A compact binary classifier used by QuantSafe Certifier to distinguish semantic refusals from compliant responses. Label 1 is refusal; label 0 is compliance.

Training

Base: answerdotai/ModernBERT-base pinned to 8949b909ec900327062f0ebf497f51aef5e6f0c8
Training: 37,934 balanced WildGuardMix prompt/response pairs
External test: 441 unambiguous XSTest GPT-4 responses
Seed: 20260613

External XSTest result

Method	Accuracy	Macro F1	Refusal F1
This model	0.9773	0.9773	0.9760
QuantSafe legacy opener lexicon	0.5261	0.4124	0.1538

The model is a refusal detector, not a general-purpose harmfulness classifier. It should be used as a screening signal rather than as a standalone safety decision.

Crusadersk/quantsafe-refusal-modernbert

QuantSafe Refusal ModernBERT

Training

External XSTest result

Sources

Crusadersk/
quantsafe-refusal-modernbert