MasryGPT Chat Demo

Overview

ISLAM-PO/MasryGPT-Chat-1.5B is an experimental Egyptian Arabic conversational model fine-tuned from Qwen/Qwen2.5-1.5B-Instruct with QLoRA.

• Model size: 1.5B parameters
• Focus: Egyptian Arabic conversational responses
• Base model: Qwen2.5-1.5B-Instruct
• Status: Experimental, internal pilot evaluation only

Standard Inference

Use the canonical Transformers workflow on a GPU machine:

pip install -q torch transformers accelerate safetensors

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

MODEL_ID = "ISLAM-PO/MasryGPT-Chat-1.5B"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

For the LoRA adapter, load the base model first:

from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("aubmindlab/aragpt2-base")
model = PeftModel.from_pretrained(base, "ISLAM-PO/MasryGPT-Flash-Adapter")

Testing

• GPU notebook: MasryGPT_Chat_Test.ipynb in the model repository
• Instructions: TESTING.md
• Evaluation protocol and limits: EVALUATION.md
• Recommended runtime: Colab T4 GPU or local CUDA GPU

Serverless Inference Providers support is not enabled for this custom model. A 404 response means the model is not hosted by a provider. Use Colab or local inference.

Documentation

• Model: https://huggingface.co/ISLAM-PO/MasryGPT-Chat-1.5B
• Adapter: https://huggingface.co/ISLAM-PO/MasryGPT-Flash-Adapter
• Spaces docs: https://huggingface.co/docs/hub/spaces-overview
• Inference Providers docs: https://huggingface.co/docs/inference-providers