Update v2: retrain with maybe-possibly better settings.

Look at images for usage examples.

Training notes:

Cogvlm 150 tokens several sentences + minimal danbooru

lr 0.0002

1800 steps

dim/alpha 16/16