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.
mogu
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ GEND ▲ S_MONO ▲ ROUG ▼ REGS ▼ EXPL ▼ S_CONV ▼ R_HEAD · emo: Concentration Pride Malevolence_Malice
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 117/586.
Best 5 of the kept top-20%
R 6.07 · genu 0.59 · blend 0.30 · qual 0.43 · WER 0.78
R 4.66 · genu 0.12 · blend 0.57 · qual 0.73 · WER 0.54
R 4.53 · genu 0.10 · blend 0.70 · qual 0.70 · WER 0.53
R 4.37 · genu 0.11 · blend 0.72 · qual 0.45 · WER 0.03
R 4.24 · genu 0.19 · blend 0.46 · qual 0.67 · WER 0.46
Weakest 5 still inside the top-20% (the selection boundary)
R 1.38 · genu 0.10 · blend 0.40 · qual 0.60 · WER 0.72
R 1.36 · genu 0.06 · blend 0.43 · qual 0.49 · WER 0.00
R 1.34 · genu 0.17 · blend 0.18 · qual 0.47 · WER 0.00
R 1.33 · genu 0.08 · blend 0.01 · qual 0.62 · WER 0.13
R 1.32 · genu 0.01 · blend 0.39 · qual 0.74 · WER 0.74
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.”