ISLAM-PO/MasryGPT_chat_FINALLY
ISLAM-PO/MasryGPT_chat_FINALLY is a 1.5 billion parameter Qwen2.5-Instruct model fine-tuned by ISLAM-PO for Egyptian Arabic (Masry) conversational AI. It specializes in understanding and generating authentic Egyptian dialect, including daily speech, humor, and idioms, addressing the gap left by models optimized for Modern Standard Arabic. With a 32768 token context length, it is designed to be lightweight and deployable on resource-constrained environments for chatbots and customer service in Egypt.
Loading preview...
MasryGPT_chat: Egyptian Arabic Conversational AI
ISLAM-PO/MasryGPT_chat is the first open-source chat model specifically fine-tuned for Egyptian Arabic (Masry), based on the Qwen2.5-1.5B-Instruct architecture. While base Qwen models excel in Modern Standard Arabic (MSA), MasryGPT_chat addresses the critical need for a model that understands and generates authentic Egyptian daily speech, humor, and idioms.
Key Capabilities & Features
- Authentic Egyptian Dialect: Fine-tuned on 80,000 Egyptian terms, achieving 92% natural Masry fluency in manual evaluations, a 48% improvement over the base Qwen model.
- Lightweight & Efficient: A 1.5 billion parameter model, optimized for deployment on devices with limited GPU resources (e.g., T4 GPUs, mobile, edge devices), requiring only 1.2GB VRAM in 4-bit quantization.
- QLoRA & Unsloth Training: Utilizes QLoRA for efficient fine-tuning and Unsloth for 2x faster training, completing 1 epoch in 2 hours 13 minutes with a loss convergence from 0.091 to 0.074.
- High Context Length: Supports a maximum position embedding of 32768 tokens, allowing for longer conversations.
Good For
- Chatbots & Customer Service: Ideal for applications requiring natural, culturally relevant interactions in the Egyptian dialect.
- Social Media & Education: Generating content or providing educational tools tailored to Egyptian users.
- Resource-Constrained Deployments: Its small size and efficiency make it suitable for startups and environments with limited computational power.
Limitations
While strong in general Masry, the model shows weakness in understanding complex idioms (40% correctness) and may produce shorter responses due to training on packed short sentences. There is also a slight degradation in MSA fluency (10%) and no robust safety filter beyond the base Qwen model.