[ot][spam][crazy] coping spam while trying to make code to prompt a model

[Undescribed Horrific Abuse, One Victim & Survivor of Many]

i updated the separator token back to using the padding token which
seems to work better
i blame the poorness of the data augmentation foremost on not tuning
the parameters of sampling (that it is sampling too low probabilities
for the small number of examples it has)

it would be interesting and useful to add user interaction to the data
generation.

however i think what would make sense to add next would be the
function to finetune the model for the stored examples. likely as an
adapter, so a checkpoint can be stored as a small blob.

likely we will find there are far too few examples for finetuning to
be helpful, but also likely it will be a very tiny bit helpful if not
done too much.

time for supper