mikaelnurminen/modernbert-large-phishing

Phishing Detection Space (ZeroGPU)

This Space is configured for Hugging Face ZeroGPU using the spaces framework pattern (@spaces.GPU).

Workflows

• —Email classification: paste email text and/or upload .eml.
• —URL classification: analyze a URL in a dedicated classifier flow.

ZeroGPU setup

• —In Space Settings, set hardware to ZeroGPU.
• —Keep SDK as Gradio.
• —The inference path in app.py is wrapped with @spaces.GPU for on-demand GPU execution.

Runtime env vars

• —MODEL_REPO_ID default mikaelnurminen/modernbert-large-phishing
• —MAX_LENGTH default 2048
• —EMAIL_THRESHOLD default 0.57
• —URL_THRESHOLD default 0.52
• —MIN_EMAIL_CHARS default 1

Notes

• —Full-model inference is the default path (CPU quantized export is not used by default).
• —The app uses separate thresholds for email and URL to reduce false positives.
• —Very short email greetings (for example hi) are treated as benign to avoid trivial false positives.