A Tale of Two Hallucinations: Empirical Evidence for Distinct Failure Modes in Large Language Models Dataset

This dataset is a comprehensive test suite designed to empirically validate and evaluate two distinct types of AI hallucinations: Type 1 (Uncertainty-driven) and Type 2 (Confidence-driven). The data was collected from controlled experiments involving six model families (including Perplexity, Claude Sonnet 4, ChatGPT 5, Deepseek, ChatGPT 4o, and Gemini 2.5 Pro) and a total of 120 test cases.