Introducing v0.5 of the AI Safety Benchmark from MLCommons

Bertie Vidgen; Adarsh Agrawal; Victor Akinwande; Namir Al-Nuaimi; Najla Alfaraj; Elie Alhajjar; Lora Aroyo; Trupti Bavalatti; Borhane Blili-Hamelin; Kurt Bollacker; Rishi Bomassani; Marisa Ferrara Boston; Siméon Campos; Kal Chakra; Canyu Chen; Cody Coleman; Zacharie Delpierre Coudert; 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 Hutiri; Joseph Marvin Imperial; 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; Zeyi Liao; Eileen Peters Long; Victor Lu; Yifan Mai; Priyanka Mary Mammen; Kelvin Manyeki; Sean McGregor; Virendra Mehta; Shafee Mohammed; Emanuel Moss; Lama Nachman; Dinesh Jinenhally Naganna; Amin Nikanjam; Besmira Nushi; Luis Oala; Iftach Orr; Alicia Parrish; Cigdem Patlak; William Pietri; Forough Poursabzi-Sangdeh; Eleonora Presani; Fabrizio Puletti; Paul Röttger; Saurav Sahay; Tim Santos; Nino Scherrer; Alice Schoenauer Sebag; Patrick Schramowski; Abolfazl Shahbazi; Vin Sharma; Xudong Shen; Vamsi Sistla; Leonard Tang; Davide Testuggine; Vithursan Thangarasa; Elizabeth Anne Watkins; 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

Introducing v0.5 of the AI Safety Benchmark from MLCommons

Bertie Vidgen, Adarsh Agrawal, Victor Akinwande, Namir Al-Nuaimi, Najla Alfaraj, Elie Alhajjar, Lora Aroyo, Trupti Bavalatti, Borhane Blili-Hamelin, Kurt Bollacker, Rishi Bomassani, Marisa Ferrara Boston, Siméon Campos, Kal Chakra, Canyu Chen, Cody Coleman, Zacharie Delpierre Coudert, Leon Derczynski, Debojyoti Dutta, Ian EisenbergJames 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 Hutiri, Joseph Marvin Imperial, 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, Zeyi Liao, Eileen Peters Long, Victor Lu, Yifan Mai, Priyanka Mary Mammen, Kelvin Manyeki, Sean McGregor, Virendra Mehta, Shafee Mohammed, Emanuel Moss, Lama Nachman, Dinesh Jinenhally Naganna, Amin Nikanjam, Besmira Nushi, Luis Oala, Iftach Orr, Alicia Parrish, Cigdem Patlak, William Pietri, Forough Poursabzi-Sangdeh, Eleonora Presani, Fabrizio Puletti, Paul Röttger, Saurav Sahay, Tim Santos, Nino Scherrer, Alice Schoenauer Sebag, Patrick Schramowski, Abolfazl Shahbazi, Vin Sharma, Xudong Shen, Vamsi Sistla, Leonard Tang, Davide Testuggine, Vithursan Thangarasa, Elizabeth Anne Watkins, 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

Research output: Working paper / Preprint › Preprint

117 Downloads (Pure)

Abstract

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

Original language	English
Publisher	arXiv
Publication status	Published - 18 Apr 2024

