hastyle/olmOCR-arabic-lora-v2

olmOCR Arabic LoRA v2

LoRA fine-tuned adapter for Arabic manuscript OCR based on allenai/olmOCR-2-7B-1025.

Training Details

• Base Model: allenai/olmOCR-2-7B-1025
• LoRA Rank: 64
• LoRA Alpha: 128
• Training Data: 1,222 full-page Arabic manuscript images from hastyle/arabic-manuscript-ocr
• Epochs: 10
• Final Loss: ~7.2

Usage

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
from peft import PeftModel

# Load base model
base_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "allenai/olmOCR-2-7B-1025",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "hastyle/olmOCR-arabic-lora-v2")
processor = AutoProcessor.from_pretrained("allenai/olmOCR-2-7B-1025")

Improvements over v1

• Trained on full-page manuscripts instead of text-line images
• Higher LoRA capacity (rank 64 vs 16)
• Better word boundary preservation