ibm-granite/granite-guardian-3.0-8b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kPublished:Oct 15, 2024License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Granite Guardian 3.0 8B is an 8 billion parameter fine-tuned Granite 3.0 Instruct model developed by IBM Research. It specializes in detecting risks within prompts and responses, including harm, social bias, jailbreaking, and hallucination risks in RAG pipelines. Trained on unique human-annotated and synthetic data, it outperforms other open-source models in its class on standard safety benchmarks. This model is primarily designed for enterprise risk detection and guardrailing applications.

Loading preview...

Overview

Granite Guardian 3.0 8B, developed by IBM Research, is an 8 billion parameter model specifically fine-tuned from Granite 3.0 Instruct for risk detection in AI applications. It is designed to identify various risks in both user prompts and model responses, as well as assess hallucination risks within Retrieval-Augmented Generation (RAG) pipelines.

Key Capabilities

  • Comprehensive Risk Detection: Identifies risks across multiple dimensions, including:
    • Harm, Social Bias, Jailbreaking, Violence, Profanity, Sexual Content, Unethical Behavior.
  • RAG-Specific Risk Assessment: Evaluates hallucination risks in RAG contexts by checking:
    • Context Relevance: Whether retrieved context is pertinent to the query.
    • Groundedness: If the response is supported by the provided context.
    • Answer Relevance: If the response directly addresses the user's input.
  • Performance: Outperforms other open-source models in its category on standard benchmarks, achieving an Aggregate F1 score of 0.76 for harm benchmarks and an Average AUC of 0.85 for RAG hallucination benchmarks.
  • Training Data: Utilizes unique human annotations and synthetic data informed by internal red-teaming, including samples from the hh-rlhf dataset.

When to Use This Model

Granite Guardian 3.0 8B is ideal for enterprise applications requiring robust AI safety and governance. It is particularly suited for:

  • Guardrailing: Detecting harm-related risks in user prompts or model-generated text.
  • Model Risk Assessment: Evaluating and monitoring AI risks.
  • RAG Pipeline Integrity: Ensuring context relevance, groundedness, and answer relevance in RAG systems.
  • Custom Risk Definitions: Applicable for use with custom risk definitions, though these require testing. The model is currently trained and tested only on English data.

Finetunes
3 models