README.md

---
language:
- bo
license: cc0-1.0
task_categories:
- image-classification
tags:
- tibetan
- uchen
- ume
- script-classification
- paleography
- buddhist-manuscripts
- BDRC
pretty_name: Uchen-Ume Classification Benchmark
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image_bytes
    dtype: image
  - name: script
    dtype:
      class_label:
        names:
          '0': uchen
          '1': ume
  splits:
  - name: train
    num_bytes: 5599672929
    num_examples: 9110
  - name: validation
    num_bytes: 645878888
    num_examples: 1000
  - name: test
    num_bytes: 567916938
    num_examples: 851
  download_size: 6813558825
  dataset_size: 6813468755
---

# Uchen–Ume Classification Benchmark

A binary image classification dataset for distinguishing two fundamental categories of Tibetan script: **Uchen** (དབུ་ཅན།, headed script with a horizontal top stroke) and **Ume** (དབུ་མེད།, headless script without a top stroke). All images are raw, unprocessed manuscript scans from the Buddhist Digital Resource Center (BDRC).

**Model:** [openpecha/uchen-ume-classifier](https://huggingface.co/openpecha/uchen-ume-classifier)

## Dataset summary

| Split | Examples | Uchen | Ume |
|-------|----------|------:|----:|
| Train | 9,110 | ~3,124 | ~5,986 |
| Validation | 1,000 | ~340 | ~660 |
| Test | 851 | ~290 | ~561 |
| **Total** | **10,961** | **~3,754** | **~7,207** |


## Format

Three Parquet files, each with three columns:

| Column | Type | Description |
|--------|------|-------------|
| `id` | string | Original filename (e.g., `W00KG09391-I00KG093950005.jpg`) |
| `image_bytes` | image | Raw manuscript scan, unprocessed (no resizing, cropping, or normalization) |
| `script` | ClassLabel | `uchen` (0) or `ume` (1) |

Images are stored in their original resolution and aspect ratio (typically 5:1 landscape pecha format). Any preprocessing (center cropping, patching, CLAHE normalization) should be applied at training or inference time, not stored in the dataset. This makes the dataset maximally reusable across different experimental setups.

## Loading the dataset

```python
from datasets import load_dataset

repo = "openpecha/uchen-ume-classification-benchmark"

train = load_dataset(repo, split="train")
val   = load_dataset(repo, split="validation")
test  = load_dataset(repo, split="test")

# Access a sample
sample = train[0]
print(sample["id"])           # "W00KG09391-I00KG093950005.jpg"
print(sample["script"])       # 0 (uchen) or 1 (ume)
sample["image_bytes"].show()  # displays the image
```

## Split methodology

Splits are stratified by class and partitioned at the **work level** to prevent data leakage. Each filename follows the pattern `{work_id}-{image_id}.jpg` (e.g., `W00KG09391-I00KG093950005.jpg`, where `W00KG09391` is the work ID). All pages from the same work (manuscript or volume) are assigned to exactly one split — no work appears in more than one of train, validation, or test. This ensures the model cannot exploit visual characteristics shared across pages of the same manuscript (paper tone, ink style, scanning conditions) to inflate evaluation scores.


## Image source

All images are digitised manuscript scans from the **[Buddhist Digital Resource Center](https://www.bdrc.io)** (BDRC), encompassing a wide range of Tibetan Buddhist texts and collections. The scans cover diverse conditions: aged paper, modern reprints, varying ink densities, different scanning equipment, and multiple centuries of manuscript production.

## Annotation process

Images were annotated through a structured process developed by **Dharmaduta**  in collaboration with **BDRC** for the BDRC Etext Corpus project, funded by the **Khyentse Foundation**:

1. **Annotation guidelines** were developed based on a multi-year typology of Tibetan scripts by Pentsok Rtsang, defining the visual criteria for Uchen (horizontal head stroke present) and Ume (head stroke absent) classification.

2. **Label mapping:** Images originally annotated as `uchen_sugthung`, `uchen_sugdring`, or `uchen_sugring` are labeled `uchen`. All other Tibetan script subcategories (`petsuk`, `peri`, `tsegdrig`, `drudring`, `druring`, `druthung`, `drathung`, `khyuyig`, `tsumachug`, `yigchung`, `tsugchung`, `trinyig`, `dhumri`, and others) are labeled `ume`. Non-script categories (`difficult`, `multi_scripts`, `non_tibetan`) are excluded.

3. **Quality control:** Ambiguous images were reviewed by multiple annotators. 

## Additional files

| File | Description |
|------|-------------|
| `splits/train_val_test_splits.json` | Full manifest with page IDs, image URLs, and split assignments |
| `benchmark/benchmark_holdout.json` | 60 holdout pages with image URLs, excluded from all splits |

## License

This dataset is released under [CC0 1.0 Universal](https://creativecommons.org/publicdomain/zero/1.0/) (Public Domain). The manuscript images are provided by BDRC for research and preservation purposes.

## Citation

```bibtex
@misc{karma2026uchenume,
    title        = {Uchen-Ume Classification Benchmark: A Binary Script Classification Dataset for Tibetan Manuscripts},
    author       = {Karma Tashi and Elie Roux},
    year         = {2026},
    publisher    = {HuggingFace},
    url          = {https://huggingface.co/datasets/openpecha/uchen-ume-classification-benchmark},
    note         = {Funded by Khyentse Foundation. Images sourced from the Buddhist Digital Resource Center (BDRC).}
}
```

## Acknowledgements

This dataset was developed by **Dharmaduta** from specifications provided by the **[Buddhist Digital Resource Center](https://www.bdrc.io)** (BDRC) for the BDRC Etext Corpus, with funding from the **Khyentse Foundation**. The annotation guidelines are based on a typology of Tibetan scripts by **Pentsok Rtsang**.