ISLAM-PO/MasryGPT-Chat-1.5B

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ISLAM-PO/MasryGPT-Chat-1.5B is a 1.56 billion parameter Qwen2.5-1.5B-Instruct based causal language model fine-tuned by ISLAM-PO for Egyptian Arabic (Masry) dialect. Optimized for conversational fluency in authentic Egyptian daily speech, humor, and idioms, it achieves a 4.6/5 human fluency rating on 100 prompts. This lightweight model is designed for efficient deployment on resource-constrained devices like mobile or T4 GPUs, making it suitable for chatbots and customer service in Egypt.

Loading preview...

MasryGPT Chat: Egyptian Dialect Conversational AI

MasryGPT Chat is an experimental open-source 1.56 billion parameter model, built on Qwen2.5-1.5B-Instruct and fine-tuned by ISLAM-PO using QLoRA and Unsloth. Its primary goal is to bridge the gap between Modern Standard Arabic (MSA) and authentic Egyptian Arabic (Masry), focusing on daily speech, humor, and idioms.

Key Capabilities & Features

  • Egyptian Dialect Fluency: Achieves a 4.6/5 human fluency rating for Egyptian dialect, significantly outperforming the base Qwen model (3.1/5).
  • Lightweight & Efficient: At 1.5B parameters, it's designed for deployment on devices with limited VRAM (e.g., T4 GPUs, mobile), requiring only 2-4GB VRAM in 4-bit mode.
  • Optimized Training: Fine-tuned on 80,000 Egyptian terms over 2,500 steps, resulting in a smooth 18.7% loss reduction (0.091 to 0.074).
  • Unsloth Integration: Leverages Unsloth for 2x faster fine-tuning, making it efficient even on free-tier GPUs.

Use Cases & Considerations

  • Chatbots & Customer Service: Ideal for applications requiring natural, colloquial Egyptian Arabic interactions.
  • Social Media & Education: Can be used for generating content or educational tools tailored to the Egyptian context.
  • Resource-Constrained Environments: Its small size and efficient inference make it suitable for startups or edge deployments.

Limitations

  • Idiom Weakness: While improved, understanding of complex Egyptian idioms remains a challenge.
  • MSA Degradation: Exhibits a slight decrease in Modern Standard Arabic fluency compared to the base model.
  • Short Responses: Tends to generate shorter responses, requiring adjustment of max_new_tokens for longer outputs.
  • Bias: Data primarily reflects Cairo dialect, potentially introducing regional bias.

Finetunes
1 models