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.
hobgoblin
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ S_STRY ▲ S_NARR ▲ GEND ▼ REGS ▼ S_CONV ▼ R_MASK ▼ R_HEAD · emo: Concentration Emotional_Numbness Interest
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 100/500.
Best 5 of the kept top-20%
R 4.53 · genu 0.15 · blend 0.96 · qual 0.64 · WER 0.80
R 4.31 · genu 0.21 · blend 0.83 · qual 0.39 · WER 0.06
R 3.59 · genu 0.22 · blend 0.84 · qual 0.59 · WER 0.74
R 3.54 · genu 0.12 · blend 0.90 · qual 0.64 · WER 0.71
R 3.40 · genu 0.11 · blend 0.78 · qual 0.65 · WER 0.54
Weakest 5 still inside the top-20% (the selection boundary)
R 1.08 · genu 0.03 · blend 0.10 · qual 0.64 · WER 0.00
R 1.07 · genu 0.06 · blend 0.00 · qual 0.70 · WER 0.10
R 1.07 · genu 0.18 · blend 0.45 · qual 0.49 · WER 0.64
R 1.06 · genu 0.08 · blend 0.22 · qual 0.57 · WER 0.05
R 1.05 · genu 0.20 · blend 0.17 · qual 0.40 · 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.”