This repository contains an additional experimental model — a LoRA fine-tuned Qari-OCR — developed during the NakbaNLP 2026 Shared Task (AR-MS) on Arabic Manuscript Understanding (Subtask 2: Systems Track).

Note: This is not our main submission. Our primary model is Ketaba-OCR, 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 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.

python

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:

Split	Samples	Description
Training	15,163	Line images with gold transcriptions (95%)
Validation	799	Line images for evaluation (5%)
Dev Test	2,095	Development test set
Blind Test	2,671	Held-out for official evaluation

Hyperparameters

Parameter	Value	Parameter	Value
Base Model	NAMAA-Space/Qari-OCR-v0.3-VL-2B-Instruct	Architecture	Qwen2-VL-2B
Model Size	~2.21B parameters	Trainable Params	37.6M (1.67%)
LoRA Rank (r)	32	LoRA Alpha (α)	64
Target Modules	q, k, v, o, gate, up, down	LoRA Dropout	0.05
DoRA	True	RSLoRA	True
Learning Rate	2×10⁻⁴	Optimizer	AdamW (fused)
LR Scheduler	Cosine	Warmup Ratio	0.03
Batch Size	2 (per GPU)	Gradient Accumulation	8
Effective Batch	16	Number of Epochs	3
Max Gradient Norm	1.0	Weight Decay	0.01
Max Sequence Length	2048	Precision	bfloat16

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

Dataset	CER	WER	Samples
Development Test	0.5413	0.8873	2,095
Blind Test	0.2635	0.5521	2,671

Comparison with Other Models

Model	Blind CER	Blind WER	Notes
Ketaba-OCR (Our Main Model)	0.0819	0.2588	1st per-line, 3rd corpus-wide
Qari-OCR LoRA (This Model)	0.2635	0.5521	Additional experiment
Qari-OCR v0.3 (Zero-Shot)	0.300	0.485	Base model
Arabic OCR 4-bit v2 (Sherif)	0.3234	0.6203	—
Qwen2.5-VL-7B (Zero-Shot)	0.6808	0.9198	—

⚠️ 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.

📜 Citation

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

bibtex

@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.

HassanB4/Qari-OCR-LoRA

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