TL;DR. Best reward comes WITHOUT the LoRA, and in fact without even the steering prompt — plain BASE (0.489) beats the evolved prompt (BASE_P 0.437) and every merge, which decline monotonically (LoRA50 0.403 to LoRA100 0.319). Pride is a failure case in this rig: both prompt-steering and the LoRA reduce reward, and emo never rises (max 0.018 at LoRA50), so keep a plain neutral prompt and leave both LoRA and steering prompt off. The LoRA's only real effect is to erode vocal-burst blend (0.453 to 0.263) for zero emotional payoff. When forced, its coloring is counterproductive: outputs correlate with Triumph, Contempt and Malevolence, collapsing dignified pride into an arrogant, contemptuous tone.
Without a LoRA
Best: neutral prompt, no LoRA → reward 0.49
GENERAL: A voice overwhelmingly expressing pride, a proud, self-assured voice, chest swelling with pride, triumphantly boastful, impossible to hide.
SCRIPT: (with swelling, self-assured pride) "<neutral sentence>" · temp 1.05 top_p 0.95 top_k 25
Best reward comes WITHOUT the LoRA, and in fact without even the steering prompt — plain BASE (0.489) beats the evolved prompt (BASE_P 0.437) and every merge, which decline monotonically (LoRA50 0.403 to LoRA100 0.319). Pride is a failure case in this rig: both prompt-steering and the LoRA reduce reward, and emo never rises (max 0.018 at LoRA50), so keep a plain neutral prompt and leave both LoRA and steering prompt off. The LoRA's only real effect is to erode vocal-burst blend (0.453 to 0.263) for zero emotional payoff. When forced, its coloring is counterproductive: outputs correlate with Triumph, Contempt and Malevolence, collapsing dignified pride into an arrogant, contemptuous tone.