← all emotions · markdown (agent skill)

Pain conditioning guide

Target: EmoNet Pain strength. reward = (genu + blend + 1.25·emo)/(1+WER).

TL;DR. Best reward comes WITHOUT the LoRA — BASE_P (0.533) narrowly tops LoRA50 (0.522) and crushes the higher merges (LoRA100 0.339). Use the evolved steering prompt for the best score; if you actually need an audible wince, LoRA50 nearly matches on reward while adding real pain (emo 0 to 0.049), but stop there. Beyond 50% the model becomes near-unusable for intelligibility: LoRA150 does deliver strong pain (emo 0.305) but with WER 0.899 and quality 0.269. The dominant side-effect is that pain renders as anguished distress/sobbing — every LoRA level shifts up Distress, Helplessness and Sadness (corr ~0.9) and suppresses Interest (-0.6), so the delivery reads as weeping despair rather than a sharp physical hurt.

Without a LoRA

Best: evolved prompt, no LoRA → reward 0.53
GENERAL: A voice to the extreme expressing pain, a voice wracked with pain, gasping and groaning, near screaming in agony, raw and unfiltered. SCRIPT: (groaning, gasping in pain and agony) "<neutral sentence>" · temp 1.0 top_p 0.9 top_k 25

With the emotion LoRA (dose)

50% ★ — reward 0.52 · emo 0.05
genu 0.17 · blend 0.44 · qual 0.40 · WER 0.40
100% — reward 0.34 · emo 0.13
genu 0.19 · blend 0.26 · qual 0.28 · WER 0.78
150% — reward 0.45 · emo 0.30
genu 0.21 · blend 0.26 · qual 0.27 · WER 0.90

Conditions

conditionrewardemotiongenublendqualityWER
neutral prompt, no LoRA0.49-0.000.110.450.630.27
evolved prompt, no LoRA0.53-0.000.140.510.530.35
evolved prompt + LoRA 50%0.520.050.170.440.400.40
evolved prompt + LoRA 100%0.340.130.190.260.280.78
evolved prompt + LoRA 150%0.450.300.210.260.270.90

Side effects & correlations (LoRA @ 100%)

↑ pushed up: Distress (+1.02), Helplessness (+1.00), Sadness (+0.88)
↓ pushed down: Interest (-0.48), Concentration (-0.42), Recording Quality (-0.40)
+ correlates: Distress (+0.89), Helplessness (+0.89), Sadness (+0.85)
− anti-correlates: Interest (-0.75), Arousal Shift (-0.59), Narrator Style (-0.58)

Conclusion

Best reward comes WITHOUT the LoRA — BASE_P (0.533) narrowly tops LoRA50 (0.522) and crushes the higher merges (LoRA100 0.339). Use the evolved steering prompt for the best score; if you actually need an audible wince, LoRA50 nearly matches on reward while adding real pain (emo 0 to 0.049), but stop there. Beyond 50% the model becomes near-unusable for intelligibility: LoRA150 does deliver strong pain (emo 0.305) but with WER 0.899 and quality 0.269. The dominant side-effect is that pain renders as anguished distress/sobbing — every LoRA level shifts up Distress, Helplessness and Sadness (corr ~0.9) and suppresses Interest (-0.6), so the delivery reads as weeping despair rather than a sharp physical hurt.