README.md

---
language: ar
license: apache-2.0
library_name: transformers
tags:
- sentiment-analysis
- arabic
- marbert
- twitter
- text-classification
datasets:
- mksaad/arabic-sentiment-twitter-corpus
metrics:
- accuracy
- f1
- precision
- recall
---

# MARBERT Model for Arabic Sentiment Analysis (Positive/Negative)

This is a fine-tuned version of `UBC-NLP/MARBERTv2` for Arabic Sentiment Analysis.
The model is trained to classify Arabic text (specifically tweets) into two categories: **Positive (`LABEL_1`)** or **Negative (`LABEL_0`)**.

## 🚀 Live Demo

You can test the model live on the Hugging Face Space:
**[https://huggingface.co/spaces/iMeshal/arabic-sentiment-app](https://huggingface.co/spaces/iMeshal/arabic-sentiment-app)**

---

## 📊 Model Performance

The model was trained on 80% of the training data and validated on 20%. The final evaluation was performed on a separate, unseen test set.

**Final Test Set Results (Accuracy: 94.40%)**

| Metric | Score |
| :--- | :---: |
| **Accuracy** | **94.40%** |
| F1 (Macro) | 94.40% |
| Precision (Macro) | 94.40% |
| Recall (Macro) | 94.40% |
| Loss | 0.1667 |

The model achieved its best validation accuracy of **93.4%** at Epoch 2, and `load_best_model_at_end` was used.

---

## 💻 Intended Use (How to use)

You can use this model directly with the `transformers` pipeline.

```python
from transformers import pipeline

# Load the pipeline
pipe = pipeline(
    "sentiment-analysis", 
    model="iMeshal/arabic-sentiment-classifier-marbert"
)

# Test with new texts
texts = [
    "هذا المنتج رائع جداً أنصح به",
    "أسوأ خدمة عملاء على الإطلاق",
    "الجو اليوم جميل"
]

results = pipe(texts)
print(results)
# Output:
# [
#   {'label': 'LABEL_1', 'score': 0.99...}, # Positive
#   {'label': 'LABEL_0', 'score': 0.99...}, # Negative
#   {'label': 'LABEL_1', 'score': 0.98...}  # Positive
# ]

```

## 📚 Training Data

The model was trained on the **[Arabic Sentiment Twitter Corpus](https://www.kaggle.com/datasets/mksaad/arabic-sentiment-twitter-corpus)** dataset from Kaggle.

* **Preprocessing:** Long/concatenated tweets (which appeared to be noise) were cleaned.
* **Training Set:** ~24,163 samples.
* **Validation Set:** ~6,041 samples.
* **Test Set:** ~11,508 samples.
* **Balance:** All datasets were perfectly balanced (approx. 50% Positive / 50% Negative).

---

## ⚙️ Training Procedure

The model was trained using the `transformers.Trainer` class with the following key hyperparameters:

* **Framework:** PyTorch
* **Base Model:** `UBC-NLP/MARBERTv2`
* **Epochs:** 3 (with Early Stopping)
* **Early Stopping:** Patience set to 2 (training stopped at Epoch 3, but Epoch 2 was the best).
* **Batch Size:** 16
* **Learning Rate:** 2e-5
* **Tokenizer:** `AutoTokenizer` (with `padding="max_length"`, `truncation=True`, `max_length=512`)

---

### 📞 Contact

* **Name:** Meshal AL-Qushaym
* **Email:** meshalqushim@outlook.com
* **Kaggle:** [kaggle.com/meshalfalah](https://www.kaggle.com/meshalfalah)