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.
hozen
Characteristic dimensions (auto, from its own voice profile): ▲ S_CASU ▲ S_DRAM ▲ VOLT ▲ GEND ▼ EXPL ▼ R_HEAD ▼ S_TECH ▼ S_WHIS · emo: Pain Malevolence_Malice Anger
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 41/208.
Best 5 of the kept top-20%
R 4.36 · genu 0.54 · blend 0.40 · qual 0.34 · WER 0.44
R 4.04 · genu 0.16 · blend 0.70 · qual 0.52 · WER 0.04
R 2.84 · genu 0.19 · blend 0.78 · qual 0.32 · WER 0.09
R 2.69 · genu 0.06 · blend 0.46 · qual 0.46 · WER 0.00
R 2.62 · genu 0.20 · blend 0.87 · qual 0.33 · WER 1.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.37 · genu 0.10 · blend 0.25 · qual 0.50 · WER 0.00
R 1.35 · genu 0.20 · blend 0.40 · qual 0.31 · WER 0.26
R 1.34 · genu 0.27 · blend 0.54 · qual 0.24 · WER 0.76
R 1.30 · genu 0.37 · blend 0.73 · qual 0.08 · WER 0.96
R 1.27 · genu 0.18 · blend 0.43 · qual 0.27 · 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.”