在训练 PPO 时遇到的问题
作者: wuyanjay创建于 2026年7月14日更新于 2026年7月14日
Hi, while studying the source code of PPO, I found that the calculation of the repetition penalty coefficient for Chinese is not normal when manually calling the rep_penalty function in train_ppo.py. I wonder if my understanding is wrong. Here is my calling process:
内容来源: jingyaogong/minimind