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.
zandalari_troll
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ AROU ▲ S_DRAM ▲ RANG ▼ S_RANT ▼ EXPL ▼ S_TECH ▼ COGL · emo: Concentration Interest Emotional_Numbness
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 126/634.
Best 5 of the kept top-20%
R 6.78 · genu 0.13 · blend 0.59 · qual 0.55 · WER 0.03
R 5.78 · genu 0.10 · blend 0.61 · qual 0.56 · WER 0.00
R 5.55 · genu 0.17 · blend 0.51 · qual 0.47 · WER 0.00
R 5.20 · genu 0.11 · blend 0.60 · qual 0.47 · WER 0.00
R 4.83 · genu 0.29 · blend 0.49 · qual 0.40 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.19 · genu 0.11 · blend 0.00 · qual 0.60 · WER 0.07
R 1.18 · genu 0.00 · blend 0.04 · qual 0.67 · WER 0.00
R 1.17 · genu 0.13 · blend 0.11 · qual 0.58 · WER 0.08
R 1.17 · genu 0.10 · blend 0.00 · qual 0.59 · WER 0.00
R 1.17 · genu 0.04 · blend 0.11 · qual 0.53 · 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.”