{
 "engine": "cosyvoice3",
 "base": "FunAudioLLM/Fun-CosyVoice3-0.5B-2512",
 "rate": "normal",
 "sampler": "silexcap",
 "one_pass": true,
 "instruct": true,
 "cfg": 2.0,
 "text_pauses": true,
 "sha256_best_pt": "87c3a7e9a806e89cf1b77cff78989ebe63938016f5d8066cf673d350c4cab3db",
 "autoresearch_artifact": "art-dbr28 (private)",
 "training": "Flash-DPO LM -> Flash-Instruct SFT (3,000 steps, ~100.9k licensed clips with audio-derived pace/pause/energy labels, natural-language instruction in the CosyVoice3 system-prompt slot, 18% default-prompt items, pause/energy control embeddings) -> DPO polish (700 prompts x 4 samples, 564 preference pairs, reward = CER + pace and pause match to native + clipped-ending penalty + small UTMOS guard; 134 steps, beta 0.1, lr 1e-6, SFT weight 0.1; final step kept, val pref-acc 0.61 (0.75 after epoch 1, 28-pair val set))",
 "data": "licensed only (no African Voices / WaZoBiaSpeech, no NaijaVoices)"
}