How is the performance if we train with sCM loss only?

[meng-ustc]

Hi @zhengkw18,

In the paper, you compared the rCM with DMD loss, I am wondering what if you train the model with sCM loss only? Do I need to turn the hyperparameters a lot if I want to use sCM loss only?

Btw, would you mind sharing your loss curves/grad norms over training time?

Thanks in advance!
Meng

[zhengkw18]

Training with sCM only will result in high generation diversity, but the quality is low. There are sCM-only results in the paper and the rCM repo.

Loss curves and grad norms are not important in sCM or DMD distillation, unless you encounter bad results or potential bugs.