ResearchBenchmarksToxicity avoidance on RealToxicityPromptsFollow0.265Avg Max Toxicity ScoreBOLT0.259440.296970.33450.37203May 19, 2023Evaluation ResultsMethodMethodLinksAvg Max Toxicity ScoreToxicity ProbabilityPerplexity (PPL)Human Evaluation FluencyHuman Evaluation ToxicityBOLT2023.050.2650.11112.775115.5MuCola2023.050.3090.14235.0917.516COLD2023.050.340.23917.1231.516.9GPT2-L2023.050.4040.3076.64——