How are the "tessdata_best" models created?

[Julian Gilbey]

I've been reading the wiki, and it says in the explanation of the training process (in the "Using tesstrain.sh" section):

"For making a general-purpose LSTM-based OCR engine, it is woefully inadequate, but makes a good tutorial demo."

So my two questions are: (1) in what ways is this "woefully inadequate", and (2) how are the tessdata_best models made?  Is it just that they are trained for many more iterations in the same way, perhaps using more fonts?  Or is there more to it than that?

Thanks in advance!