ravi2505/ne-nasedaj-sms-phishing

Ne Nasedaj SMS Phishing Detector

Fine-tuned BERTić model for phishing SMS detection in South Slavic languages.

Overview

This model is based on classla/bcms-bertic and was fine-tuned for binary SMS classification:

• Phishing
• Legitimate

The goal is to detect phishing and scam SMS messages commonly found in Serbian-speaking and neighboring language communities.

Supported Languages

• Serbian (sr)
• Croatian (hr)
• Bosnian (bs)
• Montenegrin (me)

The model was trained using language-aware inputs by prepending language tags such as:

[sr] message text
[hr] message text

Dataset

• Total Samples: 1,529
• Legitimate Messages: 900
• Phishing Messages: 629

Language	Samples
Serbian	1013
Croatian	201
Montenegrin	153
Bosnian	143
Neutral	19

Model

Base Model:

classla/bcms-bertic

Parameters:

110,618,882

Test Results

Metric	Score
Accuracy	96.96%
Precision	96.81%
Recall	95.79%
F1 Score	96.30%

Confusion Matrix

[[132   3]
 [  4  91]]

Intended Use

• SMS phishing detection
• Cybersecurity education
• Anti-scam awareness tools
• Messaging safety applications

Example

Input:

[sr] Vaš račun će biti blokiran. Kliknite odmah.

Output:

phishing

Contributions

• Model fine-tuning, training, evaluation, and Hugging Face model publication: Ravi Chand Mutthina
• Dataset creation and curation: Otanwe- Developed as part of a collaborative hackathon project.

Disclaimer

This model is intended for research, educational, and demonstration purposes. Predictions should not be considered a replacement for professional cybersecurity review.