# Create a Markdown file with the enhanced model card content

model_card_content = """\
---
license: apache-2.0
datasets:
- hadyelsahar/ar_res_reviews
language:
- ar
metrics:
- accuracy
- precision
- recall
- f1
base_model:
- aubmindlab/bert-base-arabertv02
pipeline_tag: text-classification
tags:
- arabic
- sentiment-analysis
- transformers
- huggingface
- bert
- restaurants
- fine-tuning
- nlp
---

# **🍽️ Arabic Restaurant Review Sentiment Analysis 🚀**  

## **📌 Overview**  
This **fine-tuned AraBERT model** classifies **Arabic restaurant reviews** as **Positive** or **Negative**.  
It is based on **aubmindlab/bert-base-arabertv2** and fine-tuned using **Hugging Face Transformers**.

### **🔥 Why This Model?**  
✅ **Trained on Real Restaurant Reviews** from the **Hugging Face Dataset**.  
✅ **Fine-tuned with Full Training** (not LoRA or Adapters).  
✅ **Balanced Dataset** (2418 Positive vs. 2418 Negative Reviews).  
✅ **High Accuracy & Performance** for Sentiment Analysis in Arabic.  

---

## **📥 Dataset & Preprocessing**
- **Dataset Source**: [`hadyelsahar/ar_res_reviews`](https://huggingface.co/datasets/hadyelsahar/ar_res_reviews)
- **Text Cleaning**:
  - Removed **non-Arabic text**, special characters, and extra spaces.
  - Normalized Arabic characters (`إ, أ, آ → ا`, `ة → ه`).
  - Balanced **Positive & Negative** sentiment distribution.
- **Tokenization**:
  - Used **AraBERT tokenizer** (`aubmindlab/bert-base-arabertv2`).
- **Train-Test Split**:
  - **80% Training** | **20% Testing**.

---

## **🏋️ Training & Performance**
The model was fine-tuned using **Hugging Face Transformers** with the following hyperparameters:

### **📊 Final Model Results**
| Metric       | Score  |
|-------------|--------|
| **Train Loss** | `0.470` |
| **Eval Loss**  | `0.373` |
| **Accuracy**   | `86.41%` |
| **Precision**  | `87.01%` |
| **Recall**     | `86.49%` |
| **F1-score**   | `86.75%` |

### **⚙️ Training Configuration**
```python
training_args = TrainingArguments(
    output_dir="./results",
    evaluation_strategy="epoch",
    save_strategy="epoch",
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    num_train_epochs=4,
    weight_decay=1,
    learning_rate=1e-5,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,
    fp16=True,
    save_total_limit=2,
    gradient_accumulation_steps=2,
    load_best_model_at_end=True,
    max_grad_norm=1.0,
    metric_for_best_model="eval_loss",
    greater_is_better=False,
)