ibm-granite/granite-guardian-3.2-8b-factuality-detection

Overview

Granite Guardian 3.2 8B Factuality Detection, developed by IBM Research, is an 8 billion parameter model fine-tuned from ibm-granite/granite-3.2-8b-instruct. Its primary function is to assess the factuality of an LLM-generated response relative to a given context. The model outputs a binary label: "Yes" if the response is unfactual, and "No" if it is factual. This model is part of an ongoing research effort focused on post-generation mitigation and is intended for risk detection in enterprise applications.

Key Capabilities

• Factuality Detection: Identifies if an LLM's response contains information not supported by or directly contradicted by the provided context.
• Contextual Analysis: Evaluates responses based on a specific context, making it suitable for RAG (Retrieval Augmented Generation) scenarios.
• Binary Labeling: Provides clear "Yes" (unfactual) or "No" (factual) outputs, along with a confidence level.
• Experimental Research Model: Actively under development with a commitment to continuous improvement.

Training and Limitations

The model was trained using synthetic data generated from the ELI5-Category dataset, utilizing a FactCorrector pipeline that leverages FactReasoner for factuality assessment. It is specifically optimized for factuality risk and is trained and tested exclusively on English data. Users should be aware that there is no built-in safeguard to guarantee 100% correct detection, and it is intended for use strictly in its prescribed scoring mode with a temperature of 0 for deterministic responses. For broader risk detection, users are advised to consult the latest Granite Guardian models.