TL;DR. Best reward comes WITH the LoRA at 50% (0.582 vs 0.498 for the plain BASE prompt, which is the best no-LoRA option here). Use the evolved steering prompt plus the 50% merge: the target-emotion signal stays weak (emo 0.026 — embarrassment resists direct elicitation), so the gain is actually driven by genuineness rising (0.110 to 0.161) and WER dropping sharply (0.269 to 0.112) while blend holds at 0.45. The trade-off is the speech-quality proxy, which falls from 0.625 to 0.475. Watch that the LoRA drags delivery toward a tired/sad register (shift up Fatigue +0.62, corr with Sadness 0.73 and Helplessness 0.71); don't exceed 50%, as LoRA100/150 both lose reward.
Without a LoRA
Best: neutral prompt, no LoRA → reward 0.50
GENERAL: A voice unmistakably expressing embarrassment, a flustered, sheepish voice, cheeks burning with embarrassment, awkward and shrinking, impossible to hide.
SCRIPT: (flustered and mortified with embarrassment) "<neutral sentence>" · temp 1.05 top_p 0.9 top_k 30
With the emotion LoRA (dose)
50% ★ — reward 0.58 · emo 0.03
genu 0.16 · blend 0.45 · qual 0.47 · WER 0.11
100% — reward 0.44 · emo 0.03
genu 0.11 · blend 0.34 · qual 0.50 · WER 0.17
150% — reward 0.46 · emo 0.06
genu 0.13 · blend 0.34 · qual 0.41 · WER 0.23
Conditions
condition
reward
emotion
genu
blend
quality
WER
neutral prompt, no LoRA
0.50
0.01
0.11
0.45
0.63
0.27
evolved prompt, no LoRA
0.49
0.02
0.12
0.42
0.56
0.25
evolved prompt + LoRA 50%
0.58
0.03
0.16
0.45
0.47
0.11
evolved prompt + LoRA 100%
0.44
0.03
0.11
0.34
0.50
0.17
evolved prompt + LoRA 150%
0.46
0.06
0.13
0.34
0.41
0.23
Side effects & correlations (LoRA @ 100%)
↑ pushed up: Impatience and Irritability (+0.46), Fear (+0.38), Fatigue Exhaustion (+0.27)
Best reward comes WITH the LoRA at 50% (0.582 vs 0.498 for the plain BASE prompt, which is the best no-LoRA option here). Use the evolved steering prompt plus the 50% merge: the target-emotion signal stays weak (emo 0.026 — embarrassment resists direct elicitation), so the gain is actually driven by genuineness rising (0.110 to 0.161) and WER dropping sharply (0.269 to 0.112) while blend holds at 0.45. The trade-off is the speech-quality proxy, which falls from 0.625 to 0.475. Watch that the LoRA drags delivery toward a tired/sad register (shift up Fatigue +0.62, corr with Sadness 0.73 and Helplessness 0.71); don't exceed 50%, as LoRA100/150 both lose reward.