All six completed RL runs: immediate listening preview

Saved training rollouts, NOT final-checkpoint evaluation predictions. These clips were generated by intermediate policies during the completed runs. They are known training inputs and cannot establish held-out performance or a final-checkpoint winner.

Four ordinary, non-explicit real-speech texts are shared by all six runs, with and without a reference:48 actual128-kbps mono MP3s. Candidate0 is fixed; no best-of-N and no selection by score. The latest saved occurrence of each chosen UID is used. Invalid waveforms cannot be played, but transcript-gate failures are not removed. Seeds match across recipes within a model size; Small and Large case orders can differ. Group numbers and seeds are printed below. Do not compare these training-judge scores directly with the separately MP3-scored207-case evaluation.

Open the matched final-checkpoint grid. Its four new SFT6-RL evaluations are queued and publish automatically; the two AlphaRL-v2 final evaluations already contain207 cases each. References are identified below but are not embedded in this rollout-only preview.

Completed training run	AttemptedG8groups	Actual optimizer updates	This preview
Small Alpha standard RL-v2 · FINAL · 2,256 updates	2616	2256	8 saved rollouts
Large Alpha standard RL-v2 · FINAL · 2,344 updates	2530	2344	8 saved rollouts
Small SFT6 standard RL · FINAL · 2,324 updates	2616	2324	8 saved rollouts
Small SFT6 emotion RL · FINAL · 2,384 updates	2616	2384	8 saved rollouts
Large SFT6 standard RL + 10% inverse-WER · FINAL · 2,360 updates	2530	2360	8 saved rollouts
Large SFT6 emotion RL + 10% inverse-WER · FINAL · 2,378 updates	2530	2378	8 saved rollouts

EN_B00000_S02089_W000005 · reference

I mean, George had the jacket for it and I had, you know, had my swagger going. It like, it looked good, it fit, everything was right except for our skills. We're no good.

Exact training prompt and conditioning

GENERAL: Express interest intensely. Use extremely casual delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.
SCRIPT:
(Express interest intensely. Use extremely casual delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.) I mean, George had the jacket for it and I had, you know, had my swagger going. It like, it looked good, it fit, everything was right except for our skills. We're no good.

Reference UID: EN_B00000_S02089_W000009.

Small Alpha standard RL-v2 · FINAL · 2,256 updates

TRAINING ROLLOUT · intermediate policy

CE 6.428 · genuineness 4.645 · PQ 7.899 · identity cosine 0.280 · WER 0.0%

Training group 428; seed 134450; candidate0; transcript gate passed.

Large Alpha standard RL-v2 · FINAL · 2,344 updates

TRAINING ROLLOUT · intermediate policy

CE 6.459 · genuineness 4.959 · PQ 7.720 · identity cosine 0.526 · WER 2.9%

Training group 1350; seed 141826; candidate0; transcript gate passed.

Small SFT6 standard RL · FINAL · 2,324 updates

TRAINING ROLLOUT · intermediate policy

CE 6.078 · genuineness 5.122 · PQ 7.576 · identity cosine 0.291 · WER 0.0%

Training group 428; seed 134450; candidate0; transcript gate passed.

Small SFT6 emotion RL · FINAL · 2,384 updates

TRAINING ROLLOUT · intermediate policy

CE 6.285 · genuineness 4.696 · PQ 7.837 · identity cosine 0.267 · WER 5.9%

Training group 428; seed 134450; candidate0; transcript gate passed.

Large SFT6 standard RL + 10% inverse-WER · FINAL · 2,360 updates

TRAINING ROLLOUT · intermediate policy

CE 6.475 · genuineness 4.256 · PQ 7.926 · identity cosine 0.691 · WER 0.0%

Training group 1350; seed 141826; candidate0; transcript gate passed.

Large SFT6 emotion RL + 10% inverse-WER · FINAL · 2,378 updates

TRAINING ROLLOUT · intermediate policy

CE 6.467 · genuineness 4.495 · PQ 7.871 · identity cosine 0.623 · WER 2.9%

Training group 1350; seed 141826; candidate0; transcript gate passed.

EN_B00000_S02089_W000005 · no_reference

I mean, George had the jacket for it and I had, you know, had my swagger going. It like, it looked good, it fit, everything was right except for our skills. We're no good.

Exact training prompt and conditioning

GENERAL: Use a masculine voice in a lower register, with a balanced resonance and a lightly textured timbre. Express interest intensely. Use extremely casual delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.
SCRIPT:
(Express interest intensely. Use extremely casual delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.) I mean, George had the jacket for it and I had, you know, had my swagger going. It like, it looked good, it fit, everything was right except for our skills. We're no good.

Reference UID: EN_B00000_S02089_W000009.

