KnowledgeCheck

KnowledgeCheck is a hallucination detection dataset in the medical domain, constructed based on WiMedKG~\\cite{WiMedKG} and CPubMed~\\cite{cpubmed2025}knowledge graphs.Unlike surface-level fact matching tasks, it often requires one or more reasoning steps to assess the correctness of a claim.To support such complex, inference-dependent hallucination detection, we construct the dataset based on a medical knowledge graph, which can represent reasoning processes through logical rules on nodes and edges.We categorize common reasoning types into four classes: One-hop, Conjunction, Multi-hop, and Multi-path according to structural patterns in the graph. These reasoning patterns are then instantiated with medical knowledge to generate natural language claims with specific reasoning requirements. Each claim is labeled as either supported or refuted and can be verified against the medical knowledge graph.