yhile-spam-lr-v1
Spam and scam text classifier
About
yhile-spam-lr-v1 detects spam, phishing, and cryptocurrency scam messages, including giveaway scams, pig-butchering schemes, and wallet phishing. It is a scikit-learn logistic regression model over TF-IDF word and character n-grams, plus hand-built signals for URLs, crypto wallet addresses, urgency and money keywords, and scam categories.
Intended use
Flagging spam and scam messages in SMS and email text, as a small and inspectable baseline, and as a reference for building your own text classifier. The decision threshold is tuned for at least 95% precision to keep false alarms low.
Limitations
About 3.1 to 3.6% of legitimate messages are flagged as spam, and legitimate marketing and transactional email can share vocabulary with spam. AI-generated scam text is underrepresented in the training data, and the model reads one message at a time, without conversation history or email metadata. Wallet address checks validate the format only. English only.
Demo
Runs entirely in your browser. What you type is never uploaded or stored.
A small, fast model: about 97% accurate on its test set, so it will still miss some scams and flag some real messages. It is not a security product.
Demo coming soon.
Details
| Architecture | Logistic regression (scikit-learn) on TF-IDF word 1-2-grams (8,000 features) and character 3-5-grams (4,000 features), plus engineered URL, wallet, keyword, and scam-category signals. Saved with skops. |
|---|---|
| Training data | About 41,000 deduplicated rows (59% legitimate, 41% spam): SMS Spam Collection, Enron-Spam, SpamAssassin public corpus, and about 65 handcrafted examples. |
| Metrics | On a 20% stratified test split (about 8,235 rows): accuracy 97.1%, ROC-AUC 0.995, spam precision 96%, spam recall 97%, false positive rate 3.1 to 3.6%. |
| License | MIT |
| File size | 9.0 MB (yhile-spam-lr-v1.skops) |