---
language:
  - bo
license: cc0-1.0
tags:
  - tesseract
  - ocr
  - tibetan
  - uchen
  - fine-tuned
library_name: tesseract
pipeline_tag: image-to-text
datasets:
  - openpecha/OCR-Tibetan_line_to_text_benchmark
---

# bod_uchen — Fine-Tuned Tesseract Model for Tibetan Uchen Script

A fine-tuned [Tesseract OCR](https://github.com/tesseract-ocr/tesseract) model for recognizing Tibetan Uchen (དབུ་ཅན་) script, trained on 121K+ line-level image–text pairs from historical Tibetan woodblock prints.

## Model Details

| Field | Value |
|-------|-------|
| **Model name** | `bod_uchen` |
| **Base model** | [`bod.traineddata`](https://github.com/tesseract-ocr/tessdata_best/blob/main/bod.traineddata) from `tessdata_best` |
| **Script** | Tibetan Uchen (དབུ་ཅན་) |
| **Language** | Tibetan (bod) |
| **Engine** | Tesseract LSTM |
| **PSM** | 13 (Raw line) |
| **Epochs** | 5 |
| **Learning rate** | 0.001 |
| **Target error rate** | 0.005 |

## Training Data

The model was trained on a curated subset of Tibetan OCR line-level images paired with ground-truth text transcriptions, formatted for Tesseract training. The dataset is derived from [`openpecha/OCR-Tibetan_line_to_text_benchmark`](https://huggingface.co/datasets/openpecha/OCR-Tibetan_line_to_text_benchmark) and prepared by the [`bo_tessaract_data_prep`](https://github.com/OpenPecha/Bo_Tessaract_data_prep) pipeline.

| Split | Directory | Samples |
|-------|-----------|--------:|
| Train | `train-ground-truth/` | 96,978 |
| Eval (Val) | `val-ground-truth/` | 12,123 |
| Test | `test-ground-truth/` | 12,125 |
| **Total** | | **121,226** |

## Training Results

Training converged at iteration 526,010 / 545,515 total:

| Metric | Value |
|--------|-------|
| **Best BCER (selected model)** | 9.719% |
| **Final BCER (train)** | 10.690% |
| **Mean RMS** | 1.077% |
| **Delta** | 2.498% |
| **Skip ratio** | 0.000% |



### Collections Included

Five historical Tibetan text collections are included from the source dataset:

| Collection | Description |
|------------|-------------|
| **Lithang_Kanjur** | Lithang edition of the Tibetan Buddhist canon (Kangyur) |
| **Derge_Tenjur** | Derge edition of the Tengyur commentarial collection |
| **Lhasa_Kanjur** | Lhasa edition of the Kangyur |
| **KhyentseWangpo** | Writings of Jamyang Khyentse Wangpo |
| **Karmapa8** | Texts attributed to the 8th Karmapa, Mikyö Dorje |

## Usage

### Prerequisites

Install Tesseract 5.x:

```bash
# macOS
brew install tesseract

# Ubuntu/Debian
sudo apt install tesseract-ocr
```

### Download & Install the Model

Download `bod_uchen.traineddata` from this repository and place it in your Tesseract `tessdata` directory:

```bash
# Find your tessdata path
tesseract --print-parameters | grep tessdata

# Typical locations:
#   macOS (Homebrew): /opt/homebrew/share/tessdata/
#   Linux:           /usr/share/tesseract-ocr/5/tessdata/
```

### Run OCR

```bash
tesseract input_image.png output -l bod_uchen --psm 13
```

For multi-line page segmentation:

```bash
tesseract input_image.png output -l bod_uchen --psm 6
```

### Python (pytesseract)

```python
import pytesseract
from PIL import Image

image = Image.open("tibetan_line.png")
text = pytesseract.image_to_string(image, lang="bod_uchen", config="--psm 13")
print(text)
```

## Training Command

The model was fine-tuned using [tesstrain](https://github.com/tesseract-ocr/tesstrain):

```bash
make training \
  MODEL_NAME=bod_uchen \
  START_MODEL=bod \
  TESSDATA=data \
  GROUND_TRUTH_DIR=data/bo_tesseract \
  LANG_TYPE=Indic \
  PSM=13 \
  EPOCHS=5 \
  LEARNING_RATE=0.001 \
  TARGET_ERROR_RATE=0.005 \
  WORDLIST_FILE=data/langdata/bod/bod.wordlist \
  NUMBERS_FILE=data/langdata/bod/bod.numbers \
  PUNC_FILE=data/langdata/bod/bod.punc \
  2>&1 | tee -a data/bod_uchen/training.log
```

## Intended Use

- OCR of historical Tibetan woodblock-printed texts in Uchen script
- Digitization of Kangyur, Tengyur, and other classical Tibetan collections
- Building searchable archives of Tibetan Buddhist literature

## Limitations

- Optimized for **Uchen (དབུ་ཅན་)** woodblock print style; may underperform on Umé (དབུ་མེད་) or handwritten Tibetan
- Trained primarily on the five collections listed above; generalization to other print styles or modern typeset Tibetan may vary
- Best results when input images are pre-segmented into individual text lines (PSM 13)

## Citation

If you use this model, please cite:

```bibtex
@misc{bod_uchen_tesseract,
  title   = {bod_uchen: Fine-Tuned Tesseract Model for Tibetan Uchen Script},
  author  = {Buddhist Digital Resource Center (BDRC) and Dharmaduta},
  year    = {2025},
  url     = {https://huggingface.co/bdrc/bod_uchen_tesseract},
  note    = {Fine-tuned from tessdata_best/bod.traineddata, funded by the Khyentse Foundation}
}
```

## Acknowledgements

This model was developed by [Dharmaduta](https://github.com/dharmaduta) from specifications provided by the [Buddhist Digital Resource Center (BDRC)](https://www.bdrc.io/) for the BDRC Etext Corpus, with funding from the [Khyentse Foundation](https://khyentsefoundation.org/).