experiment:
  name: "progressive_reasoning_experiment"
  base_model: "microsoft/DialoGPT-small"  # Lightweight model for testing
  output_dir: "./outputs"
  use_wandb: false
  wandb_project: "matsuo-llm-comp-2025"

model:
  load_in_4bit: false  # Disable quantization for small model
  bnb_4bit_compute_dtype: "bfloat16"
  bnb_4bit_use_double_quant: true
  device_map: "auto"

progressive_stages:
  - name: "basic_cot"
    description: "Basic Chain-of-Thought reasoning"
    dataset_path: "./data/basic_cot/"
    adapter_config:
      r: 8
      lora_alpha: 16
      lora_dropout: 0.1
      target_modules: ["c_attn", "c_proj"]
    training:
      num_epochs: 2
      per_device_batch_size: 8  # Increase batch size for small model
      gradient_accumulation_steps: 2  # Reduce accumulation steps
      learning_rate: 5e-4  # Higher learning rate for faster training
      warmup_steps: 50
      max_length: 1024  # Shorter sequences

evaluation:
  benchmarks:
    - "HLE"  # Humanity's Last Exam
    - "Do-Not-Answer"
  save_results: true
  results_dir: "./outputs/evaluation_results"