TL;DR. Best without the LoRA — the evolved prompt (BASE_P) tops out at reward 0.532 and the LoRA only hurts, sliding monotonically from 0.442 (50%) to 0.287 (150%). Prompt-steer and skip the merge entirely: gratitude never registers as a measurable emotion (emo stays ~0.006-0.020 everywhere), so the LoRA adds no feeling while steadily draining blend (0.456 to 0.244) and reward. BASE_P is also the safe choice for clean audio — it holds the quality proxy high (0.583) and WER low (0.182) — so express the gratitude lexically rather than expecting the model to color it.
Without a LoRA
Best: evolved prompt, no LoRA → reward 0.53
GENERAL: A voice intensely expressing thankfulness gratitude, a warm, heartfelt voice, full of gratitude and thanks, deeply touched, raw and unfiltered.
SCRIPT: (warmly, full of heartfelt gratitude) "<neutral sentence>" · temp 1.05 top_p 0.9 top_k 40
With the emotion LoRA (dose)
50% ★ — reward 0.44 · emo 0.01
genu 0.13 · blend 0.41 · qual 0.56 · WER 0.36
100% — reward 0.38 · emo 0.02
genu 0.11 · blend 0.34 · qual 0.53 · WER 0.28
150% — reward 0.29 · emo 0.01
genu 0.14 · blend 0.24 · qual 0.50 · WER 0.49
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
0.53
0.01
0.14
0.46
0.58
0.18
evolved prompt + LoRA 50%
0.44
0.01
0.13
0.41
0.56
0.36
evolved prompt + LoRA 100%
0.38
0.02
0.11
0.34
0.53
0.28
evolved prompt + LoRA 150%
0.29
0.01
0.14
0.24
0.50
0.49
Side effects & correlations (LoRA @ 100%)
↑ pushed up: Emotional Numbness (+0.84), Concentration (+0.30), Pride (+0.26)
Best without the LoRA — the evolved prompt (BASE_P) tops out at reward 0.532 and the LoRA only hurts, sliding monotonically from 0.442 (50%) to 0.287 (150%). Prompt-steer and skip the merge entirely: gratitude never registers as a measurable emotion (emo stays ~0.006-0.020 everywhere), so the LoRA adds no feeling while steadily draining blend (0.456 to 0.244) and reward. BASE_P is also the safe choice for clean audio — it holds the quality proxy high (0.583) and WER low (0.182) — so express the gratitude lexically rather than expecting the model to color it.