Small Alpha standard RL-v2 · FINAL · 2,256 updates

TRAINING ROLLOUT · intermediate policy

CE 6.571 · genuineness 4.922 · PQ 7.885 · WER 2.9%

Training group 429; seed 134458; candidate0; transcript gate passed.

Large Alpha standard RL-v2 · FINAL · 2,344 updates

TRAINING ROLLOUT · intermediate policy

CE 6.507 · genuineness 4.696 · PQ 7.933 · WER 2.9%

Training group 1351; seed 141834; candidate0; transcript gate passed.

Small SFT6 standard RL · FINAL · 2,324 updates

TRAINING ROLLOUT · intermediate policy

CE 6.420 · genuineness 5.148 · PQ 7.758 · WER 17.6%

Training group 429; seed 134458; candidate0; transcript gate failed.

Small SFT6 emotion RL · FINAL · 2,384 updates

TRAINING ROLLOUT · intermediate policy

CE 6.420 · genuineness 4.558 · PQ 8.063 · WER 0.0%

Training group 429; seed 134458; candidate0; transcript gate passed.

Large SFT6 standard RL + 10% inverse-WER · FINAL · 2,360 updates

TRAINING ROLLOUT · intermediate policy

CE 6.555 · genuineness 4.658 · PQ 7.899 · WER 2.9%

Training group 1351; seed 141834; candidate0; transcript gate passed.

Large SFT6 emotion RL + 10% inverse-WER · FINAL · 2,378 updates

TRAINING ROLLOUT · intermediate policy

CE 6.452 · genuineness 4.771 · PQ 8.050 · WER 2.9%

Training group 1351; seed 141834; candidate0; transcript gate passed.

EN_B00000_S08561_W000001 · reference

I feel like we are in a VIP car. We are back at the airport. Time to go. This was fun. I wish we could stay one more day here because the sea is just amazing. But it's time to go back. That's our plane!

Exact training prompt and conditioning

GENERAL: Express interest intensely. Use extremely casual delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.
SCRIPT:
(Express interest intensely. Use extremely casual delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.) I feel like we are in a VIP car. We are back at the airport. Time to go. This was fun. I wish we could stay one more day here because the sea is just amazing. But it's time to go back. That's our plane!

Reference UID: EN_B00000_S08561_W000000.

Small Alpha standard RL-v2 · FINAL · 2,256 updates

TRAINING ROLLOUT · intermediate policy

CE 6.499 · genuineness 4.457 · PQ 7.837 · identity cosine 0.060 · WER 0.0%

Training group 672; seed 136402; candidate0; transcript gate passed.

Large Alpha standard RL-v2 · FINAL · 2,344 updates

TRAINING ROLLOUT · intermediate policy

CE 6.515 · genuineness 4.620 · PQ 7.668 · identity cosine 0.084 · WER 2.2%

Training group 1588; seed 143730; candidate0; transcript gate passed.

Small SFT6 standard RL · FINAL · 2,324 updates

TRAINING ROLLOUT · intermediate policy

CE 6.602 · genuineness 4.595 · PQ 7.844 · identity cosine 0.014 · WER 0.0%

Training group 672; seed 136402; candidate0; transcript gate passed.

Small SFT6 emotion RL · FINAL · 2,384 updates

TRAINING ROLLOUT · intermediate policy

CE 6.523 · genuineness 4.087 · PQ 7.864 · identity cosine 0.087 · WER 0.0%

Training group 672; seed 136402; candidate0; transcript gate passed.

Large SFT6 standard RL + 10% inverse-WER · FINAL · 2,360 updates

TRAINING ROLLOUT · intermediate policy

CE 6.602 · genuineness 4.558 · PQ 7.940 · identity cosine 0.230 · WER 0.0%

Training group 1588; seed 143730; candidate0; transcript gate passed.

Large SFT6 emotion RL + 10% inverse-WER · FINAL · 2,378 updates

TRAINING ROLLOUT · intermediate policy

CE 6.444 · genuineness 4.482 · PQ 7.765 · identity cosine 0.158 · WER 0.0%

Training group 1588; seed 143730; candidate0; transcript gate passed.

EN_B00000_S08561_W000001 · no_reference

Exact training prompt and conditioning

GENERAL: Use a masculine voice in a lower register, with a balanced resonance and a lightly textured timbre. Express interest intensely. Use extremely casual delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.
SCRIPT:
(Express interest intensely. Use extremely casual delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.) I feel like we are in a VIP car. We are back at the airport. Time to go. This was fun. I wish we could stay one more day here because the sea is just amazing. But it's time to go back. That's our plane!

Reference UID: EN_B00000_S08561_W000000.

