Enoki · Multi-Level Hallucination Detection

Paste an evidence document and a generated answer. Enoki extracts (subject, predicate, object) facts from the answer with an encoder-based OpenIE model, verifies each one against the evidence with NLI, and projects the failures back to the exact words that caused them.

Highlighted spans mean not supported by the evidence you provided — not necessarily false.

Paper · Code · Encoder · NLI verifier · EnokiQA

Evidence / context

Generated answer to check

Answer, with unsupported spans highlighted

Every extracted fact and its verification verdict

Examples from the paper, the model card, and the EnokiQA test split

Evidence / context	Generated answer to check