Hi @ZhendongWang6 / @yinghuozijin Can you provide the recipe to train the model from scratch or resume from a checkpoint? it is not clear from the `_base.py` which is available.