SafeWorld : Geo-Diverse Safety Alignment

Da Yin, Haoyi Qiu *, Kung-Hsiang Huang, Kai-Wei Chang, Nanyun Peng · 2024

Read SafeWorld : Geo-Diverse Safety Alignment by Da Yin, Haoyi Qiu *, Kung-Hsiang Huang, Kai-Wei Chang, Nanyun Peng (2024) 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-2412-06483.txt.

Description

Da Yin and co-authors introduce SafeWorld, a benchmark for whether language models respect safety norms that vary by place, in a 2024 paper. Safety work often overlooks that cultural and legal standards differ across the world. A reply legal in one country may break norms in another.

The benchmark holds 2,342 test queries grounded in human-verified norms and laws from fifty countries and many regions and groups. An automatic evaluation frame scores replies for fit to context, accuracy, and completeness. Tests show current models struggle to meet these criteria. The authors also synthesize training data to improve alignment with geo-diverse standards.

Catalog record

CREATORDa Yin, Haoyi Qiu *, Kung-Hsiang Huang, Kai-Wei Chang, Nanyun Peng
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE78 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-06483

Browse the full library · About the Tower of Babel