TL;DR. Best without the LoRA (BASE_P reward 0.500, with LoRA50 a near-tie at 0.489), so prompt-steer and add at most a 50% merge. Do not go to 100% or 150%: WER explodes past 1.0 (LoRA100 WER 1.25, LoRA150 1.02), i.e. output becomes largely unintelligible, while blend collapses to 0.180. Those high merges do raise emo (to 0.186), but by turning triumph into arrogant Pride (shift +1.34) and Malevolence (corr ~0.82) — gloating rather than victorious. Keep it prompt-driven; triumph is fragile under merge.
Without a LoRA
Best: evolved prompt, no LoRA → reward 0.50
GENERAL: A voice unmistakably expressing triumph, a triumphant, victorious voice, roaring with the thrill of winning, exultant, in every breath.
SCRIPT: (triumphantly, roaring with victory) "<neutral sentence>" · temp 0.9 top_p 0.95 top_k 30
With the emotion LoRA (dose)
50% ★ — reward 0.49 · emo 0.01
genu 0.13 · blend 0.43 · qual 0.50 · WER 0.29
100% — reward 0.36 · emo 0.05
genu 0.14 · blend 0.30 · qual 0.45 · WER 1.25
150% — reward 0.37 · emo 0.19
genu 0.14 · blend 0.18 · qual 0.33 · WER 1.02
Conditions
condition
reward
emotion
genu
blend
quality
WER
neutral prompt, no LoRA
0.49
-0.00
0.11
0.45
0.63
0.27
evolved prompt, no LoRA
0.50
-0.00
0.13
0.42
0.55
0.23
evolved prompt + LoRA 50%
0.49
0.01
0.13
0.43
0.50
0.29
evolved prompt + LoRA 100%
0.36
0.05
0.14
0.30
0.45
1.25
evolved prompt + LoRA 150%
0.37
0.19
0.14
0.18
0.33
1.02
Side effects & correlations (LoRA @ 100%)
↑ pushed up: Malevolence Malice (+0.66), Impatience and Irritability (+0.51), Pride (+0.44)
Best without the LoRA (BASE_P reward 0.500, with LoRA50 a near-tie at 0.489), so prompt-steer and add at most a 50% merge. Do not go to 100% or 150%: WER explodes past 1.0 (LoRA100 WER 1.25, LoRA150 1.02), i.e. output becomes largely unintelligible, while blend collapses to 0.180. Those high merges do raise emo (to 0.186), but by turning triumph into arrogant Pride (shift +1.34) and Malevolence (corr ~0.82) — gloating rather than victorious. Keep it prompt-driven; triumph is fragile under merge.