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.
rumbling-golem
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ GEND ▲ S_MONO ▲ R_CHST ▼ REGS ▼ S_CONV ▼ EXPL ▼ S_TECH · 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 116/584.
Best 5 of the kept top-20%
R 8.65 · genu 0.57 · blend 0.33 · qual 0.30 · WER 0.03
R 5.38 · genu 0.12 · blend 0.61 · qual 0.63 · WER 0.00
R 4.65 · genu 0.30 · blend 0.14 · qual 0.52 · WER 0.07
R 4.24 · genu 0.13 · blend 0.42 · qual 0.51 · WER 0.00
R 4.11 · genu 0.18 · blend 0.54 · qual 0.30 · WER 0.03
Weakest 5 still inside the top-20% (the selection boundary)
R 1.30 · genu 0.12 · blend 0.32 · qual 0.32 · WER 0.08
R 1.30 · genu 0.04 · blend 0.00 · qual 0.73 · WER 0.00
R 1.28 · genu 0.03 · blend 0.12 · qual 0.53 · WER 0.00
R 1.27 · genu 0.09 · blend 0.26 · qual 0.47 · WER 0.07
R 1.25 · genu 0.08 · blend 0.00 · qual 0.56 · 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.”