Keywords

cs.CL
cs.AI

Access to Document

2404.12241v1

Cite this

Vidgen, B, Agrawal, A, Akinwande, V, Al-Nuaimi, N, Alfaraj, N, Alhajjar, E, Aroyo, L, Bavalatti, T, Blili-Hamelin, B, Bollacker, K, Bomassani, R, Boston, MF, Campos, S, Chakra, K, Chen, C, Coleman, C, Coudert, ZD, Derczynski, L, Dutta, D, Eisenberg, I, Ezick, J, Frase, H, Fuller, B, Gandikota, R, Gangavarapu, A, Gangavarapu, A, Gealy, J, Ghosh, R, Goel, J, Gohar, U, Goswami, S, Hale, SA, Hutiri, W, Imperial, JM, Jandial, S, Judd, N, Juefei-Xu, F, Khomh, F, Kailkhura, B, Kirk, HR, Klyman, K, Knotz, C, Kuchnik, M, Kumar, SH, Lengerich, C, Liao, Z, Long, EP, Lu, V, Mai, Y, Mammen, PM, Manyeki, K, McGregor, S, Mehta, V, Mohammed, S, Moss, E, Nachman, L, Naganna, DJ, Nikanjam, A, Nushi, B, Oala, L, Orr, I, Parrish, A, Patlak, C, Pietri, W, Poursabzi-Sangdeh, F, Presani, E, Puletti, F, Röttger, P, Sahay, S, Santos, T, Scherrer, N, Sebag, AS, Schramowski, P, Shahbazi, A, Sharma, V, Shen, X, Sistla, V, Tang, L, Testuggine, D, Thangarasa, V, Watkins, EA, Weiss, R, Welty, C, Wilbers, T, Williams, A, Wu, C-J, Yadav, P, Yang, X, Zeng, Y, Zhang, W, Zhdanov, F, Zhu, J, Liang, P, Mattson, P & Vanschoren, J 2024 'Introducing v0.5 of the AI Safety Benchmark from MLCommons' arXiv.

@techreport{ee92a9dae8df4a419496020210c79f33,

title = "Introducing v0.5 of the AI Safety Benchmark from MLCommons",

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

keywords = "cs.CL, cs.AI",

author = "Bertie Vidgen and Adarsh Agrawal and Victor Akinwande and Namir Al-Nuaimi and Najla Alfaraj and Elie Alhajjar and Lora Aroyo and Trupti Bavalatti and Borhane Blili-Hamelin and Kurt Bollacker and Rishi Bomassani and Boston, {Marisa Ferrara} and Sim{\'e}on Campos and Kal Chakra and Canyu Chen and Cody Coleman and Coudert, {Zacharie Delpierre} and Leon Derczynski and Debojyoti Dutta and Ian Eisenberg and James Ezick and Heather Frase and Brian Fuller and Ram Gandikota and Agasthya Gangavarapu and Ananya Gangavarapu and James Gealy and Rajat Ghosh and James Goel and Usman Gohar and Sujata Goswami and Hale, {Scott A.} and Wiebke Hutiri and Imperial, {Joseph Marvin} and Surgan Jandial and Nick Judd and Felix Juefei-Xu and Foutse Khomh and Bhavya Kailkhura and Kirk, {Hannah Rose} and Kevin Klyman and Chris Knotz and Michael Kuchnik and Kumar, {Shachi H.} and Chris Lengerich and Zeyi Liao and Long, {Eileen Peters} and Victor Lu and Yifan Mai and Mammen, {Priyanka Mary} and Kelvin Manyeki and Sean McGregor and Virendra Mehta and Shafee Mohammed and Emanuel Moss and Lama Nachman and Naganna, {Dinesh Jinenhally} and Amin Nikanjam and Besmira Nushi and Luis Oala and Iftach Orr and Alicia Parrish and Cigdem Patlak and William Pietri and Forough Poursabzi-Sangdeh and Eleonora Presani and Fabrizio Puletti and Paul R{\"o}ttger and Saurav Sahay and Tim Santos and Nino Scherrer and Sebag, {Alice Schoenauer} and Patrick Schramowski and Abolfazl Shahbazi and Vin Sharma and Xudong Shen and Vamsi Sistla and Leonard Tang and Davide Testuggine and Vithursan Thangarasa and Watkins, {Elizabeth Anne} and Rebecca Weiss and Chris Welty and Tyler Wilbers and Adina Williams and Carole-Jean Wu and Poonam Yadav and Xianjun Yang and Yi Zeng and Wenhui Zhang and Fedor Zhdanov and Jiacheng Zhu and Percy Liang and Peter Mattson and Joaquin Vanschoren",

