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.
elf
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_NARR ▲ S_MONO ▲ S_DRAM ▼ S_RANT ▼ EXPL ▼ S_NEWS ▼ S_TECH · emo: Concentration Interest Awe
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 127/637.
Best 5 of the kept top-20%
R 5.31 · genu 0.11 · blend 0.37 · qual 0.69 · WER 0.00
R 4.90 · genu 0.16 · blend 0.49 · qual 0.60 · WER 0.00
R 4.68 · genu 0.09 · blend 0.43 · qual 0.64 · WER 0.00
R 4.51 · genu 0.17 · blend 0.15 · qual 0.68 · WER 0.00
R 4.32 · genu 0.17 · blend 0.59 · qual 0.44 · WER 0.08
Weakest 5 still inside the top-20% (the selection boundary)
R 1.18 · genu 0.14 · blend 0.21 · qual 0.48 · WER 0.03
R 1.17 · genu 0.22 · blend 0.03 · qual 0.47 · WER 0.00
R 1.17 · genu 0.06 · blend 0.04 · qual 0.68 · WER 0.00
R 1.15 · genu 0.08 · blend 0.06 · qual 0.60 · WER 0.00
R 1.14 · genu 0.06 · blend 0.26 · qual 0.52 · WER 0.03
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.”