Research Preview: This is the second iteration (v2) of the Egyptian Legal fine-tuned model. It is currently in a Beta stage for evaluation and community feedback.

Project Vision

Building on the foundations of v1, this version (v2) aims to bridge the gap between complex Egyptian legislative language and AI understanding. By leveraging the advanced reasoning of Qwen2.5-3B, this model is tailored to handle the specific linguistic complexities and formal structures found in Egyptian Civil, Criminal, and Administrative codes.

Key Objective: To assist legal researchers and practitioners in navigating the Egyptian legal landscape through a highly specialized, instruction-following AI.

Technical Configuration (v2)

This version introduces a more refined training pipeline compared to its predecessor:

Foundation: Qwen/Qwen2.5-3B-Instruct
Technique: LoRA (Low-Rank Adaptation) for efficient, high-fidelity fine-tuning.
Training Loop: 4 Epochs across a diverse set of Egyptian legal corpora.
Environment: Powered by Dual NVIDIA T4 GPUs via Kaggle.
Orchestration: LLaMA-Factory.

Capabilities & Benchmarks

Compared to general-purpose models, v2 excels in:

Contextual Jurisprudence: Better understanding of how Egyptian laws interconnect.
Structural Integrity: Generating legal documents (contracts, memos) that follow the Egyptian "Legalese" template.
Semantic Search Ready: Optimized for RAG (Retrieval-Augmented Generation) applications involving Egyptian statutes.

Experimental Disclaimer

This model is a work-in-progress. * Accuracy: It may generate hallucinations or cite outdated articles.

Non-Legal Advice: Outputs must be verified by a qualified Egyptian attorney.
Arabic Support: While the model is bilingual, its primary strength lies in the formal Arabic used in the Egyptian Official Gazette.

Quick Start

python

from transformers import pipeline

pipe = pipeline("question-answering", model="tarekys5/qwen2.5-3B-egyptian-legalv2-wip")