List your company

Preference Modeling on HelpSteer2 held-out (test)

68.4Preference AccuracyMean-Var
56.9659.9362.965.87Oct 18, 2024

Evaluation results

MethodLinks
Mean-VarReward Model Type=Distributional, Loss=KL2024.1068.40.582
Bradley-TerryReward Model Type=Single-Value, Data Scaling=Agg2024.1068.30.482
Bradley-TerryReward Model Type=Single-Value, Data Scaling=All2024.1067.80.489
MSE RegressionReward Model Type=Single-Value, Data Scaling=All2024.1067.50.481
MSE RegressionReward Model Type=Single-Value, Data Scaling=Agg2024.1066.90.488
ClassificationReward Model Type=Distributional, Loss=KL2024.1065.90.648
Mean-VarReward Model Type=Distributional, Loss=NLL, Independence=Independent2024.1057.40.573