training_summary.json

{
  "lora": true,
  "lora_r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "lora_scope": "all",
  "lora_target_modules": "auto",
  "lora_target_modules_resolved": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj",
    "c_attn",
    "c_proj",
    "fc_in",
    "fc_out"
  ],
  "trainable_parameters": 33669120,
  "learning_rate": 0.0001,
  "lr_scheduler_type": "constant",
  "num_epochs": 3,
  "per_device_batch_size": 4,
  "gradient_accumulation_steps": 4,
  "mixed_precision": "bf16",
  "n_vq": null,
  "validation_split_ratio": 0.01,
  "best_val_loss": 4.311636394729263,
  "saved_as_best": true,
  "saved_epoch": 1,
  "saved_global_step": 6716,
  "saved_at": "2026-06-28 17:05:52",
  "base_model": "OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5",
  "codec_model": "OpenMOSS-Team/MOSS-Audio-Tokenizer-v2",
  "artifact": "LoRA adapter plus merged full model weights"
}