|
|
--- |
|
|
title: Arabic RAG Question Answering |
|
|
emoji: 🤖 |
|
|
colorFrom: blue |
|
|
colorTo: green |
|
|
sdk: gradio |
|
|
sdk_version: 4.0.0 |
|
|
app_file: app.py |
|
|
pinned: false |
|
|
license: apache-2.0 |
|
|
--- |
|
|
|
|
|
# 🤖 Arabic RAG Question Answering System |
|
|
|
|
|
An intelligent Arabic question answering system powered by LFM2-1.2B-RAG fine-tuned with **AdaLoRA** - enabling accurate, context-aware responses for general Arabic queries. |
|
|
|
|
|
## 🌟 Why This Model? |
|
|
|
|
|
### ⚡ Fast & Efficient (LiquidAI Architecture) |
|
|
- **Edge-optimized**: Runs efficiently on CPU, GPU, or NPU |
|
|
- **Lightning-fast inference**: 2x faster than comparable models |
|
|
- **Device-agnostic**: Deploy on smartphones, laptops, or servers |
|
|
- **Low memory footprint**: Perfect for resource-constrained environments |
|
|
|
|
|
### 🎯 Advanced Fine-tuning (AdaLoRA) |
|
|
This model uses **AdaLoRA (Adaptive Low-Rank Adaptation)** - an advanced parameter-efficient fine-tuning technique that: |
|
|
- Dynamically allocates model capacity based on importance |
|
|
- **Outperforms standard LoRA** across multiple metrics |
|
|
- Achieves better F1 scores and answer correctness |
|
|
- More efficient parameter usage for superior results |
|
|
|
|
|
### 🌍 Arabic RAG Excellence |
|
|
- General-purpose Arabic question answering |
|
|
- Context-aware responses grounded in provided information |
|
|
- Modern Standard Arabic optimization |
|
|
- Real-world RAG applications ready |
|
|
|
|
|
## 🎯 How to Use |
|
|
|
|
|
1. **Paste Context**: Add any Arabic text containing information |
|
|
2. **Ask Question**: Write your question in Arabic |
|
|
3. **Get Answer**: Receive an accurate, extracted answer instantly |
|
|
|
|
|
Perfect for: document analysis, information extraction, educational tools, customer support, and research applications. |
|
|
|
|
|
## 🔧 Model Details |
|
|
|
|
|
- **Base Model**: [LiquidAI/LFM2-1.2B-RAG](https://huggingface.co/LiquidAI/LFM2-1.2B-RAG) |
|
|
- **Fine-tuning**: AdaLoRA (Adaptive Low-Rank Adaptation) |
|
|
- **Dataset**: [ARCD](https://huggingface.co/datasets/hsseinmz/arcd) — 693 Arabic QA examples |
|
|
- **Language**: Modern Standard Arabic |
|
|
- **Architecture**: Hybrid model with multiplicative gates and convolutions |
|
|
|
|
|
> ⚡ The model can be further enhanced and evaluated on larger or similar Arabic QA datasets to improve generalization and robustness. |
|
|
|
|
|
## ⚡ Features |
|
|
|
|
|
- 🚀 Real-time answer generation |
|
|
- 🎛️ Adjustable generation parameters |
|
|
- 📝 Pre-loaded example questions |
|
|
- 🔄 Full RTL support for Arabic |
|
|
- 📋 Copy-to-clipboard functionality |
|
|
- 💻 Works on any device (CPU/GPU) |
|
|
|
|
|
## 🔗 Resources |
|
|
|
|
|
- **Model Card**: [azeddinShr/LFM2-1.2B-RAG-ARABIC-AdaLoRA](https://huggingface.co/azeddinShr/LFM2-1.2B-RAG-ARABIC-AdaLoRA) |
|
|
- **Training Dataset**: [ARCD](https://huggingface.co/datasets/hsseinmz/arcd) |
|
|
- **Base Model**: [LiquidAI/LFM2-1.2B-RAG](https://huggingface.co/LiquidAI/LFM2-1.2B-RAG) |
|
|
- **Comparison**: Also available - [LoRA variant](https://huggingface.co/azeddinShr/LFM2-1.2B-RAG-ARABIC-LoRA) |
|
|
|
|
|
## 📧 Contact |
|
|
|
|
|
Questions or feedback? Visit the [model repository](https://huggingface.co/azeddinShr/LFM2-1.2B-RAG-ARABIC-AdaLoRA) or email me directly at [azdinsahir11@gmail.com](mailto:azdinsahir11@gmail.com) ! |
|
|
|
|
|
--- |
|
|
|
|
|
**Built with ❤️ using LiquidAI, AdaLoRA, and Gradio** |
