# Model defaults for unsloth/Llasa-3B
# Based on Llasa_TTS_(3B).ipynb and Llasa_TTS_(1B).ipynb
# Also applies to: HKUSTAudio/Llasa-1B
# added inference parameters from unsloth notebook

training:
  trust_remote_code: false
  max_seq_length: 2048
  # num_epochs: 4
  num_epochs: 0
  learning_rate: 5e-4
  batch_size: 2
  gradient_accumulation_steps: 4
  warmup_steps: 5
  max_steps: 30
  save_steps: 30
  weight_decay: 0.001
  random_seed: 3407
  packing: false
  train_on_completions: true
  gradient_checkpointing: "unsloth"
  optim: "adamw_8bit"
  lr_scheduler_type: "linear"

lora:
  lora_r: 128
  lora_alpha: 128
  lora_dropout: 0.0
  target_modules:
    - "q_proj"
    - "v_proj"
  use_rslora: false
  use_loftq: false

logging:
  enable_wandb: false
  wandb_project: "llm-finetuning"
  enable_tensorboard: false
  tensorboard_dir: "runs"
  log_frequency: 10

inference:
  trust_remote_code: false
  temperature: 1.2
  top_p: 1.2