Small Alpha standard RL-v2 · FINAL · 2,256 updates

TRAINING ROLLOUT · intermediate policy

CE 6.610 · genuineness 4.419 · PQ 7.960 · WER 0.0%

Training group 673; seed 136410; candidate0; transcript gate passed.

Large Alpha standard RL-v2 · FINAL · 2,344 updates

TRAINING ROLLOUT · intermediate policy

CE 6.571 · genuineness 4.250 · PQ 7.857 · WER 0.0%

Training group 1589; seed 143738; candidate0; transcript gate passed.

Small SFT6 standard RL · FINAL · 2,324 updates

TRAINING ROLLOUT · intermediate policy

CE 6.523 · genuineness 4.658 · PQ 7.905 · WER 2.2%

Training group 673; seed 136410; candidate0; transcript gate passed.

Small SFT6 emotion RL · FINAL · 2,384 updates

TRAINING ROLLOUT · intermediate policy

CE 6.594 · genuineness 4.288 · PQ 7.940 · WER 0.0%

Training group 673; seed 136410; candidate0; transcript gate passed.

Large SFT6 standard RL + 10% inverse-WER · FINAL · 2,360 updates

TRAINING ROLLOUT · intermediate policy

CE 6.579 · genuineness 4.200 · PQ 8.043 · WER 0.0%

Training group 1589; seed 143738; candidate0; transcript gate passed.

Large SFT6 emotion RL + 10% inverse-WER · FINAL · 2,378 updates

TRAINING ROLLOUT · intermediate policy

CE 6.587 · genuineness 4.350 · PQ 7.967 · WER 0.0%

Training group 1589; seed 143738; candidate0; transcript gate passed.

DE_B00000_S00716_W000055 · reference

Um ein schönes Gericht zu machen. Das sind diese dankbaren Kräuter, die später im Jahr kommen.

Exact training prompt and conditioning

GENERAL: Express thankfulness gratitude extremely. Use intensely conversational delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.
SCRIPT:
(Express thankfulness gratitude extremely. Use intensely conversational delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.) Um ein schönes Gericht zu machen. Das sind diese dankbaren Kräuter, die später im Jahr kommen.

Reference UID: DE_B00000_S00716_W000052.

Small Alpha standard RL-v2 · FINAL · 2,256 updates

TRAINING ROLLOUT · intermediate policy

CE 6.261 · genuineness 4.783 · PQ 7.953 · identity cosine 0.212 · WER 0.0%

Training group 1844; seed 145778; candidate0; transcript gate passed.

Large Alpha standard RL-v2 · FINAL · 2,344 updates

TRAINING ROLLOUT · intermediate policy

CE 6.396 · genuineness 4.407 · PQ 8.036 · identity cosine 0.289 · WER 0.0%

Training group 538; seed 135330; candidate0; transcript gate passed.

Small SFT6 standard RL · FINAL · 2,324 updates

TRAINING ROLLOUT · intermediate policy

CE 6.269 · genuineness 4.834 · PQ 7.988 · identity cosine 0.223 · WER 0.0%

Training group 1844; seed 145778; candidate0; transcript gate passed.

Small SFT6 emotion RL · FINAL · 2,384 updates

TRAINING ROLLOUT · intermediate policy

CE 6.420 · genuineness 4.758 · PQ 7.995 · identity cosine 0.257 · WER 0.0%

Training group 1844; seed 145778; candidate0; transcript gate passed.

Large SFT6 standard RL + 10% inverse-WER · FINAL · 2,360 updates

TRAINING ROLLOUT · intermediate policy

CE 6.301 · genuineness 4.445 · PQ 8.077 · identity cosine 0.443 · WER 0.0%

Training group 538; seed 135330; candidate0; transcript gate passed.

Large SFT6 emotion RL + 10% inverse-WER · FINAL · 2,378 updates

TRAINING ROLLOUT · intermediate policy

CE 6.459 · genuineness 4.583 · PQ 8.002 · identity cosine 0.315 · WER 0.0%

Training group 538; seed 135330; candidate0; transcript gate passed.

DE_B00000_S00716_W000055 · no_reference

Um ein schönes Gericht zu machen. Das sind diese dankbaren Kräuter, die später im Jahr kommen.

Exact training prompt and conditioning

GENERAL: Use a feminine voice in a middle register, with a balanced resonance and a smooth timbre. Express thankfulness gratitude extremely. Use intensely conversational delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.
SCRIPT:
(Express thankfulness gratitude extremely. Use intensely conversational delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.) Um ein schönes Gericht zu machen. Das sind diese dankbaren Kräuter, die später im Jahr kommen.

Reference UID: DE_B00000_S00716_W000052.

Small Alpha standard RL-v2 · FINAL · 2,256 updates

TRAINING ROLLOUT · intermediate policy

CE 6.372 · genuineness 4.733 · PQ 8.153 · WER 0.0%

Training group 1845; seed 145786; candidate0; transcript gate passed.

