HallOumi-8B
| Property | Value |
|---|---|
| Parameter Count | 8 Billion |
| License | CC-BY-NC-4.0 |
| Base Model | Llama-3.1-8B-Instruct |
| F1 Score | 77.2% ± 2.2% |
| Demo Available | Yes (oumi.ai/halloumi-demo) |
What is HallOumi-8B?
HallOumi-8B represents a significant breakthrough in AI hallucination detection, achieving state-of-the-art performance with only 8 billion parameters. Developed by Oumi AI, this model outperforms much larger models including DeepSeek R1 (671B parameters), Claude Sonnet 3.5, and Google Gemini 1.5 Pro. It's specifically designed to verify content at a sentence level, providing detailed citations and human-readable explanations for its determinations.
Implementation Details
The model is built on Llama-3.1-8B-Instruct architecture and trained using multiple specialized datasets including oumi-synthetic-document-claims and oumi-anli-subset. It processes input through a structured format using context, request, and response tags, enabling precise verification of claims against provided source materials.
- Sentence-level verification capability
- Confidence scoring system
- Citation tracking and traceability
- Human-readable explanations for verification decisions
Core Capabilities
- Achieves 77.2% Macro F1 Score in hallucination detection
- Provides detailed context-based verification
- Generates human-readable explanations
- Supports document-based claim verification
- Offers confidence scoring for each verification
Frequently Asked Questions
Q: What makes this model unique?
HallOumi-8B stands out for achieving superior performance with a significantly smaller parameter count (8B) compared to competitors, while providing detailed explanations and citations for its verifications. This makes it both more efficient and more practical for real-world applications.
Q: What are the recommended use cases?
The model is specifically designed for verifying claims and detecting hallucinations in scenarios where a known source of truth is available. It's particularly useful for content verification, fact-checking, and ensuring AI-generated content remains truthful and accurate.
