---
base_model: unsloth/Qwen3.5-0.8B
language:
- ar
library_name: peft
model_name: Katib-Qwen3.5-0.8B-0.1
tags:
- generated_from_trainer
- unsloth
- sft
- trl
- ocr
- arabic
- vision-language
- handwriting-recognition
license: apache-2.0
pipeline_tag: image-text-to-text
metrics:
- wer
- cer
- bleu
datasets:
- oddadmix/qari-0.2.2-news-dataset-large
- oddadmix/qari-0.2.2-diacritics-dataset-large
- johnlockejrr/KHATT_v1.0_dataset
---

# KATIB 0.8B v0.1 — Arabic OCR Model

**KATIB** (كاتب) is a fine-tuned Arabic OCR model built on [Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B), designed to accurately transcribe Arabic text from images — including printed documents and **handwritten content**.

Despite being a 0.8B parameter model, KATIB outperforms larger 2B-class Arabic OCR models on standard benchmarks while running at **2× the speed** with **half the memory footprint**.


---

## ✨ Highlights

- 🏆 **Outperforms Qari-OCR v0.3** (2B) on WER, CER, and BLEU
- 🥈 **Competitive with Qari-OCR v0.2.2.1** (2B) — a stronger model — at half the size
- ✍️ **Enhanced handwriting support** — better generalization to real-world Arabic scripts
- ⚡ **2× faster inference** compared to 2B-parameter alternatives
- 🪶 **Lightweight** — deployable on modest hardware

---

## 📊 Benchmark Results

Evaluated on an Arabic OCR test set. Lower WER/CER is better; higher BLEU is better.

| Model | Size | WER ↓ | CER ↓ | BLEU ↑ |
|---|---|---|---|---|
| **KATIB 0.8B v0.1** *(ours)* | **0.8B** | **0.2386** | **0.0648** | **0.5819** |
| NAMAA-Space/Qari-OCR-v0.3-VL-2B-Instruct | 2B | 0.2643 | 0.0782 | 0.5520 |
| NAMAA-Space/Qari-OCR-0.2.2.1-VL-2B-Instruct | 2B | 0.1993 | 0.0498 | 0.6402 |
| Qwen/Qwen3.5-0.8B *(base, no fine-tune)* | 0.8B | 2.5834 | 1.9487 | 0.0256 |

> **WER** = Word Error Rate | **CER** = Character Error Rate | **BLEU** = Bilingual Evaluation Understudy Score

### Key Takeaways

- KATIB beats **Qari v0.3** across all three metrics — despite being 2.5× smaller.
- KATIB comes close to **Qari v0.2.2.1** on WER and CER, with only a ~6 point BLEU gap — a strong result for a model at this size.
- The base Qwen model without fine-tuning is essentially unusable for Arabic OCR (WER > 2.5), demonstrating the value of domain-specific fine-tuning.

---

## 🚀 Quick Start

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

model_id = "oddadmix/Katib-Qwen3.5-0.8B-0.1"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

image = Image.open("arabic_document.jpg")

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "Free OCR"}
        ]
    }
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=512)

result = processor.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(result)
```

---

## 🧪 Training Details

| Detail | Value |
|---|---|
| Base Model | Qwen/Qwen3.5-0.8B |
| Fine-tuning Method | Supervised Fine-Tuning (SFT) |
| Language | Arabic (Modern Standard + Handwritten) |
| Task | Optical Character Recognition (OCR) |
| Precision | float16 / bfloat16 |

---

## 📋 Intended Use

- ✅ Arabic document digitization
- ✅ Handwritten Arabic text recognition
- ✅ Arabic printed text extraction from images
- ✅ Low-resource / edge deployment scenarios
- ❌ Not intended for non-Arabic languages
- ❌ Not a general-purpose vision-language model

---

## ⚠️ Limitations

- Performance may degrade on very low-quality or heavily degraded scans.
- Dialectal Arabic and mixed-language (Arabic + Latin) text may reduce accuracy.
- Extreme cursive or stylized calligraphy has not been extensively evaluated.

---

## 📄 Citation

If you use KATIB in your research or application, please consider citing this model:

```bibtex
@misc{katib2025,
  title     = {KATIB 0.8B v0.1: A Lightweight Arabic OCR Model},
  author    = {oddadmix},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/oddadmix/Katib-Qwen3.5-0.8B-0.1}
}
```

---