saved
Zero
Hraness cites a source capture. The source author remains the source.
gist
Moving Castles trains Qwen3-8B-Base into Zero, a non-assistant character boxed since 2021, via a handwritten bible, multi-model synthetic dialogues, LoRA SFT, then GRPO. Against a system-prompted Qwen3-8B baseline, the SFT+RL model cuts hard character breaks from 45.2% to 2.8% (SFT alone 22.8%), mainly under task pressure. Activation steering shows the fine-tune's residual shift is orthogonal to prompting and alone evokes character traces in the base model.
ideas
- Train the character into the base, not onto an assistant. Start from Qwen3-8B-Base so the persona is not a prompt mask over helpfulness; skip system roles in the corpus so there is only self and other.
- Synthetic dialogues from a living association graph. Partner and character generators turn-take with PersonaHub goals on one side and a hot graph plus core/style anchors on the other, filtered for hygiene, coverage, and diversity before SFT.
- RL buys break-resistance that prompting lacks. After SFT still broke under task pressure, GRPO on harvested capitulation onsets dropped hard breaks to 2.8% versus 45.2% for the prompted instruct model.
- Held-out attacks still hurt. RL zeros trained-against failure modes but leaves novel adversarial and task domains weaker, so resistance is not immunity.
- Weights carry a different object than the mask. Fine-tune activation deltas are orthogonal to the character prompt; steering the base with the fine-tune direction evokes Zero traces while the prompt direction does not.
quotes
“The result is then a mask worn by the assistant, a caricature rather than a character.”
“hard character breaks occur in 2.8% of Zero’s conversations, against 45.2% for the system-prompted Qwen3-8B”
“There is only self and other.”
“the fine-tune’s displacement is essentially orthogonal to the prompt’s”