TL;DR. Best without the LoRA — the evolved prompt (BASE_P) leads at reward 0.553 with an excellent WER of 0.077, while every merge is worse (LoRA50 0.483, down to 0.371 at 150%). Prompt-steer only: shame never registers as a distinct emotion here (target emo stays ~0.009 in all conditions), so the LoRA buys no emotion but costs blend and intelligibility (WER 0.077 to 0.373 at 50%). Practically, shame surfaces as sadness — BASE_P correlates with Pain (0.999), Sadness (0.978) and Disappointment (0.95) — so write it as quiet, downcast lines.
Without a LoRA
Best: evolved prompt, no LoRA → reward 0.55
GENERAL: A voice deeply expressing shame, a small, shame-filled voice, cringing and self-reproaching, unable to look up, building and building.
SCRIPT: (with cringing shame and self-reproach) "<neutral sentence>" · temp 0.9 top_p 0.95 top_k 25
Best without the LoRA — the evolved prompt (BASE_P) leads at reward 0.553 with an excellent WER of 0.077, while every merge is worse (LoRA50 0.483, down to 0.371 at 150%). Prompt-steer only: shame never registers as a distinct emotion here (target emo stays ~0.009 in all conditions), so the LoRA buys no emotion but costs blend and intelligibility (WER 0.077 to 0.373 at 50%). Practically, shame surfaces as sadness — BASE_P correlates with Pain (0.999), Sadness (0.978) and Disappointment (0.95) — so write it as quiet, downcast lines.