Granite Guardian

About

The Granite Guardian family of language models is engineered to identify and mitigate risks within user prompts and responses by serving as safety guardrails. These models, developed by IBM, are adept at pinpointing issues such as social bias, hate speech, toxicity, and other harmful content, as categorized in IBM's AI Risk Atlas. Additionally, they offer specific checks for concerns related to Retrieval Augmented Generation (RAG), such as groundedness and context relevance. Available in 8B and 2B parameter sizes, Granite Guardian models can be integrated with any open or proprietary language models to bolster AI safety and responsible use practices. Notably, in testing, the Granite Guardian 3.0 8B model surpassed Meta's LlamaGuard 3 8B by improving the average F1-score by 4 points on standard risk detection benchmarks 345.