🐪 llama3-arabic-instruct

llama3-arabic-instruct is an instruction-tuned, alignment-optimized Large Language Model tailored for Modern Standard Arabic (MSA) and Egyptian dialect instruction following. Developed by Yousef Khaled Gouda (Senior AI Systems Architect & Principal ML Engineer).

🚀 Key Highlights

• Base Architecture: Meta LLaMA 3 8B / 70B Instruct.
• Fine-Tuning Paradigm: QLoRA SFT on 80,000 curated high-entropy Arabic instruction samples + DPO alignment on 1,000 pairwise human preference rankings.
• Human Preference Score: 82.0% blind win rate over GPT-3.5-turbo in Egyptian and Gulf conversational settings.
• Serving Benchmark: Optimized for vLLM PagedAttention serving with P95 latency under 420ms on GCP A100.

💻 Usage Example

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Yousef1258852/llama3-arabic-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

prompt = "اشرح لي باختصار كيف تعمل أنظمة استرجاع المعلومات RAG في معالجة المستندات القانونية."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📜 Citation & Credits

• Author: Yousef Khaled Gouda
• Organization: github.com/yousefkhaled2005