Introducing v0.5 of the AI Safety Benchmark from MLCommons

• DOI: 10.48550/arxiv.2404.12241
• Publication Date: 2024-04-18
• Authors: Bertie Vidgen; Adarsh Agrawal; Ahmed Mohamed Ahmed; Victor Akinwande; Namir Al-Nuaimi; Najla Alfaraj; Elie Alhajjar; Lora Aroyo; Trupti Bavalatti; Borhane Blili-Ham...; Kurt Bollacker; Rishi Bomassani; Marisa Ferrara Bo...; Siméon Campos; Kal Chakra; Canyu Chen; Cody Coleman; Zacharie Delpierr...; Leon Derczynski; Debojyoti Dutta; Ian Eisenberg; James Ezick; Heather Frase; Brian Fuller; Ram Gandikota; Agasthya Gangavarapu; Ananya Gangavarapu; James Gealy; Rajat Ghosh; James Goel; Usman Gohar; Sujata Goswami; Scott A. Hale; Wiebke Toussaint ...; Joseph Marvin Imp...; Surgan Jandial; Nick Judd; Felix Juefei-Xu; Foutse Khomh; Bhavya Kailkhura; Hannah Rose Kirk; Kevin Klyman; Chris Knotz; Michael Kuchnik; Shachi H Kumar; Chris Lengerich; Bo Li; Zeyi Liao; Eileen Peters Long; Victor M. Lu; Yifan Mai; Priyanka Mary Mammen; Kelvin Manyeki; Sean McGregor; Virendra Mehta; Shafee Mohammed; Emanuel Moss; Lama Nachman; Dinesh Jinenhally...; Amin Nikanjam; Besmira Nushi; Luis Oala; Iftach Orr; Alicia Parrish; Cigdem Patlak; William Pietri; Forough Poursabzi...; E. Presani; Fabrizio Puletti; Paul Röttger; Saurav Sahay; Tim Santos; Nino Scherrer; Alice Schoenauer ...; Patrick Schramowski; Abolfazl Shahbazi; Vin Sharma; Xudong Shen; Vamsi Sistla; Leonard Tang; Davide Testuggine; Vithursan Thangarasa; Elizabeth Anne Wa...; Rebecca Weiss; Chris Welty; Tyler Wilbers; Adina Williams; Carole-Jean Wu; Poonam Yadav; Xianjun Yang; Yi Zeng; Wenhui Zhang; Fedor Zhdanov; Jiacheng Zhu; Percy Liang; Peter Mattson; Joaquin Vanschoren
• Keywords: Benchmark (surveying)

This paper introduces v0.5 of the AI Safety Benchmark, which has been created by MLCommons Working Group. The Benchmark designed to assess safety risks systems that use chat-tuned language models. We introduce a principled approach specifying and constructing benchmark, for covers only single case (an adult chatting general-purpose assistant in English), limited set personas (i.e., typical users, malicious vulnerable users). new taxonomy 13 hazard categories, 7 have tests benchmark. plan release version 1.0 end 2024. v1.0 benchmark will provide meaningful insights into systems. However, should not be used sought fully document limitations, flaws, challenges v0.5. includes (1) comprises cases, types under test (SUTs), context, personas, tests, items; (2) categories with definitions subcategories; (3) seven each comprising unique items, i.e., prompts. There are 43,090 items total, we templates; (4) grading system against benchmark; (5) an openly available platform, downloadable tool, called ModelBench can evaluate on (6) example evaluation report benchmarks performance over dozen models; (7) specification