MedHal: An Evaluation Dataset for Medical Hallucination Detection

Gaya Mehenni, Amal Zouaq · 2025

Read MedHal: An Evaluation Dataset for Medical Hallucination Detection by Gaya Mehenni, Amal Zouaq (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-2504-08596.txt.

Description

Fabrice Lamarche, Gaya Mehenni, and co-authors release MedHal, a dataset for catching fabricated content in medical AI text, in work from twenty twenty-five. Hallucinated medical statements can mislead care with real harm. General detectors stumble on specialised medical language and facts. MedHal gathers varied medical sources and task formats at scale. Both planted and naturally arising falsehoods are included.

A baseline detector trained on MedHal beats general-purpose rivals. Training data volume supports proper model learning in-domain. Expert review demands drop when detectors pre-screen outputs. Safer medical text generation gets a dedicated proving ground. The resource invites stronger hallucination research in medicine.

Catalog record

CREATORGaya Mehenni, Amal Zouaq
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-08596

Browse the full library · About the Tower of Babel