resect-ai/veritas-0.6B-fact-checker-non-thinking-1.0
Veritas-0.6B-Fact-Checker-Non-Thinking-1.0 by Resect Research Labs is a 0.6 billion parameter model built on the Qwen3 architecture, specifically fine-tuned for fact-checking and factual consistency verification. It achieves an average score of 72.30% on the LLM-AggreFact benchmark, representing a 7.37% improvement over the base Qwen3-0.6B model in non-thinking mode. This model is optimized for binary classification tasks to verify claims against provided documents, with a context length of 32768 tokens.
Loading preview...
Veritas-0.6B-Fact-Checker-Non-Thinking-1.0 Overview
Resect Research Labs has developed Veritas-0.6B-Fact-Checker-Non-Thinking-1.0, a 0.6 billion parameter model based on the Qwen3 architecture. This model is specifically fine-tuned and optimized for fact-checking and factual consistency verification, aiming to improve factual grounding and mitigate hallucinations in AI models through proprietary reinforcement learning and novel fine-tuning techniques.
Key Capabilities
- Specialized Fact-Checking: Optimized for binary classification to determine the factual consistency of claims against documents.
- Enhanced Factual Accuracy: Achieves an average score of 72.30% on the LLM-AggreFact benchmark, an improvement of 7.37% over the base Qwen3-0.6B model in non-thinking mode.
- Benchmark Performance: Demonstrates strong performance across various datasets within LLM-AggreFact, including significant gains in CNN, MediaS, MeetB, WiCE, Claim Verify, and RAG Truth categories.
- Integration with MiniCheck: Designed for use with the MiniCheck library, requiring a specific operating mode for optimal performance.
When to Use This Model
- Factual Verification: Ideal for applications requiring precise, binary classification of factual claims.
- Hallucination Mitigation: Useful in pipelines designed to reduce or detect AI model hallucinations by verifying generated content.
- Specific Scoring Tasks: Must be used strictly in its prescribed scoring mode, generating a binary classification based on a specified template.