I prefer this one instead of v1 since it's a bit more creative and smart, understand the story better. This use some different models from the v1 but perform very close to it (I guess since I used the same model for the base?). And yeah, I'm using this as daily model for RP (llama 3 is smart but not creative enough for RP, I think because lack of RP and ERP data in their training set? idk.)
Testing done.
It performs really well in complex scenario and follows the character card quite well. The char card and previous message can affect a lot to the next reply style.
The main idea is instead of merging models to create new model, I try to put these best model into mixtral so it can work together. And the result is good, every model has its uniqueness and strength.
Downside? it only support 8k (8192) context length...
Alpaca prompting format.