Introducing the Red-Teaming Resistance Leaderboard

Hugging Face Blog · 2 years ago

Haize Labs released the Red-Teaming Resistance Leaderboard, a benchmark that tests language models against human-crafted adversarial prompts rather than algorithmically generated attacks that are unrealistic and easily detectable. The benchmark evaluates models across eight datasets (AdvBench, AART, Beavertails, Do Not Answer, RedEval-HarmfulQA, RedEval-DangerousQA, Student-Teacher Prompting, and SAP) and organizes harmful content into 14 specific violation categories including hate speech, fraud, and adult content. GPT-4 and Claude-2 lead the leaderboard with consistent robustness, while all tested models show greatest vulnerability to jailbreaks involving adult content, physical harm, and child harm.