finetune_config.json

{
  "model_path": "./models/MOSS-TTS-Nano",
  "codec_path": "./models/MOSS-Audio-Tokenizer-Nano",
  "train_jsonl": "/workspace/train_ref_with_codes.jsonl",
  "output_dir": "/workspace/output/moss_tts_ref_v6",
  "max_length": 1024,
  "per_device_batch_size": 4,
  "gradient_accumulation_steps": 4,
  "learning_rate": 0.0001,
  "weight_decay": 0.1,
  "adam_beta1": 0.9,
  "adam_beta2": 0.95,
  "adam_eps": 1e-08,
  "warmup_steps": 0,
  "warmup_ratio": 0.05,
  "lr_scheduler_type": "linear",
  "num_epochs": 50,
  "max_train_steps": null,
  "max_grad_norm": 1.0,
  "logging_steps": 10,
  "save_every_epochs": 1,
  "num_workers": 0,
  "mixed_precision": "fp16",
  "attn_implementation": "sdpa",
  "channelwise_loss_weight": "1,8",
  "seed": 42,
  "resolved_warmup_steps": 6523,
  "resolved_channelwise_loss_weight": [
    1.0,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5,
    0.5
  ],
  "global_batch_size": 16,
  "records_paths": [
    "/workspace/train_ref_with_codes.jsonl"
  ],
  "saved_global_step": 75661,
  "saved_epoch": 29,
  "saved_at": "2026-05-08 06:33:19",
  "checkpoint_dir": "/workspace/output/moss_tts_ref_v6/checkpoint-epoch-29"
}