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.
furbolg
Characteristic dimensions (auto, from its own voice profile): ▲ GEND ▲ S_MONO ▲ S_STRY ▲ S_CASU ▼ EXPL ▼ R_HEAD ▼ DARC ▼ VALN · emo: Concentration Emotional_Numbness Awe
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 8/37.
Best 5 of the kept top-20%
R 3.46 · genu 0.25 · blend 0.52 · qual 0.64 · WER 0.10
R 2.80 · genu 0.25 · blend 0.59 · qual 0.44 · WER 0.08
R 2.39 · genu 0.11 · blend 0.59 · qual 0.61 · WER 0.00
R 2.36 · genu 0.24 · blend 0.57 · qual 0.37 · WER 0.00
R 1.95 · genu 0.20 · blend 0.35 · qual 0.61 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 2.36 · genu 0.24 · blend 0.57 · qual 0.37 · WER 0.00
R 1.95 · genu 0.20 · blend 0.35 · qual 0.61 · WER 0.00
R 1.83 · genu 0.21 · blend 0.74 · qual 0.39 · WER 0.04
R 1.69 · genu 0.22 · blend 0.82 · qual 0.66 · WER 0.92
R 1.61 · genu 0.16 · blend 0.67 · qual 0.35 · 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.”