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.
mantid
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_MONO ▲ S_CASU ▲ S_DRAM ▼ S_NEWS ▼ EXPL ▼ S_CONV ▼ DARC · emo: Concentration Interest Awe
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 126/631.
Best 5 of the kept top-20%
R 6.21 · genu 0.54 · blend 0.76 · qual 0.25 · WER 0.88
R 4.54 · genu 0.31 · blend 0.24 · qual 0.44 · WER 0.04
R 4.49 · genu 0.20 · blend 0.69 · qual 0.39 · WER 0.00
R 4.39 · genu 0.19 · blend 0.47 · qual 0.54 · WER 0.00
R 4.38 · genu 0.15 · blend 0.94 · qual 0.57 · WER 0.79
Weakest 5 still inside the top-20% (the selection boundary)
R 1.46 · genu 0.14 · blend 0.15 · qual 0.44 · WER 0.07
R 1.44 · genu 0.22 · blend 0.12 · qual 0.42 · WER 0.00
R 1.44 · genu 0.15 · blend 0.09 · qual 0.49 · WER 0.00
R 1.44 · genu 0.19 · blend 0.01 · qual 0.54 · WER 0.03
R 1.43 · genu 0.18 · blend 0.08 · qual 0.47 · 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.”