AbhinavMaurya/SpamShield-Datasets

SpamShield Multilingual Training Datasets

Comprehensive multilingual spam classification and threat categorization datasets used to train the SpamShield ONNX model family.

---

📋 Dataset Summary

This dataset collection contains over 200,000+ curated and anonymized conversational samples across multiple languages (English, German, Russian, Hinglish, Spanish, Arabic, etc.), categorized into binary labels (Ham vs. Spam) and 6 fine-grained threat categories.

Threat Categories:

1. phishing — Spoofed domain links, fake authentication portals, credential theft attempts.
2. job_scams — Task-based payment scams, fake remote freelance offers, deposit-based hiring.
3. crypto — Airdrop drainers, wallet phrase harvesters, fake exchange bots.
4. adult — NSFW text, solicitations, illicit media promotions.
5. marketing — Bulk promotional messaging, channel shilling, spam affiliate schemes.
6. giveaway — Fake rewards, lottery scams, fraudulent contest announcements.

---

📊 Dataset Distribution

Language	Total Messages	Normal (Ham)	Spam
English	188,182	96,209	91,973
German	2,487	1,106	1,381
Russian	1,486	636	850
Hinglish	910	825	85
Other Multilingual	10,000+	~5,500	~4,500

---

🗂️ Data Schema

Each sample is structured as JSON / JSONL with the following fields:

{
  "text": "Join our official airdrop pool to claim 500 USDT: https://fake-crypto.xyz",
  "label": 1,
  "category": "crypto",
  "language": "en"
}

• text (string): The preprocessed message content.
• label (integer): 0 for legitimate (ham), 1 for spam.
• category (string): Threat category (phishing, job_scams, crypto, adult, marketing, giveaway, or ham).
• language (string): ISO 639-1 language code.

---

🔗 Related Models

• AbhinavMaurya/SpamShield — Pre-trained ONNX models (Lite, Flash, Edge, and NSFW Vision).

📜 License

Released under the MIT License.