experiment:
  name: "progressive_reasoning_gemma3_1b_minimal"
  base_model: "google/gemma-3-1b-pt"
  output_dir: "./outputs"
  use_wandb: true
  wandb_project: "matsuo-llm-comp-2025"

model:
  load_in_4bit: false
  device_map: "auto"
  gradient_checkpointing: true
  use_flash_attention_2: false
  use_eager_attention: true

progressive_stages:
  - name: "basic_cot"
    description: "Basic Chain-of-Thought reasoning"
    dataset_path: "./data/basic_cot/"
    adapter_config:
      r: 4  # Extremely minimal rank
      lora_alpha: 8
      lora_dropout: 0.1
      target_modules: ["q_proj"]  # Only one module
      init_lora_weights: true
    training:
      num_epochs: 1  # Reduced epochs
      per_device_batch_size: 1
      gradient_accumulation_steps: 4
      learning_rate: 5e-4
      warmup_steps: 10
      max_length: 128  # Very short sequences
      bf16: true
      max_grad_norm: 1.0
      weight_decay: 0.001
      save_steps: 100
      logging_steps: 10
      dataloader_num_workers: 1
      dataloader_pin_memory: false

evaluation:
  benchmarks:
    - "HLE"
    - "Do-Not-Answer"
  save_results: true
  results_dir: "./outputs/evaluation_results"