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.
granite-titan
Characteristic dimensions (auto, from its own voice profile): ▲ S_STRY ▲ S_MONO ▲ S_DRAM ▲ AROU ▼ EXPL ▼ S_RANT ▼ S_ASMR ▼ 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 127/635.
Best 5 of the kept top-20%
R 11.41 · genu 0.60 · blend 0.69 · qual 0.39 · WER 0.07
R 7.69 · genu 0.39 · blend 0.62 · qual 0.40 · WER 0.00
R 5.34 · genu 0.31 · blend 0.47 · qual 0.37 · WER 0.00
R 4.93 · genu 0.14 · blend 0.62 · qual 0.54 · WER 0.00
R 4.72 · genu 0.21 · blend 0.44 · qual 0.54 · WER 0.00
Weakest 5 still inside the top-20% (the selection boundary)
R 1.32 · genu 0.13 · blend 0.08 · qual 0.52 · WER 0.00
R 1.31 · genu 0.15 · blend 0.12 · qual 0.50 · WER 0.00
R 1.31 · genu 0.16 · blend 0.27 · qual 0.40 · WER 0.00
R 1.30 · genu 0.15 · blend 0.29 · qual 0.47 · WER 0.04
R 1.30 · genu 0.17 · blend 0.08 · qual 0.50 · 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.”