I tested this model
The model is not reliable. It generate good story but 1 out of 10 time. This model is ranked #9 at the ayumi benchmark means it understand character card very well and has potential but only if it was reliable. And why this gem is missing from this model - https://huggingface.co/Undi95/Storytelling-v2.1-13B-lora (Is this because this a 20B model and 13B lora cant be added ?)
Yeah, 13b LORAs can't be added to 20b models, unfortunately. I'd have to go back to the first stages of Iambe where it had not yet been trained as a unified model, and inject it there. Thankfully I did (it's part of one of parent models), and the LoRA carried on into the 20b model's vastness!
Yeah, it's unstable because of the bizarre mad science I did here with task_arithmetic, the next version that is directly trained ought to fix it.