year = "2024",

month = apr,

day = "18",

language = "English",

publisher = "arXiv",

type = "WorkingPaper",

institution = "arXiv",

}

TY - UNPB

T1 - Introducing v0.5 of the AI Safety Benchmark from MLCommons

AU - Vidgen, Bertie

AU - Agrawal, Adarsh

AU - Akinwande, Victor

AU - Al-Nuaimi, Namir

AU - Alfaraj, Najla

AU - Alhajjar, Elie

AU - Aroyo, Lora

AU - Bavalatti, Trupti

AU - Blili-Hamelin, Borhane

AU - Bollacker, Kurt

AU - Bomassani, Rishi

AU - Boston, Marisa Ferrara

AU - Campos, Siméon

AU - Chakra, Kal

AU - Chen, Canyu

AU - Coleman, Cody

AU - Coudert, Zacharie Delpierre

AU - Derczynski, Leon

AU - Dutta, Debojyoti

AU - Eisenberg, Ian

AU - Ezick, James

AU - Frase, Heather

AU - Fuller, Brian

AU - Gandikota, Ram

AU - Gangavarapu, Agasthya

AU - Gangavarapu, Ananya

AU - Gealy, James

AU - Ghosh, Rajat

AU - Goel, James

AU - Gohar, Usman

AU - Goswami, Sujata

AU - Hale, Scott A.

AU - Hutiri, Wiebke

AU - Imperial, Joseph Marvin

AU - Jandial, Surgan

AU - Judd, Nick

AU - Juefei-Xu, Felix

AU - Khomh, Foutse

AU - Kailkhura, Bhavya

AU - Kirk, Hannah Rose

AU - Klyman, Kevin

AU - Knotz, Chris

AU - Kuchnik, Michael

AU - Kumar, Shachi H.

AU - Lengerich, Chris

AU - Liao, Zeyi

AU - Long, Eileen Peters

AU - Lu, Victor

AU - Mai, Yifan

AU - Mammen, Priyanka Mary

AU - Manyeki, Kelvin

AU - McGregor, Sean

AU - Mehta, Virendra

AU - Mohammed, Shafee

AU - Moss, Emanuel

AU - Nachman, Lama

AU - Naganna, Dinesh Jinenhally

AU - Nikanjam, Amin

AU - Nushi, Besmira

AU - Oala, Luis

AU - Orr, Iftach

AU - Parrish, Alicia

AU - Patlak, Cigdem

AU - Pietri, William

AU - Poursabzi-Sangdeh, Forough

AU - Presani, Eleonora

AU - Puletti, Fabrizio

AU - Röttger, Paul

AU - Sahay, Saurav

AU - Santos, Tim

AU - Scherrer, Nino

AU - Sebag, Alice Schoenauer

AU - Schramowski, Patrick

AU - Shahbazi, Abolfazl

AU - Sharma, Vin

AU - Shen, Xudong

AU - Sistla, Vamsi

AU - Tang, Leonard

AU - Testuggine, Davide

AU - Thangarasa, Vithursan

AU - Watkins, Elizabeth Anne

AU - Weiss, Rebecca

AU - Welty, Chris

AU - Wilbers, Tyler

AU - Williams, Adina

AU - Wu, Carole-Jean

AU - Yadav, Poonam

AU - Yang, Xianjun

AU - Zeng, Yi

AU - Zhang, Wenhui

AU - Zhdanov, Fedor

AU - Zhu, Jiacheng

AU - Liang, Percy

AU - Mattson, Peter

AU - Vanschoren, Joaquin

PY - 2024/4/18

Y1 - 2024/4/18

N2 - This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.

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

KW - cs.CL

KW - cs.AI

M3 - Preprint

BT - Introducing v0.5 of the AI Safety Benchmark from MLCommons

PB - arXiv

ER -

Introducing v0.5 of the AI Safety Benchmark from MLCommons

Abstract

Keywords

Access to Document

Fingerprint

Cite this