ResearchBenchmarksFine-tuning ranking evaluation on English fine-tuning corpus 60B tokensFollow0.817Observed LossLinear-H0.816640.819070.82150.82393Sep 7, 2024Evaluation ResultsMethodMethodLinksObserved LossPredicted ScoreActual RankPredicted RankLinear-HSchedule=Linear-H, ηma...Schedule=Linear-H, ηmax=2×10−3, a1=512, a3=5122024.090.8171.97211Linear-SSchedule=Linear-S, ηma...Schedule=Linear-S, ηmax=5×10−4, a1=128, a3=1282024.090.8222.05222Two-phaseSchedule=Two-phase, ηm...Schedule=Two-phase, ηmax=3×10−4, a1=128, a3=70002024.090.8262.17833