Skip to content
ModelListedExternal
Download

Trocr Traditional Chinese Historical Finetune

This is a TrOCR model finetuned on Lo Chia-Luen historical manuscripts for Traditional Chinese historical document recognition.

Sourcehuggingface.co/ZihCiLin/trocr-traditional-chinese-historical-finetune/tree/d7bf47416d26f02113e8517d8c44c222a59c9d11
KindOtherBase modelZihCiLin/trocr-traditional-chinese-baselineVersionRevision d7bf474

Model source

Kind
Other
File size
1.15 GB
Base model
ZihCiLin/trocr-traditional-chinese-baseline
Version
vRevision d7bf474
Updated
Sep 17, 2026
Downloads
70
Tasks
Captioning
Source
Hugging Face

Source description

TrOCR Traditional Chinese Historical Finetune

Model Description

This is a TrOCR model finetuned on Lo Chia-Luen historical manuscripts for Traditional Chinese historical document recognition.

Intended Use

This finetuned model is optimized for:

  • Historical manuscript recognition
  • Semi-classical Chinese text
  • Degraded document images
  • Archaic character variants

Training Details

Finetuning Data

Finetuning Hyperparameters

  • Batch size: 8 (per device)
  • Learning rate: 2e-5
  • Epochs: 30
  • Data augmentation: First 20 epochs (ColorJitter, RandomAffine, GaussianBlur)
  • Clean training: Last 10 epochs

Usage

from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image

# Load model and processor
processor = TrOCRProcessor.from_pretrained("ZihCiLin/trocr-traditional-chinese-historical-finetune")
model = VisionEncoderDecoderModel.from_pretrained("ZihCiLin/trocr-traditional-chinese-historical-finetune")

# Load historical document image
image = Image.open("historical_manuscript.jpg").convert("RGB")

# Generate text
pixel_values = processor(image, return_tensors="pt").pixel_values
generated_ids = model.generate(pixel_values)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

print(generated_text)

With Decoding-Time Fusion (Recommended)

For best results, use with LLM-guided decoding:

# See: https://github.com/Jason9339/traditional-chinese-historical-document-ocr-llm-fusion
# Fusion with Breeze-7B improves CER by 10% on historical documents

Limitations

  • Optimized for Lo Chia-Luen collection style (20th century Chinese manuscripts)
  • Performance may vary on other historical periods or writing styles
  • Requires reasonable image quality (severe degradation may still cause errors)

Citation

If you use this model, please cite:

@inproceedings{lin2026decoding,
  title={Decoding-Time Fusion of OCR and Large Language Models for Traditional Chinese Historical Document Recognition},
  author={Lin, Zih-Ci and Liao, Wen-Hung},
  booktitle={International Conference on Pattern Recognition (ICPR)},
  year={2026}
}

Acknowledgments

This work is based on the Lo Chia-Luen Manuscripts from the Special Collection Center of NCCU Libraries.

Related Resources

Sources

1 source
Hugging FacePrimaryVersion d7bf47416d26f02113e8517d8c44c222a59c9d11
huggingface.co/ZihCiLin/trocr-traditional-chinese-historical-finetune/tree/d7bf47416d26f02113e8517d8c44c222a59c9d11

Verified Sep 17

Model artifacts

1 artifact

model.safetensors

safetensors · 1.15 GB · SHA-256 7432edf5d1e9…cb4d · Hugging Face

Download

Source excerpts

2 excerpts
ReadmeSep 17huggingface.co/ZihCiLin/trocr-traditional-chinese-historical-finetune

--- language: - zh tags: - trocr - ocr - traditional-chinese - historical-documents - vision - image-to-text datasets: - ZihCiLin/traditional-chinese-historical-ocr-lo-chia-luen base_model: - ZihCiLin/trocr-traditional-chinese-baseline metrics: - cer library_name: transformers --- # TrOCR Traditional Chinese Historical Finetune ## Model Description This is a TrOCR model finetuned on Lo Chia-Luen historical manuscripts for Traditional Chinese historical document recognition. - **Base Model**: [ZihCiLin/trocr-traditional-chinese-baseline](https://huggingface.co/ZihCiLin/trocr-traditional-chinese-baseline) - **Finetuning Data**: 690 annotated historical manuscript samples - **Task**: Historical OCR for Traditional Chinese documents ## Intended Use This finetuned model is optimized for: - Historical manuscript recognition - Semi-classical Chinese text - Degraded document images - Archaic character variants ## Training Details ### Finetuning Data - **Dataset**: [ZihCiLin/traditional-chinese-historical-ocr-lo-chia-luen](https://huggingface.co/datasets/ZihCiLin/traditional-chinese-historical-ocr-lo-chia-luen) - **Train**: 690 samples - **Validation**: 46 samples - **Source**: Lo Chia-Luen Manuscripts (20th century Chinese historical documents) ### Finetuning Hyperparameters - Batch size: 8 (per device) - Learning rate: 2e-5 - Epochs: 30 - Data augmentation: First 20 epochs (ColorJitter, RandomAffine, GaussianBlur) - Clean training: Last 10 epochs ## Usage ```python from transformers import TrOCRProcessor, VisionEncoderDecoderModel from PIL import Image # Load model and processor processor = TrOCRProcessor.from_pretrained("ZihCiLin/trocr-traditional-chinese-historical-finetune") model = VisionEncoderDecoderModel.from_pretrained("ZihCiLin/trocr-traditional-chinese-historical-finetune") # Load historical document image image = Image.open("historical_manuscript.jpg").convert("RGB") # Generate text pixel_values = processor(image, return_tensors="pt").pixel_values generated_ids = model.generate(pixel_values) generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0] print(generated_text) ``` ### With Decoding-Time Fusion (Recommended) For best results, use with LLM-guided decoding: ```python # See: https://github.com/Jason9339/traditional-chinese-historical-document-ocr-llm-fusion # Fusion with Breeze-7B improves CER by 10%...

NoteSep 17huggingface.co/ZihCiLin/trocr-traditional-chinese-historical-finetune

Source context: 70 downloads · 0 likes · Pipeline image-to-text · Library transformers · Repo ZihCiLin/trocr-traditional-chinese-historical-finetune