Large Alpha standard RL-v2 · FINAL · 2,344 updates

TRAINING ROLLOUT · intermediate policy

CE 6.388 · genuineness 4.859 · PQ 8.015 · WER 0.0%

Training group 539; seed 135338; candidate0; transcript gate passed.

Small SFT6 standard RL · FINAL · 2,324 updates

TRAINING ROLLOUT · intermediate policy

CE 6.459 · genuineness 4.633 · PQ 8.002 · WER 0.0%

Training group 1845; seed 145786; candidate0; transcript gate passed.

Small SFT6 emotion RL · FINAL · 2,384 updates

TRAINING ROLLOUT · intermediate policy

CE 6.348 · genuineness 4.608 · PQ 7.947 · WER 0.0%

Training group 1845; seed 145786; candidate0; transcript gate passed.

Large SFT6 standard RL + 10% inverse-WER · FINAL · 2,360 updates

TRAINING ROLLOUT · intermediate policy

CE 6.404 · genuineness 4.319 · PQ 8.077 · WER 0.0%

Training group 539; seed 135338; candidate0; transcript gate passed.

Large SFT6 emotion RL + 10% inverse-WER · FINAL · 2,378 updates

TRAINING ROLLOUT · intermediate policy

CE 6.301 · genuineness 4.608 · PQ 7.960 · WER 0.0%

Training group 539; seed 135338; candidate0; transcript gate passed.

DE_B00000_S01186_W000110 · reference

Und so, so unvorhergesehen war es für mich mit, mit Snoke und, äh, Kylo. Also ich hätte nie im Leben gedacht, dass das so passiert.

Exact training prompt and conditioning

GENERAL: Express contemplation intensely. Use intensely conversational delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.
SCRIPT:
(Express contemplation intensely. Use intensely conversational delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.) Und so, so unvorhergesehen war es für mich mit, mit Snoke und, äh, Kylo. Also ich hätte nie im Leben gedacht, dass das so passiert.

Reference UID: DE_B00000_S01186_W000103.

Small Alpha standard RL-v2 · FINAL · 2,256 updates

TRAINING ROLLOUT · intermediate policy

CE 5.832 · genuineness 4.859 · PQ 7.682 · identity cosine 0.567 · WER 20.0%

Training group 1286; seed 141314; candidate0; transcript gate failed.

Large Alpha standard RL-v2 · FINAL · 2,344 updates

TRAINING ROLLOUT · intermediate policy

CE 6.054 · genuineness 4.633 · PQ 7.751 · identity cosine 0.663 · WER 24.0%

Training group 2226; seed 148834; candidate0; transcript gate failed.

Small SFT6 standard RL · FINAL · 2,324 updates

TRAINING ROLLOUT · intermediate policy

CE 6.023 · genuineness 4.809 · PQ 7.665 · identity cosine 0.689 · WER 20.0%

Training group 1286; seed 141314; candidate0; transcript gate failed.

Small SFT6 emotion RL · FINAL · 2,384 updates

TRAINING ROLLOUT · intermediate policy

CE 5.903 · genuineness 4.595 · PQ 7.775 · identity cosine 0.731 · WER 20.0%

Training group 1286; seed 141314; candidate0; transcript gate failed.

Large SFT6 standard RL + 10% inverse-WER · FINAL · 2,360 updates

TRAINING ROLLOUT · intermediate policy

CE 5.721 · genuineness 4.633 · PQ 7.411 · identity cosine 0.705 · WER 20.0%

Training group 2226; seed 148834; candidate0; transcript gate failed.

Large SFT6 emotion RL + 10% inverse-WER · FINAL · 2,378 updates

TRAINING ROLLOUT · intermediate policy

CE 5.991 · genuineness 4.671 · PQ 7.607 · identity cosine 0.811 · WER 16.0%

Training group 2226; seed 148834; candidate0; transcript gate failed.

DE_B00000_S01186_W000110 · no_reference

Und so, so unvorhergesehen war es für mich mit, mit Snoke und, äh, Kylo. Also ich hätte nie im Leben gedacht, dass das so passiert.

Exact training prompt and conditioning

GENERAL: Use a masculine voice in a lower register, with a balanced resonance and a lightly textured timbre. Express contemplation intensely. Use intensely conversational delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.
SCRIPT:
(Express contemplation intensely. Use intensely conversational delivery. Keep every spoken word intelligible and preserve spontaneous human phrasing.) Und so, so unvorhergesehen war es für mich mit, mit Snoke und, äh, Kylo. Also ich hätte nie im Leben gedacht, dass das so passiert.

Reference UID: DE_B00000_S01186_W000103.