Profile-reward recipe: generate 640 (batched) → ASR-filter+trim → SIDON → keep the top 20% by a per-character quality+profile reward → warm-start FT 3 epochs. Below: the quality range of the selected training set, then example generations after each epoch.
snarling-beastman
Characteristic dimensions (auto, from its own voice profile): ▲ S_CASU ▲ S_STRY ▲ S_DRAM ▲ VFLX ▼ EXPL ▼ S_TECH ▼ S_ASMR ▼ S_NEWS · emo: Concentration Hope_Enthusiasm_Optimism Interest
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 18/93.
Best 5 of the kept top-20%
R 3.94 · genu 0.41 · blend 0.77 · qual 0.42 · WER 0.86
R 3.50 · genu 0.37 · blend 0.21 · qual 0.63 · WER 0.07
R 3.46 · genu 0.17 · blend 0.95 · qual 0.60 · WER 0.74
R 3.21 · genu 0.42 · blend 0.35 · qual 0.38 · WER 0.07
R 3.04 · genu 0.29 · blend 0.60 · qual 0.63 · WER 0.93
Weakest 5 still inside the top-20% (the selection boundary)
R 1.89 · genu 0.24 · blend 0.50 · qual 0.47 · WER 0.00
R 1.88 · genu 0.13 · blend 0.68 · qual 0.60 · WER 0.24
R 1.84 · genu 0.14 · blend 0.01 · qual 0.67 · WER 0.00
R 1.71 · genu 0.16 · blend 0.36 · qual 0.57 · WER 0.07
R 1.58 · genu 0.12 · blend 0.24 · qual 0.61 · WER 0.03
Fantasy line — refined ep1 / ep2 / ep3
“The old road winds past the ruined tower where the last dragon knights once kept their watch.”
Sci-fi line — refined ep1 / ep2 / ep3
“The station drifted in low orbit while the reactor cooled and the scanners swept the empty dark.”
Neutral line — refined ep1 / ep2 / ep3
“The meeting is scheduled for three o'clock and the report is on the desk by the printer.”