🦅 Shaheen-Gemma4-Urdu 🦅

"نہیں تیرا نشیمن قصرِ سلطانی کے گنبد پر"
"تو شاہیں ہے، بسیرا کر پہاڑوں کی چٹانوں میں"

---

📖 Overview

Shaheen-Gemma4-Urdu is a high-performance Urdu language model developed by Khurram Pervez (Khurram123). It is fine-tuned on 51,686 high-quality Urdu instruction samples to provide deep linguistic understanding, formal vocabulary, and cultural nuance.

This repository contains both the Full 16-bit Safetensors and the Quantized GGUF versions.

🚀 Key Features

• Dual-Format: Includes Safetensors for transformers and GGUF for llama.cpp.
• Architecture: Based on the state-of-the-art Gemma 4 (2B).
• Urdu Fluency: Specifically tuned to handle complex Urdu grammar and formal literature.
• Speed: Delivers an impressive ~94 tokens per second on an NVIDIA RTX 4060 Ti.
• Academic Precision: Leveraging a background in Mathematics to ensure logical consistency in Urdu reasoning.

🛠️ Technical Specifications

• Dataset: large-traversaal/urdu-instruct
• Training Time: ~2 hours (1 full epoch).
• Final Loss: 1.118, showing strong generalization.
• Format 1: model.safetensors (10.2 GB)
• Format 2: Shaheen-Gemma4-Urdu-Q4_K_M.gguf (3.43 GB)

💻 Usage Instructions

1. Using with llama.cpp (GGUF)

Run the quantized version directly on your GPU:

./llama-cli -m Shaheen-Gemma4-Urdu-Q4_K_M.gguf \
    -p "<start_of_turn>user\nاسلام علیکم! آپ کیسے ہیں؟<end_of_turn>\n<start_of_turn>model\n" \
    -n 128 --n-gpu-layers 33