HassanB4/Qari-OCR-LoRA

Qari-OCR-LoRA: Additional Model for NakbaNLP 2026 Shared Task

<p align="center"> <img src="https://placehold.co/800x200/e0f2fe/0369a1?text=Qari-Arabic-OCR" alt="Qari Arabic OCR"> </p>

This repository contains an additional experimental model — a LoRA fine-tuned Qari-OCR — developed during the [NakbaNLP 2026 Shared Task](https://acrps.ai/nakba-nlp-manu-understanding-2026) (AR-MS) on Arabic Manuscript Understanding (Subtask 2: Systems Track).

Note: This is not our main submission. Our primary model is [Ketaba-OCR](https://huggingface.co/HassanB4/Ketaba-OCR-LoRA), which ranks 1st on per-line evaluation (CER 0.0819, WER 0.2588) and 3rd on the official (corpus-wide) leaderboard (CER 0.0938, WER 0.2996).

By: Hassan Barmandah, Fatimah Emad Eldin, Khloud Al Jallad, Omer Nacar — NAMAA Community (with Umm Al-Qura University, Trouve Labs, Syrian Society for Startups and Research, Tuwaiq Academy)

  

Model Description

This is an additional experimental model that fine-tunes Qari-OCR using Low-Rank Adaptation (LoRA) with DoRA and RSLoRA for Arabic handwritten text recognition. The base model is NAMAA-Space's Qari-OCR v0.3, built on Qwen2-VL-2B architecture.

While this model achieves reasonable results (CER 0.2635 on blind test), our main submission [Ketaba-OCR](https://huggingface.co/HassanB4/Ketaba-OCR-LoRA) significantly outperforms it (CER 0.0819 per-line; 1st on per-line, 3rd on corpus-wide).

The model transcribes cropped line images from Arabic manuscripts into machine-readable text, optimized for the Omar Al-Saleh Memoir Collection (1951-1965) written in Ruq'ah and Naskh script variants.

Key Features

• —Parameter Efficiency: LoRA fine-tuning with only ~37.6M trainable parameters (1.67% of total)
• —DoRA + RSLoRA: Weight-Decomposed Low-Rank Adaptation with rank stabilization for improved training
• —Lightweight Base: 2.2B parameter model (Qwen2-VL-2B) for faster inference
• —Experimental: Additional model for comparison with our main HRT-based approach

🚀 How to Use

You can use the fine-tuned model directly with the transformers and peft libraries.

import torch
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from peft import PeftModel
from PIL import Image

# Load base model
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "NAMAA-Space/Qari-OCR-v0.3-VL-2B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
 
# Load LoRA adapter
model = PeftModel.from_pretrained(model, "HassanB4/Qari-OCR-LoRA")
model.eval()

# Load processor
processor = AutoProcessor.from_pretrained(
    "NAMAA-Space/Qari-OCR-v0.3-VL-2B-Instruct",
    trust_remote_code=True
)

# Example inference
image = Image.open("manuscript_line.png").convert("RGB")

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": "Below is the image of one page of a document. Just return the plain text representation of this document as if you were reading it naturally. Do not hallucinate."}
    ]
}]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt")
inputs = {k: v.to(model.device) for k, v in inputs.items()}

with torch.no_grad():
    output_ids = model.generate(**inputs, max_new_tokens=512, do_sample=False)

transcription = processor.decode(output_ids[0][len(inputs['input_ids'][0]):], skip_special_tokens=True)
print(transcription)

⚙️ Training Procedure

The system employs LoRA/DoRA fine-tuning of the Qari-OCR model.

Training Data

The model was fine-tuned on the official NakbaNLP 2026 dataset from the Omar Al-Saleh Memoir Collection:

Hyperparameters

Frameworks

• —PyTorch 2.5+
• —Hugging Face Transformers ≥4.45.0
• —PEFT ≥0.14.0
• —bitsandbytes ≥0.43.0

📊 Evaluation Results

The model was evaluated on both development and blind test sets provided by the NakbaNLP 2026 organizers.

Test Set Scores

Comparison with Other Models

⚠️ Limitations

• —Domain Specificity: Optimized for 1950s Ruq'ah/Naskh manuscripts; requires adaptation for other periods/styles
• —Higher Error Rate: CER of 0.26 is higher than the HRT-based Ketaba-OCR (0.08–0.09), suggesting the smaller model capacity limits performance
• —Degraded Images: Performance degrades on severely faded or damaged manuscript regions
• —No Ensemble: Results are from a single model without ensemble techniques

🙏 Acknowledgements

We thank the NakbaNLP 2026 organizers for access to the Omar Al-Saleh Memoir Collection. We acknowledge NAMAA-Space for the Qari-OCR pretrained model, and the Hugging Face community for PEFT libraries.

Related Links

• —NakbaNLP 2026 Workshop
• —AR-MS Shared Task Website
• —Base Model (Qari-OCR v0.3)

📜 Citation

If you use this work, please cite our main paper:

@inproceedings{barmandah2026ketaba,
    title={{Ketaba-OCR at AR-MS NakbaNLP 2026: Efficient Adaptation of Vision-Language Models for Hand Written Recognition}},
    author={Barmandah, Hassan and Eldin, Fatimah Emad and Al Jallad, Khloud and Nacar, Omer},
    year={2026},
    booktitle={Proceedings of LREC 2026},
    note={NakbaNLP 2026 Shared Task}
}

📄 License

This project is licensed under the Apache 2.0 License.