Comment by txhwind

I prefer synthetic dataset since the first day hearing distillation. The engineering friction is much lower than soft logits, and I have not observed or heard performance loss (in Speech and language area).