Nice writing but !
#3
by Bumoch - opened
The model keeps writing for User constantly ... no mater how often i repeated the model instruction or deleted everything it wrote for me, it just keeps doing it.
So if there is a versio V2, maybe consider to add some safeguards in the lora or Qlora about keeping its fingers away from the user and write only from char POV
but overall, nice one !!
I dont have a way to finetune on 24B yet, saving up for the hardware to do this. Limited to 8B right now. But the method i use (SFT+Qlora) might be able to train it on that with some exmaples so ill keep that in mind
Update: Now with v2 released you can see if it fixes this issue or not. I dont know how to test for it properly or I would check
Bumoch changed discussion status to closed