TL;DR. Best without the LoRA — the evolved steering prompt (BASE_P) tops the table at reward 0.624, ahead of the best merge (LoRA100, 0.517). Prompt-steer and stop there: the prompt alone lifts genuineness from 0.110 to 0.174 and blend to 0.499, whereas any merge trades a negligible emo gain (0.049 vs 0.039) for a large blend drop (0.499 to 0.369) and lower reward. Note the side-effect: relief reads as a weary exhale here, correlating with Fatigue/Exhaustion (0.41) and Vulnerability (0.37) rather than bright relief, so pair it with tired-sounding text.
Without a LoRA
Best: evolved prompt, no LoRA → reward 0.62
GENERAL: A voice deeply expressing relief, a voice flooding with relief, exhaling the tension, grateful it is finally over, in every breath.
SCRIPT: (exhaling in profound relief) "<neutral sentence>" · temp 1.2 top_p 0.95 top_k 30
Best without the LoRA — the evolved steering prompt (BASE_P) tops the table at reward 0.624, ahead of the best merge (LoRA100, 0.517). Prompt-steer and stop there: the prompt alone lifts genuineness from 0.110 to 0.174 and blend to 0.499, whereas any merge trades a negligible emo gain (0.049 vs 0.039) for a large blend drop (0.499 to 0.369) and lower reward. Note the side-effect: relief reads as a weary exhale here, correlating with Fatigue/Exhaustion (0.41) and Vulnerability (0.37) rather than bright relief, so pair it with tired-sounding text.