LabelSmoothingLoss for finetuning

[tensorjackal]

Hey! It seems that CosyVoice2 uses lsm_weight: 0 i.e., it uses cross-entropy for training on the pre-trained corpora.

However, for fine-tuning, is it preferred to use LabelSmoothingLoss, i.e. setting lsm_weight to something like 0.05 or 0 .1?