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.
voice-05
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ S_STRY ▲ S_NARR ▲ STRU ▼ EXPL ▼ S_RANT ▼ DARC ▼ DFLU · 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 127/639.
Best 5 of the kept top-20%
R 6.30 · genu 0.27 · blend 0.69 · qual 0.37 · WER 0.00
R 5.30 · genu 0.15 · blend 0.22 · qual 0.75 · WER 0.00
R 5.27 · genu 0.21 · blend 0.55 · qual 0.49 · WER 0.00
R 4.90 · genu 0.15 · blend 0.27 · qual 0.76 · WER 0.00
R 4.69 · genu 0.08 · blend 0.44 · qual 0.64 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.47 · genu 0.13 · blend 0.30 · qual 0.51 · WER 0.00
R 1.45 · genu 0.09 · blend 0.12 · qual 0.70 · WER 0.08
R 1.44 · genu 0.00 · blend 0.12 · qual 0.70 · WER 0.00
R 1.42 · genu 0.05 · blend 0.12 · qual 0.66 · WER 0.03
R 1.39 · genu 0.04 · blend 0.12 · qual 0.63 · 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.”