PechaBridgeOCR

Tibetan line OCR model (323.24M parameters, VIT encoder + TROCR decoder in a VisionEncoderDecoder architecture) fine-tuned for traditional Tibetan pecha scans with PechaBridge.

Important: This checkpoint expects PechaBridge's gray preprocessing followed by a fixed 256×1024 resize. The gray pipeline uses the minimum RGB channel as grayscale and does not binarize the image. Use the PechaBridge CLI for the supported end-to-end path; raw image input or a generic AutoTokenizer path will not reproduce the training pipeline.

Recommended usage — PechaBridge CLI

git clone https://github.com/CodexAITeam/PechaBridge.git && cd PechaBridge
pip install -r requirements.txt
python cli.py download-models

python cli.py batch-ocr \
    --ocr-model     models/ocr/PechaBridgeOCR \
    --line-model    models/line_segmentation/PechaBridgeLineSegmentation.pt \
    --layout-engine yolo_line \
    --ocr-engine    donut \
    --input-dir     /path/to/pecha/images

Each page image produces a .txt transcript and an *_overlay.jpg preview.

Advanced: standalone Python usage

The tokenizer is backed by BoSentencePiece plus PechaBridge-specific special tokens. Load it with PechaBridge's adapter rather than AutoTokenizer.

from pathlib import Path
import torch
from huggingface_hub import snapshot_download
from PIL import Image
from transformers import AutoImageProcessor, VisionEncoderDecoderModel
from pechabridge.ocr.preprocess_bdrc import BDRCPreprocessConfig, preprocess_image_bdrc
from pechabridge.ocr.sentencepiece_tokenizer_adapter import load_sentencepiece_tokenizer

model_dir = Path(snapshot_download("TibetanCodexAITeam/PechaBridgeOCR"))
model = VisionEncoderDecoderModel.from_pretrained(model_dir).eval()
image_processor = AutoImageProcessor.from_pretrained(model_dir, use_fast=False)
tokenizer = load_sentencepiece_tokenizer(model_dir)
image = Image.open('line_crop.png').convert('RGB')
cfg = BDRCPreprocessConfig.vit_defaults()
cfg = BDRCPreprocessConfig.from_dict({
    **cfg.to_dict(), 'binarize': False, 'gray_mode': 'min_rgb',
})
prepared = preprocess_image_bdrc(image=image, config=cfg).convert('RGB')
pixel_values = image_processor(images=prepared, return_tensors='pt').pixel_values
with torch.inference_mode():
    generated_ids = model.generate(pixel_values)
text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(text)

Model details

• Checkpoint: checkpoint-184000 (step 184000)
• Image preprocessing pipeline: gray
• Input geometry: fixed 256×1024 resize, RGB tensor normalized with mean/std 0.5
• Architecture: VIT encoder + TROCR decoder, 323.24M parameters
• Repro configuration included: yes — preprocessing, generation, normalization, and metric definitions are in repro/
• Training framework: PechaBridge
• Training data: Tibetan pecha line images from OpenPecha and BDRC collections

Evaluation

Checkpoint	Internal validation CER	Valid samples
checkpoint-184000	0.5751%	241

CER is the global character edit distance divided by total normalized reference length. This is a training-time internal validation result, not an independent cross-collection benchmark.

External multi-source evaluation

On a separate 3,306-sample evaluation set, mean CER was 0.80%. 1.63% of samples exceeded 10% CER and 0.57% exceeded 20% CER.

Source dataset	Samples	Mean CER
OCR-Norbuketaka	2,350	0.43%
OCR-Google_Books	812	1.59%
OCR-Lhasakanjur	118	1.92%
OCR-Drutsa	14	3.54%
OCR-Betsug	12	3.86%

This externally supplied result summary was not independently reproduced from artifacts included in this release.

Intended use and limitations

• Intended for OCR of individual Tibetan pecha text-line crops.
• Page-level use requires a separate line-segmentation/layout stage.
• Performance may degrade on other scripts, modern book layouts, handwriting, severe blur, unusual scan colors, or collections not represented during training.
• Outputs should be reviewed before scholarly, archival, or other high-impact use.

License

MIT. See the PechaBridge repository for the project license.