The model runs entirely in your browser, so messages you check never leave your device. It's an 82 MB download the first time; after that it's cached.
Known weakness: genuine service messages (one-time codes, delivery updates, bank alerts) are often flagged, because the public training data has the scams that imitate them but almost none of the real ones. Why
Why
Each word is scored by removing it and measuring how far the scam log-odds move. Red words pushed towards scam and green words pushed away from it.
Fine-tuned model vs Gemini
12 everyday messages, written by hand
Genuine service texts the training data barely contains, and scams with no link. Illustrative, not a benchmark, but it shows where each approach fails.
Exact prompts sent to Gemini
Model card
Loading…
How it compares (campaigns held out of training)
Macro-F1 averages the three classes. "2023-25 phishing" is the share of phishing emails from 2023 to 2025 flagged as scams by models trained only on older data. The methodology, figures and error analysis are in the report.