vn_VALN_

Valence — high

Evolving prompt+sampling to push VoiceNet dimension VALN in the high direction with the vn_VALN__high LoRA. Steering phrase: bright, positive, cheerful affect.

Best in-direction score 0.676 at 100% vs neutral 0.256 (shift +0.420).

Method & reproducibility — how the evolution ran

Base model: laion/moss-tts-local-transformer-4.55b-voice-acting-v2 · LoRA: the new VoiceNet-dimension LoRA vn_<CODE>__<high|low> (retrained with ≤25% game-character data, rank 32), applied via PEFT adapter scaling at strength λ.
Algorithm: genetic search — population 12, 10 generations, elitism (top 5). During evolution every genome is generated best-of-6 with the LoRA at λ=100%; its fitness = the mean over those 6 candidates of the target VoiceNet dimension score in the intended direction (for a __high LoRA the raw score; for __low, 1−score), minus 0.2·max(0, WER−0.6).
Genome (what evolves): the GENERAL voice description, the SCRIPT cue, the carrier sentence, and the sampling params (temperature, top_p, top_k). λ is fixed at 100% during the search.
Mutation operators: 40% regenerate description, 25% swap carrier, 17% jitter temperature ±0.1, 18% change top_k. Offspring = 60% crossover + mutation, 40% mutation.
Dose measurement (below): the winning recipe is then re-generated at λ=0 (LoRA OFF), λ=50% and λ=100% (each best-of-6) to show exactly what 50% vs 100% of the LoRA brings over the no-LoRA baseline.

Dose — what 50% vs 100% brings (top recipe)

λ=0 (no LoRA) — dir 0.263 (raw 0.26)

WER 0.17 · genu 0.10

λ=50% — dir 0.449 (raw 0.45)

WER 0.00 · genu 0.16

λ=100% — dir 0.676 (raw 0.68)

WER 0.23 · genu 0.27

GENERAL: A voice that is markedly bright, positive, cheerful affect. SCRIPT: (clearly bright, positive, cheerful affect) "Please remember to lock the door before you head out." · temp 1.2 top_p 0.95 top_k 30

Other evolved recipes (100%)

— dir 0.668 (raw 0.67)