wnkh/surya-medieval-latin-lora

Surya OCR - Medieval Latin (LoRA)

LoRA adapters for fine-tuning Surya OCR on medieval Latin manuscripts.

Model Details

• Base Model: vikp/surya_ocr2
• Dataset: wnkh/medieval-latin
• Training Method: LoRA (rank=32)
• Language: Latin (Medieval)

Usage

from surya.foundation import FoundationPredictor
from peft import PeftModel

# Load base model
predictor = FoundationPredictor()

# Load LoRA adapters
model = PeftModel.from_pretrained(predictor.model, "wnkh/surya-medieval-latin-lora")
predictor.model = model

# Run OCR
results = predictor([image], task_name="ocr_with_boxes")
text = results[0]["text"]

Training

Trained using LoRA (Low-Rank Adaptation) on medieval Latin manuscripts. The adapters can be applied to the base Surya OCR model for improved performance on historical Latin text.