shariul-islam/bengali-sms-smishing-dataset

Bengali SMS Smishing Dataset

A multilingual SMS dataset for phishing (smishing) detection, covering Bengali, English, Banglish, and Code-Mixed linguistic varieties. Developed as part of the SmishDetect-LLM research framework.

Hugging Face: shariul-islam/bengali-sms-smishing-dataset

---

Dataset Summary

This dataset contains 7,005 SMS messages annotated across three classification categories and four linguistic varieties. It is the first publicly available smishing dataset to cover Banglish and Code-Mixed SMS alongside Bengali and English, making it uniquely suited for multilingual cybersecurity research.

---

Dataset Statistics

Splits

Split	Examples	Size
Train	5,604	~880 KB
Validation	700	~109 KB
Test	1,401	~215 KB
Total	7,005	~1.2 MB

Labels

Each message is assigned one of three categories:

Label	Description
smish	Fraudulent SMS designed to deceive recipients into revealing sensitive information or clicking malicious links
promo	Legitimate marketing or commercial SMS messages
normal	Regular, benign SMS messages with no malicious or promotional intent

Linguistic Varieties

Variety	Description
Bengali	SMS messages written in native Bengali (Bangla) script
English	SMS messages written in standard English
Banglish	Bengali language content written using Roman/Latin script
Code-Mixed	Messages that mix Bengali and English within the same text

---

Dataset Structure

Features

{
    "label":  str,   # Classification category: "smish", "promo", or "normal"
    "text":   str,   # Raw SMS message content
    "source": str,   # Linguistic variety: Bengali, English, Banglish, or Code-Mixed
}

Data Files

data/
├── train-*
├── validation-*
└── test-*

---

Usage

Load with 🤗 Datasets

from datasets import load_dataset

dataset = load_dataset("shariul-islam/bengali-sms-smishing-dataset")

# Access individual splits
train_data      = dataset["train"]
validation_data = dataset["validation"]
test_data       = dataset["test"]

Load a specific split

train = load_dataset("shariul-islam/bengali-sms-smishing-dataset", split="train")

Filter by linguistic variety

bengali_only = dataset["train"].filter(lambda x: x["source"] == "Bengali")

Filter by label

smishing_only = dataset["train"].filter(lambda x: x["label"] == "smishing")

---

Example Entries

{
    "label": "smish",
    "text": "আপনার ব্যাংক অ্যাকাউন্ট সাময়িকভাবে বন্ধ করা হয়েছে। এখনই যাচাই করুন: http://fake-link.com",
    "source": "Bengali"
}

{
    "label": "normal",
    "text": "Apni ki kal free acho? Amra ektu bairer khabar khete jabo.",
    "source": "Banglish"
}

{
    "label": "promo",
    "text": "Exclusive offer! Get 50% off on all items this weekend only. Visit our store now.",
    "source": "English"
}

---

Intended Uses

• Smishing / SMS phishing detection — binary or multi-class classification
• Multilingual NLP — model evaluation across Bengali, English, Banglish, and Code-Mixed text
• Low-resource language research — Banglish and Code-Mixed data are scarce in existing benchmarks
• Cybersecurity research — studying phishing patterns across linguistic and cultural contexts

---

Associated Framework

This dataset was developed alongside the SmishDetect-LLM framework, which benchmarks BERT-based and GPT-based models on this data.

• 📦 Code Repository: github.com/shariul-islam-au/SmishDetect-LLM
• 📓 Experiments: Four Google Colab notebooks covering Baseline and Proposed versions for both model families are available in the repository.

---

Citation

If you use this dataset in your research, please cite:

@dataset{islam2025smishing,
  title     = {Bengali SMS Smishing Dataset: A Multilingual Corpus for SMS Phishing Detection},
  author    = {Shariul Islam},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/shariul-islam/bengali-sms-smishing-dataset}
}

---

License

This dataset is released under the MIT License.

---

Contact

Shariul Islam
Email: shariul.islam.au@gmail.com Hugging Face: @shariul-islam