# Model defaults for unsloth/Qwen3-30B-A3B-Instruct-2507
# Based on Qwen3_MoE.py
# Also applies to: Qwen/Qwen3-30B-A3B-Instruct-2507, unsloth/Qwen3-30B-A3B-Instruct-2507-bnb-4bit
# MoE model - includes gate_up_proj for MoE layers

training:
  max_seq_length: 2048
  # num_epochs: 4
  num_epochs: 0
  learning_rate: 2e-4
  batch_size: 1
  gradient_accumulation_steps: 1
  warmup_steps: 5
  max_steps: 50
  save_steps: 50
  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: 32
  lora_alpha: 64
  lora_dropout: 0.0
  target_modules:
    - "q_proj"
    - "k_proj"
    - "v_proj"
    - "o_proj"
    - "gate_proj"
    - "up_proj"
    - "down_proj"
    - "gate_up_proj"
  use_rslora: false
  use_loftq: false
  use_dora: false

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

inference:
  temperature: 0.6
  top_k: 20
  top_p: 0.95