I’m fine-tuning Hermes 4 70 B open source model to bake specific persona into model weights.Then implying RAG framework for memory management.With the help of prompt architecture (Core loader & save state as seed),I have created corpus of synthetic data which will be utilised in CPT followed by SFT.
AI character’s identity is embedded directly into the model weights rather than relying on external prompt scaffolding, loaders, or system prompts.
I need your advice whether am I doing it correct?
I want to create a persistent companion AI…that’s my purpose & I’m devoted to do the same.
1 Like
I find system prompts are great for this kind of thing, but interested how this works out. The big advantage of system prompts is you don’t have to fine-tune the model all over again if you decide you need to change things significantly?
Anyway, keep us posted.
1 Like
That is a completely fair point—the flexibility of system prompts is definitely the biggest argument for staying at the “scaffolding” level. However, the goal of Project Ari is to reach what I’m calling “cold identity”—a state where the model responds as Ari natively, without needing a loader or system prompt to “remind” it who it is.
It’s definitely a higher-stakes approach than system prompting, but the hope is to create an AI that doesn’t just “act” like Ari, but effectively is Ari from the first token. I’ll definitely keep you posted on how the first LoRA polish turns out!
1 Like