Omartificial-Intelligence-Space / SA-BERT-Classifier

SA-BERT-Classifier: Saudi Dialect Classifier

Model Description

SA-BERT-Classifier is a binary classifier that distinguishes between Saudi and non-Saudi Arabic dialects. Built on top of the SA-BERT-V1 embeddings, this model achieves high accuracy in identifying Saudi dialectal expressions across various domains and contexts.

Intended Use

This model is designed for:

• Dialect identification in Arabic text
• Content filtering for region-specific applications
• Improving NLP pipelines for Saudi audience targeting
• Research on dialectal variations in Arabic

Performance Metrics

The model achieves the following performance on our test set:

Usage

Using the Hugging Face Transformers Pipeline

import os
import torch
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    TextClassificationPipeline
)

# Configuration
MODEL_ID = "Omartificial-Intelligence-Space/SA-BERT-Classifier"
HF_TOKEN = os.getenv("HUGGINGFACE_HUB_TOKEN", "<YOUR_TOKEN_HERE>")
DEVICE = 0 if torch.cuda.is_available() else -1

# Load tokenizer & model
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_auth_token=HF_TOKEN)
model = AutoModelForSequenceClassification.from_pretrained(
    MODEL_ID, use_auth_token=HF_TOKEN
).to("cuda" if DEVICE == 0 else "cpu")

# Build the pipeline
pipeline = TextClassificationPipeline(
    model=model,
    tokenizer=tokenizer,
    device=DEVICE,
    return_all_scores=True
)

# Example
text = "السلام عليكم ورحمة الله كيف حالك اليوم؟"
results = pipeline(text)[0]

# Format results
scores = {int(item["label"].split("_")[-1]): item["score"] for item in results}
p_non_saudi = scores.get(0, 0.0)
p_saudi = scores.get(1, 0.0)
prediction = "Saudi" if p_saudi > p_non_saudi else "Non-Saudi"

print(f"Text: {text}")
print(f"P(Non-Saudi): {p_non_saudi:.4f}")
print(f"P(Saudi): {p_saudi:.4f}")
print(f"Prediction: {prediction}")

Training Parameters

• Embedding model : Omartificial-Intelligence-Space/SA-BERT-V1
• Max sequence length : 256
• Classifier : Logistic Regression with balanced class weights
• Training split : 80% train, 20% test (stratified)

Example Results

Here are some example predictions from our test set:

Analysis

The classifier demonstrates several noteworthy characteristics:

1. High confidence predictions : The model often predicts with very high confidence (near 0.0 or 1.0)
2. Dialectal markers : Expressions like "مو" (not), "وش" (what), "عشان" (because) are strong Saudi dialect indicators
3. MSA (Modern Standard Arabic) sensitivity : Formal, MSA-heavy sentences tend to be classified as non-Saudi, regardless of content
4. Lexical features : Saudi-specific vocabulary (e.g., references to places like "جازان", "العلا") increases Saudi classification probability

Limitations

• The model may perform less effectively on mixed-dialect text or code-switching between MSA and dialect
• Very short text with limited dialectal markers may yield less reliable results
• Performance may vary for specialized domains not well-represented in the training data
• The binary classification (Saudi/non-Saudi) does not distinguish between specific non-Saudi dialects

Citation

If you use this model in your research or applications, please cite:

@misc{nacar2025marbertv2saclassifier,
  title={SA-BERT-Classifier: Saudi Dialect Classifier},
  author={Nacar, Omer},
  year={2025},
  publisher={Omartificial-Intelligence-Space},
  howpublished={\url{https://huggingface.co/Omartificial-Intelligence-Space/MarBERTv2-SA-Classifier}},
}