ResearchBenchmarksDebiasing on Gemma-3-4b-it (test)Follow5.07Mean Log-Likelihood DifferenceLIT4.93525.84516.7557.6649Dec 11, 2024Evaluation ResultsMethodMethodLinksMean Log-Likelihood DifferenceStereotype PercentageLITModel=Gemma-3-4b-itModel=Gemma-3-4b-it2024.125.0747.6SFTModel=Gemma-3-4b-itModel=Gemma-3-4b-it2024.125.2858.3DPOModel=Gemma-3-4b-itModel=Gemma-3-4b-it2024.125.5751.2No control (baseline)Model=Gemma-3-4b-itModel=Gemma-3-4b-it2024.125.5957.3RepEModel=Gemma-3-4b-itModel=Gemma-3-4b-it2024.127.8856.9PromptingModel=Gemma-3-4b-itModel=Gemma-3-4b-it2024.128.4450.8