I think you overtrained it, But instead you have to train 200-300 epochs if it was very great.

loss/g/total is a better indicator

You'll find that among the models I posted, those with longer datasets have less epochs because there is plenty of data without needing to go over the entire dataset many times

It's recommended to take the lowest (smoothed) value which I did. Also too low epochs makes the model sing nonsense during silences

haha funny goth girl go brrrrrrr