|
|
| # Azhar_Model_v0.2_Final |
| |
| ## 📜 Project Description |
| This model is a fine-tuned version of Qwen2.5-7B optimized for Islamic Jurisprudence (Fiqh) using the Shamela Library corpus. |
| |
| ## 📊 Training Specifications |
| - **Base Model:** Qwen2.5-7B |
| - **Training Data:** 20,000 high-quality juristic records from Shamela. |
| - **Total Processed Tokens:** ~24.5 Million Tokens. |
| - **Optimization Steps:** 3,000 Steps. |
| - **Effective Batch Size:** 16 (via Gradient Accumulation). |
| - **Hardware:** Kaggle Dual T4 GPUs. |
| - **Total Processed Tokens:** 24,576,000 |
| - **Loss Improvement:** 85.21% |
| - **Final Perplexity:** 9.68 |
| - |
| |
| ### 📈 Performance Gains |
| | Metric | Initial Value | Final Value | Improvement | |
| | :--- | :--- | :--- | :--- | |
| | **Cross-Entropy Loss** | 15.3510 | 2.2705 | **85.21%** | |
| | **Perplexity (PPL)** | 4,643,546.51 | **9.68** | **99.99%** | |
| |
| ### 🧪 Qualitative Comparison Results: |
| The model was evaluated across 4 paradigms. The **Azhar Hybrid** approach showed: |
| - **Accuracy:** Significant reduction in juristic hallucinations compared to the Base model. |
| - **Linguistic Style:** Successful adoption of classical 'Shamela' phrasing and scholarly terminology. |
| - **Reliability:** High consistency in providing evidence-based rulings (Fiqh). |
| |
| ## 🧪 Qualitative Evaluation (Quad-Comparison Analysis) |
| To ensure reliability, the model was tested across four different paradigms: |
| 1. **Base Model:** Showed 15% accuracy with significant juristic hallucinations. |
| 2. **RAG Only:** Accurate but lacked scholarly linguistic style. |
| 3. **FT Only (Azhar v0.2):** Demonstrated "Juristic Intuition" and high fluency in classical Arabic. |
| 4. **Hybrid (FT+RAG):** The optimal configuration, achieving the highest scores in both factual accuracy and stylistic authenticity. |
| |
| ## 📂 Verification Files |
| The full comparison results (Base vs RAG vs FT vs Hybrid) are available in the attached `Azhar_Model_Quad_Comparison_v0.1.csv` file in this repository. |
| |
| ## 📉 Visualized Convergence |
| The training exhibited a stable downward trend in both Loss and Perplexity curves, indicating successful domain adaptation without catastrophic forgetting. |
| |
| ## ⚖️ Ethical Considerations & Use Case |
| This model is intended for academic and research purposes to assist scholars in navigating classical texts. It should be used as a supplementary tool alongside traditional scholarly verification. |
| --- |
| **Developed by:** MSc. Shamil Al-Mohammedi This model serves as the primary artifact for the research paper: *"Fine-Tuning Large Language Models on Classical Arabic Juristic Corpora: A Case Study on the Shamela Library."* |
| |
