rudycaz/modernbert-phish-detector

ModernBERT Phishing Detector

Model description

This model is a fine-tuned ModernBERT-base binary sequence classifier for phishing email detection. It takes a full email as input text and predicts whether the email is safe or phishing.

The training backbone is answerdotai/ModernBERT-base, and the final release includes:

• a fine-tuned PyTorch checkpoint
• an ONNX export
• a quantized INT8 ONNX export
• a calibration file for mapping logits to a user-facing phishing score

Intended use

This model is intended for:

• phishing detection in email text
• mobile or backend inference through ONNX Runtime
• UI risk scoring, such as a 0–10 or 1–10 phishing scale
• evidence extraction via sentence-level rescoring

This model is not intended for:

• malware analysis
• attachment sandboxing
• URL detonation
• image/PDF threat inspection
• general prompt-injection detection
• fully explainable token-level rationale extraction

Inputs

The model expects a single text string representing the email content.

Example format:

Subject: Urgent Account Notice

Body:
Your account has been locked. Please reply with your password immediately to restore access.