v2: step 13000 as gpt.pth (Freya-TR-Eval WER 1.32%), v1 kept in v1/, listening samples and benchmark
981f795 verified | { | |
| "version": "v2", | |
| "base_model": "IndexTeam/IndexTTS-2.5", | |
| "fine_tuning_repo": "https://github.com/oberkayb/index-tts-2.5-ft", | |
| "checkpoint_file": "gpt.pth", | |
| "sha256": "de38009df508fd67e0abb40bb168f184b607395d9ef08675b8975767405baa2f", | |
| "step": 13000, | |
| "epoch_note": "~1.45 epochs at effective batch 32; best validation loss (5.4353) of the run, early-stopping on a 20-speaker held-out set", | |
| "mode": "lora", | |
| "lora_rank": 32, | |
| "lora_alpha": 64, | |
| "lang": "tr", | |
| "lang_token": "<|tr|>", | |
| "lang_embedding_row": 9, | |
| "lang_init_from": "es", | |
| "train_text_rows": 1530, | |
| "text_row_init": "subword", | |
| "text_lr_multiplier": 10, | |
| "lang_lr_multiplier": 10, | |
| "learning_rate": 1e-4, | |
| "normalizer": "turkish", | |
| "case": "tr_lower", | |
| "amp": "bf16", | |
| "train_hours": 486.5, | |
| "train_utterances": 143609, | |
| "train_speakers": 3652, | |
| "train_pairs": 286606, | |
| "hardware": "RTX 5070 Ti + RTX 4070 (single-process two-GPU trainer)", | |
| "freya_tr_eval": {"wer": 1.32, "wer_sd": 0.10, "cer": 0.26, "clean": 92.2, "utmos": 3.17}, | |
| "previous_version": "v1/ (step 29000, 2026-08-19)", | |
| "exported_at": "2026-10-08" | |
| } | |