This model is a fine-tuned version of Meta-Llama-3.1-8B-Instruct, optimized using the Unsloth framework and LoRA adapters. The final weights have been fully merged into a standalone 16-bit format for seamless, direct inference.

The model was trained on a custom dataset containing Egyptian legal documents, contracts, and localized conversational pairs. It is designed to understand complex legal contexts and respond naturally in the Egyptian Arabic dialect.

🚀 Key Features

Dialect Adaptation: Fine-tuned to process and generate responses in pure, natural Egyptian Arabic.
Fully Merged Standalone Model: No need to download separate LoRA adapters or use specialized frameworks. It runs out of the box with standard Hugging Face tools.
MLOps Integration: The training process was rigorously monitored and tracked using Weights & Biases for experiment logging and resource tracking.

🛠️ Usage / How to Run

You can use this model directly with the official Hugging Face transformers library without needing any extra complex frameworks:

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# 1. Define model repository
model_name = "Ahmed792/llama-3.1-8b-egyptian-legal"

# 2. Load the tokenizer and the 16-bit merged model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# 3. Formulate the prompt in Egyptian Arabic
messages = [
    {"role": "user", "content": "لو سمحت، إيه هي الإجراءات القانونية لتأسيس شركة في مصر؟"}
]

# 4. Apply the Llama-3 conversational chat template
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt").to("cuda")

# 5. Generate the response
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.5)

# 6. Decode and print the clean output
print(tokenizer.decode(outputs[0], skip_special_tokens=True))