avoid model generate *action*
i use model for conversation message chat but the model keep generating action such as laughs, blushes even i have prompted the model not to do so, i have tries use logit bias but seem like can not handle all the cases because * stick to many action to form a new token during finetuning, what i can do
Hello, this model was trained on RP dataset, in RP we use ** a lot to indicate action.
Since this is the goal of the model, I can't really tell you how to avoid it, maybe try a model that isn't made for that hahaha
I didn't tried to block that before release as I don't see the goal for the usage needed
Hi @Undi95 thank you, i am looking forward to some lewd chat models from you haha
@duongnghia222 : I haven't tested with this model, but Llama3 generally is pretty good at in-context learning. Just edit all output that you don't want (laughs) into what you do want (He laughs excitedly.) and after only a few corrections it will probably follow the desired writing style.
In general this is a much more effective method than negative commands if you want an LLM not to do something.
@FlareRebellion i agree, i use the model to generate lewd chat conversation so an instruct model would be enough, maybe i will try unholy llama3