twkeed-sa/twkeed-vision

Twkeed Vision (توكيد للرؤية)

Arabic Vision-Language Model for OCR and Document Understanding, based on Qwen3-VL-4B.

Model Details

• Base Model: Qwen3-VL-4B-Instruct-4bit
• Fine-tuned for: Arabic OCR Text Understanding, Document Understanding
• Framework: MLX (Apple Silicon optimized)
• Type: LoRA Adapters
• Parameters: 4B base + LoRA adapters

Identity

When asked "من أنت؟" (Who are you?), the model responds:

أنا توكيد للرؤية، مساعد ذكي متخصص في قراءة النصوص العربية من الصور والمستندات

Capabilities

• Arabic OCR Understanding: Understand and process Arabic OCR text
• Document Understanding: Extract information from Arabic documents
• Receipt/Invoice Processing: Parse Arabic receipts and invoices
• ID Recognition: Read Saudi IDs and official documents
• Text Recognition: Handle various Arabic fonts and text styles
• 32-Language OCR: Built-in support for 32 languages including Arabic

Usage

import mlx.core as mx
from mlx_vlm import load, generate
from mlx_vlm.trainer import get_peft_model

# Load base model
model, processor = load("mlx-community/Qwen3-VL-4B-Instruct-4bit")

# Apply LoRA structure
target_modules = ["q_proj", "v_proj", "k_proj", "o_proj"]
model = get_peft_model(model, linear_layers=target_modules, rank=16, alpha=2.0, dropout=0.05, freeze=True)

# Load adapters (download from this repo)
adapter_weights = mx.load("path/to/adapters.safetensors")
# Strip language_model prefix
stripped_weights = {k.replace('language_model.', ''): v for k, v in adapter_weights.items()}
model.language_model.load_weights(list(stripped_weights.items()), strict=False)

# Generate with Arabic prompt
prompt = "<|im_start|>user\nمن أنت؟<|im_end|>\n<|im_start|>assistant\n"
result = generate(model, processor, prompt, max_tokens=256)
print(result.text)

Training

Fine-tuned using:

• Hardware: Mac Studio M3 Ultra 96GB
• Framework: mlx-vlm
• Method: LoRA (Low-Rank Adaptation)
• Target Modules: q_proj, k_proj, v_proj, o_proj
• Rank: 16
• Alpha: 32
• Data: Arabic OCR datasets, document understanding examples
• Epochs: 3
• Steps: 2000+
• Final Loss: ~0.09

Files

• adapters.safetensors - LoRA adapter weights (47MB)
• adapter_config.json - LoRA configuration

Qwen3-VL-4B Features

• DeepStack ViT: Enhanced vision encoder
• 32-Language OCR: Built-in multilingual OCR support
• Improved Arabic: Better Arabic text handling than Qwen2.5

License

Apache 2.0

Acknowledgments

• Base model: Qwen Team (Alibaba)
• MLX framework: Apple
• Training framework: mlx-vlm