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.
cavernous-gravel-beast
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ S_NARR ▲ S_STRY ▲ GEND ▼ S_RANT ▼ EXPL ▼ DARC ▼ S_CONV · emo: Concentration Emotional_Numbness Interest
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 5.69 · genu 0.11 · blend 0.83 · qual 0.66 · WER 0.00
R 5.49 · genu 0.14 · blend 0.83 · qual 0.63 · WER 0.00
R 4.70 · genu 0.20 · blend 0.43 · qual 0.73 · WER 0.10
R 4.70 · genu 0.15 · blend 0.62 · qual 0.56 · WER 0.00
R 4.63 · genu 0.11 · blend 0.66 · qual 0.57 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.48 · genu 0.07 · blend 0.01 · qual 0.71 · WER 0.00
R 1.47 · genu 0.02 · blend 0.09 · qual 0.72 · WER 0.00
R 1.46 · genu 0.07 · blend 0.06 · qual 0.66 · WER 0.00
R 1.46 · genu 0.03 · blend 0.14 · qual 0.73 · WER 0.00
R 1.43 · genu 0.00 · blend 0.18 · qual 0.72 · 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.”