MultiHal: MultiLingual Dataset for Knowledge-Graph Grounded Evaluation of LLM Hallucinations

Ernests Lavrinovics Russa Biswas Katja Hose Johannes Bjerva · 2025

Read MultiHal: MultiLingual Dataset for Knowledge-Graph Grounded Evaluation of LLM Hallucinations by Ernests Lavrinovics Russa Biswas Katja Hose Johannes Bjerva (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-14101.txt.

Description

Lavrinovics, Biswas, Hose, and Bjerva built MultiHal in 2025, a multilingual test set for catching made up facts in language model replies. Most fact checks lean on English text or loose web passages. Fact graphs, which store who did what to whom in clean form, offer firmer ground but had not been paired with many tongues.

The team mined one hundred forty thousand fact paths from open graphs and cleaned them down to about twenty six thousand sound ones. Each path can test a model reply step by step. First runs showed clear score gains when models were checked against the graphs. The set spans many languages.

Catalog record

CREATORErnests Lavrinovics Russa Biswas Katja Hose Johannes Bjerva
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-14101

Browse the full library · About the Tower of Babel