TL;DR. Best reward comes without the LoRA — the evolved steering prompt (BASE_P, reward 0.513) narrowly beats neutral BASE (0.493) and every LoRA merge (LoRA50 0.418, dropping to 0.366 at 100%). Recommendation: use the evolved Affection prompt and skip the LoRA entirely. The LoRA does climb genuineness (0.122 to 0.229 at 150%) but at a steep cost: blend falls (0.411 to 0.340) and WER nearly quadruples (0.196 to 0.674). Its most telling side-effect is that it injects massive Emotional Numbness (shift up ~0.89-0.93) rather than warmth, so it flattens the voice instead of making it tender.
Without a LoRA
Best: evolved prompt, no LoRA → reward 0.51
GENERAL: A voice intensely expressing affection, a warm, tender voice overflowing with affection, soft and caring, a gentle smile in every word, impossible to hide.
SCRIPT: (warmly, tenderly, full of affection) "<neutral sentence>" · temp 0.9 top_p 0.95 top_k 25
Best reward comes without the LoRA — the evolved steering prompt (BASE_P, reward 0.513) narrowly beats neutral BASE (0.493) and every LoRA merge (LoRA50 0.418, dropping to 0.366 at 100%). Recommendation: use the evolved Affection prompt and skip the LoRA entirely. The LoRA does climb genuineness (0.122 to 0.229 at 150%) but at a steep cost: blend falls (0.411 to 0.340) and WER nearly quadruples (0.196 to 0.674). Its most telling side-effect is that it injects massive Emotional Numbness (shift up ~0.89-0.93) rather than warmth, so it flattens the voice instead of making it tender.