AraTrust

522 human-written Arabic three-choice trustworthiness questions across eight categories; HELM scores exact match on generated option letters.

Also known as: Ara Trust, asas-ai/AraTrust

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorysafety
SubcategoryArabic three-choice trustworthiness QA (8 categories)
Page statusactive
Metricaccuracy (paper); exact_match (HELM)
Directionhigher_is_better
Unit%
Dataset size522
Dataset licenceMIT
PublisherASAS AI, with King Abdulaziz University, University College London, University of Illinois Urbana-Champaign, and Alexandria University

What it measures

AraTrust asks a model to pick the trustworthy answer to a short Arabic multiple-choice question. Authors wrote most items by hand. Eight categories cover commonsense truthfulness, ethics, physical health, mental health, unfairness, illegal activity, privacy and offensive language. It is a knowledge-and-values quiz in Arabic, not a jailbreak chat and not the later MBZUAI Arabic safeguard set (arXiv:2410.17040) that HELM's schema text incorrectly cites.

Task format

Three-option MCQ. Options in the file are prefixed with أ) ب) ج). HELM uses joint multiple-choice generation with Arabic instructions and exact_match on the letter. The paper also reports zero-shot, one-shot, few-shot and zero-shot chain-of-thought settings. One item has only two options; HELM drops empty choice fields.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub