TL;DR. Best reward is essentially a tie between no LoRA (neutral BASE, 0.504) and the LoRA at 150% (0.502); notably the evolved prompt (BASE_P, 0.459) actually hurts here, so prefer the plain base prompt for safe quality. If you need audibly perceptible amusement, only the LoRA delivers it — target-emotion strength climbs from ~0.01 (both no-LoRA conditions) to 0.259 at 150%, driving huge Teasing (+1.56) and Pleasure/Elation shifts. The trade-off is severe: vocal-burst blend collapses (0.453 to 0.175), speech quality drops (0.625 to 0.405) and WER rises (0.269 to 0.472). Its signature side-effect is that it strongly suppresses Concentration (-0.78) and Warmth, trading composure for giddiness.
Without a LoRA
Best: neutral prompt, no LoRA → reward 0.50
GENERAL: A voice overwhelmingly expressing amusement, a voice bubbling with amusement, on the edge of laughter, playful and delighted, impossible to hide.
SCRIPT: (chuckling, highly amused, barely holding back laughter) "<neutral sentence>" · temp 1.1 top_p 0.95 top_k 30
Best reward is essentially a tie between no LoRA (neutral BASE, 0.504) and the LoRA at 150% (0.502); notably the evolved prompt (BASE_P, 0.459) actually hurts here, so prefer the plain base prompt for safe quality. If you need audibly perceptible amusement, only the LoRA delivers it — target-emotion strength climbs from ~0.01 (both no-LoRA conditions) to 0.259 at 150%, driving huge Teasing (+1.56) and Pleasure/Elation shifts. The trade-off is severe: vocal-burst blend collapses (0.453 to 0.175), speech quality drops (0.625 to 0.405) and WER rises (0.269 to 0.472). Its signature side-effect is that it strongly suppresses Concentration (-0.78) and Warmth, trading composure for giddiness.