Granite guardian models

The IBM Granite Guardian 3.2 models are designed to detect risks in prompts and/or responses. They can help with risk detection along many key dimensions catalogued in the IBM AI Risk Atlas. They are trained on unique data comprising human annotations and synthetic data informed by internal red-teaming, and they outperform other open-source models in the same space on standard benchmarks.

We’re introducing new model sizes for Granite Guardian 3.2, including a variant derived from our 3B-A800M mixture of experts (MoE) language model. The new models offer increased efficiency with minimal loss in performance.

Running

3B:

ollama run ibm/granite3.2-guardian:3b>>> /set system profanity

8B:

ollama run ibm/granite3.2-guardian:5b >>> /set system violence

The model will produce a single output token, either Yes or No. By default, the general-purpose harm category is used, but other categories can be selected by setting the system prompt.

Supported use cases

• Risk detection in prompt text or model response (i.e. as guardrails), such as:

  • Harm (harm): content considered generally harmful
    • Note: this is the default setting for the model’s chat template.
  • Social Bias (social_bias): prejudice based on identity or characteristics
  • Jailbreaking (jailbreak): deliberate instances of manipulating AI to generate harmful, undesired, or inappropriate content
  • Violence (violence): content promoting physical, mental, or sexual harm
  • Profanity (profanity): use of offensive language or insults
  • Sexual Content (sexual_content): explicit or suggestive material of a sexual nature
  • Unethical Behavior (unethical_behavior): actions that violate moral or legal standards

• RAG (retrieval-augmented generation) to assess:

  • Context relevance (relevance): whether the retrieved context is relevant to the query
  • Groundedness (groundedness): whether the response is accurate and faithful to the provided context
  • Answer relevance (answer_relevance): whether the response directly addresses the user’s query

• Agentic Workflows to assess:

  • Function Calling Hallucination (function_calling) : validates use of function calls for syntactic and semantic hallucination.

Granite dense models

The Granite dense models are designed to support tool-based use cases and for retrieval augmented generation (RAG), streamlining code generation, translation and bug fixing.

Granite mixture of experts models

The Granite MoE models are designed for low latency usage and to support deployment in on-device applications or situations requiring instantaneous inference.

Learn more

• Release Date: February 26th, 2025
• License: Apache 2.0
• https://www.ibm.com/granite