RAGTruth

RAGTruth is a word-level hallucination corpus for Retrieval-Augmented Generation (RAG) used to train and evaluate language models. It contains nearly 18,000 naturally generated responses from diverse LLMs employing RAG, with meticulous manual annotations at the case- and word-level and evaluations of hallucination intensity. The project provides training and evaluation code alongside the annotated dataset, with updates expanding the data and meta annotations.

AI named RAGTruth in February 2026.