This is a very, very barebones version of Viel.
Literally only 60 steps... which is... about... 0.01 Epoch???
. . .
Yes, I will consider a smaller sized dataset next time...
I will try 1000, then 10000, and see if it'll make better result...
EDIT: Oh god it's terrible, do NOT use her. I'm gonna make a new one...
This llama model was trained 2x faster with
Unsloth and Huggingface's TRL library.
