waqf-ocr-hand-written-v1

waqf-ocr-hand-written-v1

Reads handwritten Arabic text from images and returns it as text. A fine-tuned PaddleOCR-VL model for people digitizing Arabic handwriting.

Made by Waqf-AI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is an optical character recognition model for handwritten Arabic. It was fine-tuned from PaddleOCR-VL-1.6 on a set of augmented Arabic handwriting images. You give it a picture of a line of handwriting and it returns the text.

Who it suits: It suits people who need Arabic handwriting turned into editable text, and researchers who work on Arabic handwriting recognition.

Industry | Who uses it | How they use it | Example
Archives and libraries | Digitization teams | Convert handwritten Arabic pages into searchable text | Scanning a handwritten notebook into text
Education | Teachers and students | Transcribe handwritten Arabic notes or exercises | Turning a handwritten lesson into a text file
Research | Document analysis researchers | Use the model as a baseline for handwriting recognition | Comparing it with another vision model on handwritten lines
Publishing | Editors | Get a first draft of text from handwritten manuscripts | Preparing a manuscript page for proofreading

You can skip it if your text is printed rather than handwritten, or is in a language other than Arabic.

Where to use it:

• Python code that uses the transformers library and PIL for images
• Windows, macOS or Linux, with Python for the download
• Batch jobs over scanned lines of handwriting

Use case | Who | What they get back
Transcribe a handwritten line | Anyone with scans | The Arabic text of the image
Digitize handwritten documents | Archives | Text ready to search or proofread
Benchmark an OCR model | Researchers | Error rates to compare against other models

How to set it up:

• Python on your computer
• The transformers library, PyTorch and the PIL imaging package
• Disk space, about 1.9 GB for the half-precision weights

1. Install the Hugging Face hub package with pip.
2. Download the model files to your computer.
3. Load the processor and model in Python with trust_remote_code turned on.

pip install -U huggingface_hub
hf download Waqf-AI/waqf-ocr-hand-written-v1

How to use it: You open an image of handwriting, load the processor and model through transformers, and ask the model to read the image. The part of the card shown here begins the usage code like this.

from PIL import Image
import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("Waqf-AI/written_ocr_paddle1.6", trust_remote_code=True)

Example: A person has a scanned image of one line of handwritten Arabic and loads it with PIL. They pass it to the model. What comes back is the Arabic text of that line. The card shows sample lines where the output matched the reference exactly.

For newcomers and experts: If you are new to this, begin with one clean image of a single line of handwriting, because the model was tested on text lines. Compare the output with the image yourself.

If you know the field, the card tests on 200 held-out handwritten line images from the same augmented dataset. It reports a word error rate of 0.0077 and a character error rate of 0.0056, against 0.8057 and 0.3690 for the base PaddleOCR-VL-1.6 and higher rates for the Qwen vision models it tried.

Scan or photograph a line of handwriting
Load the image with PIL
Load the processor and model
The model reads the image
Read the Arabic text it returns

Good to know:

Model | Word error rate | Character error rate
PaddleOCR-VL-1.6 base | 0.8057 | 0.3690
Qwen2.5-VL-7B-Instruct | 0.6322 | 0.3292
This fine-tuned model | 0.0077 | 0.0056

• Limits: the scores come from 200 line images of the same kind as the training data, so other handwriting may do worse.
• Trust: the usage code turns on trust_remote_code, which runs code from the model repository.
• Language: the record lists Arabic only.
• License: Apache 2.0.
• Cost to run: none named.
• Data it reaches: it copies the model files to your computer and reads the images you give it.

On Hugging Face, 13 November 2026

• Task: image text to text
• Parameters: 959M
• Weights: about 1.9 GB in F16; about 479 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 266
• Likes: 7
• License: apache-2.0
• Languages: ar
• Tags: safetensors, paddleocr_vl, paddleocr, arabic, handwritten, image-text-to-text, conversational

Links

• Model: https://huggingface.co/Waqf-AI/waqf-ocr-hand-written-v1
• Files: https://huggingface.co/Waqf-AI/waqf-ocr-hand-written-v1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Waqf-AI/waqf-ocr-hand-written-v1

Then follow the usage in its model card. Nothing in this kit runs by itself.