library_name: transformers
base_model: aubmindlab/bert-base-arabertv02
tags:
- generated_from_trainer
metrics:
- accuracy
- f1
model-index:
- name: arabic-sentiment-model
results: []
language:
- ar
pipeline_tag: text-classification
datasets:
- ramybaly/arsentd_lev
arabic-sentiment-model
This model is a fine-tuned version of aubmindlab/bert-base-arabertv02 on an ramybaly/arsentd_lev dataset. It achieves the following results on the evaluation set:
- Loss: 0.1512
- Accuracy: 0.9454
- F1: 0.9454
Model description
This model is a fine-tuned version of aubmindlab/bert-base-arabertv02 , adapted for Arabic Sentiment Analysis.
The model is trained to classify Arabic text into binary sentiment classes (Positive / Negative). It is suitable for analyzing opinions expressed in Modern Standard Arabic (MSA) as well as dialectal Arabic, commonly found in social media posts, product reviews, and user feedback.
The model benefits from AraBERT’s strong contextual understanding of Arabic morphology and syntax, resulting in high classification accuracy.
Intended uses & limitations
This model can be used for:
Arabic sentiment analysis
Social media opinion mining
Customer feedback analysis
Academic research and NLP experiments
Graduation and portfolio projects
It is designed for inference on short to medium-length Arabic texts.
Limitations
The model performs binary sentiment classification only (no neutral class).
Performance may degrade on very long documents.
Training and evaluation data
Training and Evaluation Data
The model was trained and evaluated using the ramybaly/arsentd_lev dataset dataset, which consists of Arabic text labeled for sentiment polarity.
Dataset Characteristics
Language: Arabic
Labels: Positive, Negative
Text Type: Short Arabic opinions and statements
Domains: General opinionated text
The dataset was split into training, evaluation, and test sets following standard supervised learning practices.
Training procedure
Preprocessing
Arabic text normalization handled by AraBERT tokenizer
Tokenization using the AraBERT v02 tokenizer
Padding and truncation applied to ensure fixed input length
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 64
- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 50
- num_epochs: 3
Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy | F1 |
|---|---|---|---|---|---|
| 0.2134 | 1.0 | 588 | 0.1978 | 0.9274 | 0.9274 |
| 0.1571 | 2.0 | 1176 | 0.1482 | 0.9438 | 0.9438 |
| 0.1217 | 3.0 | 1764 | 0.1512 | 0.9454 | 0.9454 |
Framework versions
- Transformers 4.57.3
- Pytorch 2.9.0+cu126
- Datasets 4.0.0
- Tokenizers 0.22.1