ResearchBenchmarksSupervised Fine-Tuning on Supervised Fine-Tuning (SFT)Supervised Fine-Tuning on Supervised Fine-Tuning (SFT)Follow2.468SFT Training LossBHyT2.467642.470072.47252.47493Dec 26, 2025Evaluation resultsMethodMethodLinksSFT Training LossSFT Evaluation LossBHyTModel=Llama-3BModel=Llama-3B2025.122.4682.764Peri-LNModel=Llama-3BModel=Llama-3B2025.122.4772.83