Granite Guardian 3.2 5B

About

IBM Granite Guardian 3.2 5B is a 'thinned' safety model with 30% fewer parameters than Granite Guardian 3.1 8B while maintaining comparable performance. Detects: harm, social bias, jailbreaking, violence, profanity, sexual content, unethical behavior/evasiveness, and RAG/function-calling hallucinations. F1 0.784 on aggregated harm benchmarks. Trained on human-annotated data (DataForce) + synthetic red-team data. English only. Apache 2.0.

Capabilities

Rankings