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.
scorpion
Characteristic dimensions (auto, from its own voice profile): ▲ GEND ▲ S_STRY ▲ S_MONO ▲ S_DRAM ▼ S_TECH ▼ S_ASMR ▼ S_WHIS ▼ S_RANT · emo: Concentration Interest Malevolence_Malice
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 126/634.
Best 5 of the kept top-20%
R 6.02 · genu 0.16 · blend 0.51 · qual 0.62 · WER 0.00
R 5.89 · genu 0.24 · blend 0.62 · qual 0.49 · WER 0.00
R 5.87 · genu 0.14 · blend 0.82 · qual 0.58 · WER 0.04
R 5.87 · genu 0.15 · blend 0.63 · qual 0.54 · WER 0.00
R 5.50 · genu 0.10 · blend 0.79 · qual 0.52 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.23 · genu 0.08 · blend 0.02 · qual 0.57 · WER 0.00
R 1.21 · genu 0.00 · blend 0.23 · qual 0.57 · WER 0.03
R 1.20 · genu 0.15 · blend 0.12 · qual 0.47 · WER 0.06
R 1.20 · genu 0.14 · blend 0.44 · qual 0.33 · WER 0.00
R 1.18 · genu 0.04 · blend 0.32 · qual 0.48 · WER 0.00
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.”