Request
Can you make a model that can play a role and talk as someone online? Like a discord user but without roleplay strings.. Also do you think you could train a model to obey instructions in the user prompt not system prompt? maybe there same thing i havent trained models so i do not really understand. The issue i also have is models sometimes talk alot especially models like nemo they just ramble on and on. If you could make a 8 - 14b model finetune with these attributes i would greatly appreciate it. God speed!
I am confused, you want a model that can talk like a discord user, or similar IRC chat. Why not just use L3.1 8b with system prompt that points that this is a regular internet reply style chat? I doubt you need a 70b model for that kind of stuff.
To be honest, I do understand what @Skorcht means. I've had some success with doing what @GhostGate suggested but it always kinda falls apart relatively easily. Most apps/services aren't too great at it either imo. There's one called AI Peeps which does ok but is kinda shallow. Only service I've found that does a surprisingly good job is AtheneGPT. I guess they have some sort of very uniquely finetuned big model for it cause I haven't been able to reproduce a consistent and immersive experience to that extent anywhere else.
Can you make a model that can play a role and talk as someone online? Like a discord user but without roleplay strings..
Yeah? pretty easily if you have human data. tradeoff will always be intelligence. Most human data is not 'smart'. It'll talk like a person sure, but it will be dumber than other models.
Also do you think you could train a model to obey instructions in the user prompt not system prompt? maybe there same thing i havent trained models so i do not really understand.
Instruct Following? I have that pipeline. Not obeying system prompt would be a problem though. Pretty easily achievable for bigger models, tougher on smaller ones.
The issue i also have is models sometimes talk alot especially models like nemo they just ramble on and on.
Guided generation / grammar helps with this. Also example messages help reinforce message length and all.
so, maybe. not confirmed though
Trying to run tells me that the word breaker is broken, am I the only one that has this error?