LemonTea03/BinaryOPD-Data

Do We Really Need KL Divergence for On-Policy Distillation of Large Language Models? This dataset supports the paper Do We Really Need KL Divergence for On-Policy Distillation of Large Language Models?. It contains training and evaluation data for on-policy distillation (OPD) experiments, including BinaryOPD, positive/negative reward experiments, and consensus multi-teacher on-policy distillation (C-MOPD). Code and additional details are available in the GitHub repository.… See the full description on the dataset page: https://huggingface.co/datasets/LemonTea03/BinaryOPD-Data.