size of checkpoint not right

[nolangz]

I use default config(normal size, not large) to train the model, but it return a 300M+ checkpoint. what is the possible reason? Training Accuracy setting in accelerate config? Do you have some detail about accelerate config

[starrytong]

#6. The checkpoint contains multiple intermediate results, but only the “best_state” is useful for inference. You can extract it after the training is completed.