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.
ogre
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_MONO ▲ GEND ▲ R_CHST ▼ S_CONV ▼ EXPL ▼ REGS ▼ S_TECH · 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 123/618.
Best 5 of the kept top-20%
R 5.99 · genu 0.15 · blend 0.48 · qual 0.70 · WER 0.00
R 5.32 · genu 0.51 · blend 0.28 · qual 0.28 · WER 0.00
R 4.87 · genu 0.09 · blend 0.58 · qual 0.55 · WER 0.00
R 4.79 · genu 0.27 · blend 0.63 · qual 0.45 · WER 0.33
R 4.60 · genu 0.17 · blend 0.66 · qual 0.65 · WER 0.43
Weakest 5 still inside the top-20% (the selection boundary)
R 1.38 · genu 0.16 · blend 0.17 · qual 0.42 · WER 0.03
R 1.37 · genu 0.08 · blend 0.15 · qual 0.55 · WER 0.00
R 1.37 · genu 0.25 · blend 0.42 · qual 0.14 · WER 0.03
R 1.36 · genu 0.09 · blend 0.00 · qual 0.60 · WER 0.00
R 1.35 · genu 0.12 · blend 0.04 · qual 0.51 · 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.”