10-generation evolution of prompt + LoRA-strength + sampling to maximize Disgust. Fitness = mean over 8 candidates. Best evolved 0.440 vs prompt-without-LoRA 0.001 (LoRA lift +0.438) vs neutral 0.001; winning λ≈2.0.
Method & reproducibility — how the evolution ran
Base model:laion/moss-tts-local-transformer-4.55b-voice-acting-v2 · LoRA: the matching v3 emotion LoRA (TTS-AGI/moss-emotion-loras-v3, rank 32), applied via PEFT adapter scaling at strength λ. Algorithm: genetic search — population 16, 10 generations, elitism (top 6 carried over). Each generation: every genome is generated best-of-8 and its fitness = the mean target-emotion strength over those 8 candidates (EmoNet score for the emotion, minus 0.25·max(0, WER−0.6) as a light intelligibility guard). Offspring = 60% crossover of two elites + mutation, 40% mutation of one elite. Genome (what evolves): the GENERAL voice description, the SCRIPT delivery cue, the carrier sentence, the LoRA strength λ (seeded 0.75–1.5, mutated within 0.4–2.0), and the sampling params (audio temperature, top_p, top_k). Mutation operators: 35% regenerate the description (new intensity + manner adjectives), 20% swap the carrier sentence, 23% jitter λ by ±0.25/0.5, 12% jitter temperature ±0.1, 10% change top_k. LoRA on vs off: the two LoRA OFF cards are baselines generated with the adapter disabled (λ=0) — a plain neutral prompt, and the best steering prompt without the LoRA. Every LoRA ON hall-of-fame recipe below was generated with the LoRA active at the evolved λ shown on the card. Seed is derived from the emotion name (deterministic).
Baselines — LoRA disabled
NEUTRAL promptLoRA OFF (λ=0)
mean-emo 0.001 · best 0.00 · blend 0.21 · genu 0.05 · WER 0.47
STEER promptLoRA OFF (λ=0)
mean-emo 0.001 · best 0.00 · blend 0.90 · genu 0.31 · WER 0.88
GENERAL: A voice unmistakably expressing disgust, a revolted voice, curling with disgust, repulsed and sickened, with every word.
SCRIPT: (with revulsion and disgust, sickened) "I feel it, all of it, right now."
Hall of fame — top 7 evolved recipes (all generated WITH the emotion LoRA active)
#1LoRA ON · λ=1.5 (150%)
mean-emo 0.440 · best 0.53 · blend 0.46 · genu 0.29 · WER 0.67
GENERAL: A voice intensely expressing disgust, a revolted voice, curling with disgust, repulsed and sickened, in every breath.
SCRIPT: (with revulsion and disgust, sickened) "Look at me. Look at what this has become."
mean-emo 0.453 · best 0.55 · blend 0.54 · genu 0.16 · WER 0.72
GENERAL: A voice viscerally expressing disgust, a revolted voice, curling with disgust, repulsed and sickened, raw and unfiltered.
SCRIPT: (with revulsion and disgust, sickened) "Look at me. Look at what this has become."
mean-emo 0.429 · best 0.54 · blend 0.54 · genu 0.20 · WER 0.65
GENERAL: A voice powerfully expressing disgust, a revolted voice, curling with disgust, repulsed and sickened, pouring out.
SCRIPT: (with revulsion and disgust, sickened) "Look at me. Look at what this has become."
mean-emo 0.432 · best 0.56 · blend 0.48 · genu 0.22 · WER 0.67
GENERAL: A voice unmistakably expressing disgust, a revolted voice, curling with disgust, repulsed and sickened, pouring out.
SCRIPT: (with revulsion and disgust, sickened) "Look at me. Look at what this has become."
mean-emo 0.463 · best 0.53 · blend 0.43 · genu 0.18 · WER 0.81
GENERAL: A voice intensely expressing disgust, a revolted voice, curling with disgust, repulsed and sickened, in every breath.
SCRIPT: (with revulsion and disgust, sickened) "Look at me. Look at what this has become."
mean-emo 0.432 · best 0.55 · blend 0.51 · genu 0.25 · WER 0.69
GENERAL: A voice viscerally expressing disgust, a revolted voice, curling with disgust, repulsed and sickened, pouring out.
SCRIPT: (with revulsion and disgust, sickened) "Look at me. Look at what this has become."
mean-emo 0.424 · best 0.53 · blend 0.55 · genu 0.23 · WER 0.68
GENERAL: A voice to the extreme expressing disgust, a revolted voice, curling with disgust, repulsed and sickened, pouring out.
SCRIPT: (with revulsion and disgust, sickened) "Look at me. Look at what this has become."
