MedFabric and EtHER: A Data-Centric Framework for Word-Level Fabrication Generation and Detection in Medical LLMs

A new data-centric pipeline generates realistic, word-level fabrications for medical LLMs, preserving syntactic and stylistic fidelity while introducing subtle factual deviations. The resulting dataset, MedFabric, addresses limitations in existing medical hallucination datasets, which inadequately capture fabrication phenomena due to limited coverage, stylistic disparities, and distributional drift. Building on MedFabric, the modular detector EtHER integrates Text2Table Decomposition, Word Masking and Filling, and Hybrid Sentence Pair Evaluation to enhance factual accuracy. The framework targets the high-risk problem of LLMs generating factually incorrect yet fluent statements in medical contexts.