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.
guttural-abyssal-fiend
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_MONO ▲ ROUG ▲ GEND ▼ REGS ▼ EXPL ▼ R_MASK ▼ S_CONV · emo: Concentration Malevolence_Malice Pride
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 122/610.
Best 5 of the kept top-20%
R 6.96 · genu 0.43 · blend 0.44 · qual 0.38 · WER 0.00
R 6.96 · genu 0.54 · blend 0.31 · qual 0.37 · WER 0.07
R 5.85 · genu 0.28 · blend 0.56 · qual 0.48 · WER 0.03
R 5.23 · genu 0.24 · blend 0.67 · qual 0.48 · WER 0.10
R 5.17 · genu 0.18 · blend 0.45 · qual 0.60 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.62 · genu 0.16 · blend 0.30 · qual 0.35 · WER 0.00
R 1.61 · genu 0.10 · blend 0.18 · qual 0.60 · WER 0.00
R 1.61 · genu 0.15 · blend 0.31 · qual 0.39 · WER 0.00
R 1.61 · genu 0.24 · blend 0.31 · qual 0.32 · WER 0.03
R 1.58 · genu 0.12 · blend 0.38 · qual 0.42 · 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.”