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.
golem
Characteristic dimensions (auto, from its own voice profile): ▲ S_MONO ▲ GEND ▲ S_STRY ▲ R_CHST ▼ REGS ▼ EXPL ▼ S_CONV ▼ S_TECH · 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 126/630.
Best 5 of the kept top-20%
R 5.76 · genu 0.31 · blend 0.23 · qual 0.38 · WER 0.00
R 5.29 · genu 0.27 · blend 0.08 · qual 0.50 · WER 0.00
R 5.28 · genu 0.26 · blend 0.53 · qual 0.30 · WER 0.18
R 5.07 · genu 0.19 · blend 0.21 · qual 0.49 · WER 0.00
R 4.83 · genu 0.15 · blend 0.31 · qual 0.45 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.24 · genu 0.14 · blend 0.13 · qual 0.32 · WER 0.00
R 1.23 · genu 0.10 · blend 0.25 · qual 0.39 · WER 0.00
R 1.23 · genu 0.00 · blend 0.08 · qual 0.52 · WER 0.00
R 1.21 · genu 0.13 · blend 0.10 · qual 0.43 · WER 0.13
R 1.21 · genu 0.04 · blend 0.05 · qual 0.49 · 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.”