TL;DR. Best reward is without any LoRA, use the evolved steering prompt (BASE_P 0.520, just ahead of BASE 0.492); every LoRA merge lowers reward (0.458 to 0.357 to 0.302). The LoRA nudges target emotion up only marginally (0.006 to 0.058) while WER explodes to 0.79 and blend/quality sink. Its worst side-effect is category drift: LoRA50 correlates almost perfectly with Disgust and Sourness (~1.0) and by 150% shifts hard into Anger, Malevolence and Impatience, so it stops sounding like contempt and becomes generic hostility. Note that actual contempt strength stays near zero (0.001) even with the prompt — this emotion is intrinsically hard to elicit, so lean on the prompt's low WER (0.134) and clean genuineness rather than chasing the emo score.
Without a LoRA
Best: evolved prompt, no LoRA → reward 0.52
GENERAL: A voice to the extreme expressing contempt, a cold, sneering voice, full of contempt and disdain, looking down with scorn, pouring out.
SCRIPT: (with cold, sneering contempt) "<neutral sentence>" · temp 1.1 top_p 0.9 top_k 30
Best reward is without any LoRA, use the evolved steering prompt (BASE_P 0.520, just ahead of BASE 0.492); every LoRA merge lowers reward (0.458 to 0.357 to 0.302). The LoRA nudges target emotion up only marginally (0.006 to 0.058) while WER explodes to 0.79 and blend/quality sink. Its worst side-effect is category drift: LoRA50 correlates almost perfectly with Disgust and Sourness (~1.0) and by 150% shifts hard into Anger, Malevolence and Impatience, so it stops sounding like contempt and becomes generic hostility. Note that actual contempt strength stays near zero (0.001) even with the prompt — this emotion is intrinsically hard to elicit, so lean on the prompt's low WER (0.134) and clean genuineness rather than chasing the emo score.