HallOumi-8B: A Specialized Hallucination Detection Model

HallOumi-8B, developed by Oumi AI, is an 8 billion parameter model designed for state-of-the-art hallucination detection. It significantly outperforms larger and closed-source models such as DeepSeek R1 (671B), Claude Sonnet 3.5, OpenAI o1, and Google Gemini 1.5 Pro, achieving a Macro F1 Score of 77.2% ± 2.2%.

Key Capabilities

Per-sentence verification: Analyzes content (AI or human-generated) at a granular sentence level.
Contextual support determination: Identifies whether a statement is supported or unsupported by provided context, along with a confidence score.
Sentence-level citations: Provides relevant context sentences to justify its determination.
Human-readable explanations: Offers explanations for why a claim is supported or unsupported, aiding human review.
Trust building: Aims to address the critical issue of AI hallucinations by enabling verifiable and traceable outputs.

Use Cases

Claim verification: Ideal for scenarios where a known source of truth is available to verify claims.
Enhancing AI trust: Helps in safely and responsibly deploying generative models by ensuring output veracity.
Mitigating risks: Addresses issues like AI-generated misinformation in legal, customer service, and general information contexts.

This model is fine-tuned from Llama-3.1-8B-Instruct and utilizes a combination of synthetic and ANLI/C2D-D2C subsets for training. It is intended for claim verification and should not be used for purposes outside of this scope due to the inherent limitations of smaller LLMs.