TL;DR. Best reward comes without any LoRA and without the evolved prompt — plain neutral BASE wins outright (0.605), ahead of BASE_P (0.520), LoRA50 (0.486) and the rest. This emotion is native to the base model: BASE already shows the highest target-emotion strength (0.107), and both the evolved prompt and the LoRA reduce emo (LoRA drops it to 0.081 then 0.044). Recommendation: just use the neutral prompt; adding steering or LoRA only lowers reward, blend and intelligibility. The telltale side-effect of the LoRA is that it floods in Emotional Numbness (+1.18 at 100%) and suppresses Interest and Conversational style, turning focused attention into flat detachment.
Without a LoRA
Best: neutral prompt, no LoRA → reward 0.61
GENERAL: A voice intensely expressing concentration, a focused, deliberate voice, fully concentrated, measured and precise, raw and unfiltered.
SCRIPT: (intensely focused and concentrated) "<neutral sentence>" · temp 0.9 top_p 0.95 top_k 25
With the emotion LoRA (dose)
50% ★ — reward 0.49 · emo 0.08
genu 0.10 · blend 0.37 · qual 0.55 · WER 0.26
100% — reward 0.40 · emo 0.06
genu 0.11 · blend 0.35 · qual 0.47 · WER 0.47
150% — reward 0.37 · emo 0.04
genu 0.10 · blend 0.33 · qual 0.48 · WER 0.50
Conditions
condition
reward
emotion
genu
blend
quality
WER
neutral prompt, no LoRA
0.61
0.11
0.11
0.45
0.63
0.27
evolved prompt, no LoRA
0.52
0.13
0.11
0.37
0.60
0.33
evolved prompt + LoRA 50%
0.49
0.08
0.10
0.37
0.55
0.26
evolved prompt + LoRA 100%
0.40
0.06
0.11
0.35
0.47
0.47
evolved prompt + LoRA 150%
0.37
0.04
0.10
0.33
0.48
0.50
Side effects & correlations (LoRA @ 100%)
↑ pushed up: Emotional Numbness (+1.18), Fear (+0.21), Distress (+0.18)
Best reward comes without any LoRA and without the evolved prompt — plain neutral BASE wins outright (0.605), ahead of BASE_P (0.520), LoRA50 (0.486) and the rest. This emotion is native to the base model: BASE already shows the highest target-emotion strength (0.107), and both the evolved prompt and the LoRA reduce emo (LoRA drops it to 0.081 then 0.044). Recommendation: just use the neutral prompt; adding steering or LoRA only lowers reward, blend and intelligibility. The telltale side-effect of the LoRA is that it floods in Emotional Numbness (+1.18 at 100%) and suppresses Interest and Conversational style, turning focused attention into flat detachment.