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.
ancient
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_MONO ▲ ESTH ▲ S_DRAM ▼ S_TECH ▼ S_RANT ▼ EXPL ▼ S_CONV · emo: Concentration Interest Awe
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 125/626.
Best 5 of the kept top-20%
R 5.74 · genu 0.23 · blend 0.54 · qual 0.63 · WER 0.00
R 5.61 · genu 0.10 · blend 0.62 · qual 0.67 · WER 0.00
R 4.97 · genu 0.22 · blend 0.48 · qual 0.60 · WER 0.08
R 4.64 · genu 0.22 · blend 0.65 · qual 0.41 · WER 0.04
R 4.50 · genu 0.24 · blend 0.64 · qual 0.43 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.43 · genu 0.14 · blend 0.11 · qual 0.58 · WER 0.00
R 1.43 · genu 0.11 · blend 0.13 · qual 0.60 · WER 0.00
R 1.42 · genu 0.11 · blend 0.74 · qual 0.67 · WER 0.77
R 1.42 · genu 0.22 · blend 0.39 · qual 0.35 · WER 0.04
R 1.41 · genu 0.14 · blend 0.32 · qual 0.42 · 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.”