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.
dragon
Characteristic dimensions (auto, from its own voice profile): ▲ GEND ▲ S_MONO ▲ S_STRY ▲ RCQL ▼ S_RANT ▼ EXPL ▼ S_TECH ▼ DARC · emo: Concentration Interest Awe
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 120/604.
Best 5 of the kept top-20%
R 6.00 · genu 0.25 · blend 0.86 · qual 0.49 · WER 0.00
R 4.74 · genu 0.10 · blend 0.35 · qual 0.79 · WER 0.00
R 4.34 · genu 0.09 · blend 0.45 · qual 0.69 · WER 0.00
R 4.33 · genu 0.17 · blend 0.43 · qual 0.62 · WER 0.04
R 4.29 · genu 0.17 · blend 0.82 · qual 0.60 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.53 · genu 0.08 · blend 0.18 · qual 0.61 · WER 0.08
R 1.53 · genu 0.31 · blend 0.89 · qual 0.47 · WER 0.79
R 1.52 · genu 0.12 · blend 0.26 · qual 0.50 · WER 0.00
R 1.51 · genu 0.07 · blend 0.14 · qual 0.58 · WER 0.00
R 1.51 · genu 0.04 · blend 0.22 · qual 0.65 · 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.”