aelmah/arabert-sahmnad-sentiment

AraBERT Fine-tuned on SAHMNAD — Arabic News Headlines Sentiment Analysis

Fine-tuned aubmindlab/bert-base-arabertv02 for Arabic sentiment classification (positive/negative) on Moroccan news headlines from the SAHMNAD dataset.

Model Description

This model classifies Arabic news headlines into positive or negative sentiment with 90.47% accuracy. Built using transfer learning on bert-base-arabertv02 with task-specific fine-tuning for binary sentiment analysis on Moroccan press headlines (Hibapress).

Quick Start

from transformers import pipeline

# Load model
classifier = pipeline("text-classification", model="aelmah/arabert-sahmnad-sentiment")

# Predict
result = classifier("فاز المنتخب المغربي بنتيجة كبيرة")
print(result)

# Batch prediction
texts = [
    "فاس.. دورة استثنائية بجماعة زواغة تغيب عنها المعارضة",
    "بعد إطلاق صاروخ.. الطيران الإسرائيلي يقصف مواقع في قطاع غزة"
]
results = classifier(texts)

Performance on SAHMNAD Test Set

Metric	Value
Accuracy	90.47%
F1-score	90.31%

Training Progress

Epoch	Training Loss	Validation Loss	Accuracy	F1
1	—	0.2867	88.53%	87.60%
2	0.3412	0.2704	90.23%	89.65%
3	0.2066	0.2789	90.47%	90.31%

Training Details

Base Model

• Architecture: BERT-base (12 layers, 768 hidden, 12 attention heads)
• Pre-trained: aubmindlab/bert-base-arabertv02
• Parameters: ~110M

Training

• Optimizer: AdamW
• Learning Rate: 2e-5
• Batch Size: 32
• Epochs: 3
• Max Length: 128 tokens
• Hardware: GPU (fp16 mixed precision)
• Framework: Hugging Face Trainer API — Google Colab

Training Data

SAHMNAD (Sentiment-Annotated Hibapress M-NAD) — Arabic news headlines from Hibapress, annotated with binary sentiment labels (0 = negative, 1 = positive).

Split	Examples
Train	13,174
Test	3,295

• Source: Kaggle — SAHMNAD

Intended Use

Direct Use

• Arabic news headlines sentiment monitoring
• Moroccan press opinion mining
• Binary sentiment classification for Arabic short text

Limitations

• Binary classification only (positive/negative) — neutral not supported
• Trained on Moroccan press headlines — may underperform on other domains
• Best with short text under 128 tokens

Links

• Base Model: aubmindlab/bert-base-arabertv02
• Dataset: SAHMNAD on Kaggle
• Code: Google Colab Notebook

License

MIT License