future7/CogniDet

CogniDet: Cognitive Faithfulness Detector

CogniDet is an 8 billion parameter model, built on a Llama3 backbone, specifically engineered to identify factual and cognitive hallucinations in Large Language Model (LLM) outputs. Developed within the CogniBench framework, it addresses the complex challenge of evaluating inference-based statements, moving beyond simple fact-checking.

Key Capabilities

• Dual Hallucination Detection: Accurately identifies both factual inconsistencies (claims contradicting context) and cognitive hallucinations (unsupported inferences or evaluations).
• Legal-Inspired Rigor: Employs a unique tiered evaluation framework (Rational → Grounded → Unequivocal) derived from legal evidence standards, enhancing the depth of analysis.
• Efficient Inference: Provides single-pass detection, making it faster than traditional NLI-based methods for hallucination identification.
• Strong Performance: Achieves an overall F1 score of 70.30, with a notable 73.80 F1 score for cognitive hallucination detection, outperforming baselines such as SelfCheckGPT and RAGTruth.
• Large-Scale Training: Trained on the extensive CogniBench-L dataset, comprising over 24,000 dialogues and 234,000+ annotated sentences across diverse domains.

Good For

• Evaluating LLM Faithfulness: Ideal for developers and researchers needing to assess how faithfully an LLM adheres to provided context and avoids generating unsupported inferences.
• Research in LLM Reliability: Useful for studies focusing on the robustness and trustworthiness of LLM outputs, particularly in critical applications.
• Context-Grounded Applications: Best suited for scenarios where LLM responses must be strictly grounded in given information, such as knowledge retrieval or summarization tasks.

Limitations

• Primarily optimized for English knowledge-grounded dialogues.
• May require fine-tuning for highly domain-specific applications.
• Context window is limited to 8K tokens.