@sabaki-js/phishing-detection

Sabaki plugin — multilingual phishing detection (latin scripts: de/en/es/fr/id/tr/vi).

• Bundled mini mode: MiniLM-L6 cross-encoder ONNX FP32 (~90 MB in models/phishing-latin-fp32, no download at runtime).
• Question suspicious (also accepts Is this suspicious?), options yes / no.
• Specialist trained on 4 100 rows. Held-out eval (100/lang, never seen in training):

| lang | acc | P | R | F1 | |------|-----|---|---|----| | de | 0.970 | 0.961 | 0.980 | 0.970 | | es | 0.990 | 1.000 | 0.980 | 0.990 | | id | 1.000 | 1.000 | 1.000 | 1.000 | | tr | 1.000 | 1.000 | 1.000 | 1.000 | | vi | 0.930 | 0.877 | 1.000 | 0.935 | | en | 0.850 | 0.973 | 0.720 | 0.828 | | fr | 0.830 | 0.789 | 0.900 | 0.841 |

Out of scope (documented): it/pt (weak recall), non-latin scripts (base limitation — see multilingual track).

Install

npm i @sabaki-js/phishing-detection sabaki onnxruntime-node

import plugin from "@sabaki-js/phishing-detection";
const r = await plugin.classifyAsync({ state: "SMS: régularisez 96 EUR via ce lien" });
// { decision: "yes", confidence: 0.97, domain: "phishing", ... }

Swap the model

SABAKI_PHISHING_DIR env (same layout: generalist.onnx + tokenizer/):

SABAKI_PHISHING_DIR=/path/to/my-weights node app.js

License: MIT.