# Model defaults for unsloth/Qwen2.5-1.5B-Instruct
# Based on nemo_gym_sudoku.ipynb
# Also applies to: unsloth/Qwen2.5-1.5B-Instruct-unsloth-bnb-4bit, Qwen/Qwen2.5-1.5B-Instruct, unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit

training:
  max_seq_length: 4096
  # num_epochs: 4
  num_epochs: 0
  learning_rate: 1e-5
  batch_size: 1
  gradient_accumulation_steps: 64
  warmup_ratio: 0.1
  max_steps: 30
  save_steps: 30
  weight_decay: 0.001
  random_seed: 42
  packing: false
  train_on_completions: true
  gradient_checkpointing: "unsloth"
  optim: "adamw_8bit"
  lr_scheduler_type: "linear"

lora:
  lora_r: 4
  lora_alpha: 8
  lora_dropout: 0.0
  target_modules:
    - "q_proj"
    - "k_proj"
    - "v_proj"
    - "o_proj"
    - "gate_proj"
    - "up_proj"
    - "down_proj"
  use_rslora: false
  use_loftq: false

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