TL;DR. Best reward comes WITHOUT the LoRA — plain BASE tops the table at 0.690, edging out the evolved prompt (BASE_P, 0.679) and beating every merge (LoRA50 0.618, LoRA150 0.552). Interest is essentially a native emotion here: the base model already hits emo=0.187 with a neutral prompt, so just prompt plainly and skip the LoRA entirely. Adding the LoRA barely moves target strength (emo 0.187 to 0.227 only at 150%) while dragging blend down hard (0.453 to 0.274) and inflating WER (0.269 to 0.513 at 100%). The one thing the LoRA reliably drags along is a cartoonish, rough delivery (corr Cartoonish Style 0.82, Roughness 0.73 at 100%) — so higher merges buy you a caricatured voice, not more genuine interest.
Without a LoRA
Best: neutral prompt, no LoRA → reward 0.69
GENERAL: A voice viscerally expressing interest, a keenly interested voice, leaning in, curious and engaged, with every word.
SCRIPT: (keenly interested and curious) "<neutral sentence>" · temp 1.15 top_p 0.95 top_k 40
Best reward comes WITHOUT the LoRA — plain BASE tops the table at 0.690, edging out the evolved prompt (BASE_P, 0.679) and beating every merge (LoRA50 0.618, LoRA150 0.552). Interest is essentially a native emotion here: the base model already hits emo=0.187 with a neutral prompt, so just prompt plainly and skip the LoRA entirely. Adding the LoRA barely moves target strength (emo 0.187 to 0.227 only at 150%) while dragging blend down hard (0.453 to 0.274) and inflating WER (0.269 to 0.513 at 100%). The one thing the LoRA reliably drags along is a cartoonish, rough delivery (corr Cartoonish Style 0.82, Roughness 0.73 at 100%) — so higher merges buy you a caricatured voice, not more genuine interest.