model:
  name: /kaggle/working/stack-2.9/base_model_qwen7b
  trust_remote_code: true
  torch_dtype: float16

data:
  input_path: /kaggle/working/stack-2.9/data/final/train.jsonl
  train_dir: null
  eval_dir: null
  max_length: 2048
  train_split: 0.9
  test_split: 0.1

lora:
  r: 16
  alpha: 32
  dropout: 0.05
  target_modules:
    - q_proj
    - k_proj
    - v_proj
    - o_proj
  bias: none
  task_type: CAUSAL_LM

training:
  num_epochs: 1
  batch_size: 2
  gradient_accumulation: 4
  learning_rate: 0.0002
  warmup_steps: 50
  weight_decay: 0.01
  max_grad_norm: 1.0
  logging_steps: 5
  eval_steps: 100
  save_steps: 200
  save_total_limit: 2
  fp16: true
  bf16: false
  gradient_checkpointing: true

output:
  lora_dir: /kaggle/working/stack-2.9/training_output/lora
  merged_dir: /kaggle/working/stack-2.9/training_output/merged
  awq_dir: /kaggle/working/stack-2.9/training_output/awq

quantization:
  enabled: false
  bits: 4
  group_size: 128

logging:
  report_to: none
  wandb_project: stack-2.9-training

hardware:
  device: cuda
  num_gpus: 1
  use_4bit: false
  use_8bit: false