archive scans (~85.6%, including a small set of archive files produced from born-digital PDFs) and PDF-extracted, then augmented pages (14.4%, train only). In detail they fall into three groups:

Source	Images	Volumes (`i_id`)	Series (`w_id`)	Share	Splits
Archive scans	6,832	4,053	3,507	82.1%	train 5,317 / val 717 / test 798
BDRC archive, PDF-derived (not augmented)	296	129	105	3.6%	train 229 / val 32 / test 35
PDF-extracted, then augmented	1,197	351	223	14.4%	train only
Total	8,325	4,533	3,807	100%

Volume ids are disjoint across the three groups. Series (w_id) can appear in more than one group, so the series column does not sum to the total.

Archive scans — page photographs from the BDRC image archive (filename prefixes b3__, b4__, v10__). Their volume ids (I…) are not on BDRC’s inventory of volumes whose archive images were produced from born-digital PDFs.
BDRC archive, PDF-derived (not augmented) — the same archive JPEG pipeline, but the volume (i_id) is on that PDF-volume inventory (129 volumes, 105 series). These are the original archive files, not geometrically/photometrically augmented copies.
PDF-extracted, then augmented — pages rendered from born-digital PDFs in a later annotation batch and then augmented so they resemble camera or scanner captures. These are confined to training (filename prefix b11__, __p#### page marker, __aug in the stem).

Volume ids for the first two groups are read from the BDRC filename (batch__W…__I…__page for b3/b4; page file with a 4-digit page suffix for v10) and compared to the i_num column of BDRC’s PDF-volume list. Series ids (w_id) come from the same filenames, or from the volume→work map when the filename only carries an i_id (v10). For the PDF-augmented batch, the etext reproduction IE… is mapped to the corresponding W… series, and the volume is the I… token when present, otherwise the etext volume VE….

Classes

ID	Name	Annotations	% of total
0	header	8155	32.0%
1	text-area	10705	42.0%
2	footnote	367	1.4%
3	footer	6233	24.5%

Headers and footers occur either as a top/bottom running title or page number, or as a side-margin box. Side-margin boxes are uncommon (141 of 14388 header/footer boxes) and were stratified across splits.

Annotation Process

Annotations were created on the Ultralytics HUB platform in a two-stage workflow:

Annotation — annotators drew bounding boxes for each of the four layout classes (header, text-area, footnote, footer) on every page image.
Quality control — a reviewer inspected every image, verifying label correctness, box tightness, and class assignment.
Automated consistency audit — a final geometric/logical audit flagged likely mistakes (near-duplicate or conflicting-class boxes, impossible header/footer/footnote orderings, out-of-bounds boxes).

Split Methodology

The train / val / test split is series-level and leakage-free: pages are grouped at the series (w_id) level and every page of a series is assigned together. Because each volume (i_id) belongs to one series, the split is also volume-clean. After assignment, no page (or an augmented copy of it), no volume, and no series appears in more than one split.

We tried to select a dataset that is as diverse as possible, but it is possible that near-identical layouts appear in different splits.

Series were assigned with a greedy residual balancer aimed at ~81% / ~9% / ~10% of images, matching the global mix of the four classes (with extra weight on the rare footnote class) and of side-margin vs top/bottom headers and footers.

Augmented data — any series that contains a geometrically/photometrically augmented page is assigned to train, so validation and test contain no __aug images. They may still include unaugmented BDRC-archive pages from PDF-derived volumes (32 val, 35 test).
Approximate ratio: ~81% train / ~9% val / ~10% test by image count.

Split Statistics

Split	Images	Volumes	Series
train	6743	3917	3549
val	749	437	92
test	833	179	166

(train includes 1197 augmented images; val and test include 0 and 0. Of the non-augmented images, 229 / 32 / 35 in train / val / test come from BDRC archive volumes that were produced from PDFs.)

Annotation Distribution per Split

Class	train	val	test	Total
header	6431	848	876	8155
text-area	8185	1096	1424	10705
footnote	295	34	38	367
footer	4991	575	667	6233

Header/footer placement (box counts):

Placement	train	val	test	Total
top	6420	848	871	8139
bottom	4887	563	658	6108
side	115	12	14	141

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

Directory Structure

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

Usage

Point your YOLO training config at data.yaml:

yolo detect train data=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 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.

Copyright & Usage Notice

This dataset does not come with an open-content license. The page images are scans of Tibetan texts from the BDRC digital library and are distributed on a fair-use basis for research and non-commercial layout-analysis work.

No copyright license is granted over the underlying page images.
You are solely responsible for performing your own copyright / rights analysis for your jurisdiction and intended use before using this material.
BDRC accepts no liability for any misuse of this material.

By accessing the gated dataset you accept these terms.

Acknowledgements

Developed by the Buddhist Digital Resource Center (BDRC) for the BDRC Etext Corpus project, funded by the Khyentse Foundation. Thanks to the annotators and reviewers from Dharmaduta who produced and consolidated the layout annotations.

Downloads last month: 37

Total file size:

4.98 GB

Models trained or fine-tuned on BDRC/TiBLAD

Collection including BDRC/TiBLAD

TiBLAD (Tibetan Book Layout Detection)

Collection

Detect headers, footers and footnotes in Tibetan books • 4 items • Updated about 12 hours ago