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.
felguard
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_DRAM ▲ EMPH ▲ STNC ▼ REGS ▼ S_CONV ▼ S_TECH ▼ S_ASMR · emo: Pride Triumph Malevolence_Malice
Training set = top-20% by reward = z(genu)+z(blend)+z(quality)−z(WER)+0.7·z(profile-match). Kept 127/636.
Best 5 of the kept top-20%
R 5.14 · genu 0.12 · blend 0.51 · qual 0.51 · WER 0.00
R 4.76 · genu 0.30 · blend 0.38 · qual 0.23 · WER 0.00
R 4.50 · genu 0.20 · blend 0.51 · qual 0.23 · WER 0.03
R 4.48 · genu 0.24 · blend 0.45 · qual 0.23 · WER 0.03
R 4.32 · genu 0.16 · blend 0.45 · qual 0.33 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.17 · genu 0.08 · blend 0.25 · qual 0.40 · WER 0.05
R 1.17 · genu 0.09 · blend 0.16 · qual 0.33 · WER 0.00
R 1.16 · genu 0.13 · blend 0.26 · qual 0.30 · WER 0.00
R 1.16 · genu 0.06 · blend 0.14 · qual 0.40 · WER 0.03
R 1.16 · genu 0.13 · blend 0.46 · qual 0.22 · 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.”