experiment:
  name: "progressive_reasoning_8gpu"
  base_model: "google/gemma-2-2b-it"  # Can scale up to larger models
  output_dir: "./outputs"
  use_wandb: true
  wandb_project: "matsuo-llm-comp-2025"

model:
  load_in_4bit: false  # Can use FP16/BF16 with multiple GPUs
  bnb_4bit_compute_dtype: "bfloat16"
  bnb_4bit_use_double_quant: true
  device_map: "balanced"  # Distribute across all GPUs
  gradient_checkpointing: true
  use_flash_attention_2: true  # Enable if available for speed
  use_eager_attention: false

# Multi-GPU specific settings
distributed:
  strategy: "ddp"  # Distributed Data Parallel
  find_unused_parameters: false
  gradient_as_bucket_view: true

progressive_stages:
  - name: "basic_cot"
    description: "Basic Chain-of-Thought reasoning"
    dataset_path: "./data/basic_cot/"
    adapter_config:
      r: 32  # Larger rank since we have more memory
      lora_alpha: 64
      lora_dropout: 0.1
      target_modules: ["q_proj", "k_proj", "v_proj", "o_proj"]
      init_lora_weights: true
    training:
      num_epochs: 2
      per_device_batch_size: 16  # Large batch size per GPU
      gradient_accumulation_steps: 1  # No need for accumulation with 8 GPUs
      learning_rate: 5e-4
      warmup_steps: 100
      max_length: 2048
      bf16: true
      max_grad_norm: 1.0
      weight_decay: 0.001
      save_steps: 50
      logging_steps: 10
      dataloader_num_workers: 4  # More workers for data loading
      dataloader_pin_memory: true
      remove_unused_columns: false

  - name: "math_reasoning"
    description: "Mathematical reasoning with OpenR1-Math-220k dataset"
    dataset_path: "open-r1/OpenR1-Math-220k"
    inherit_from: "basic_cot"
    adapter_config:
      r: 64
      lora_alpha: 128
      lora_dropout: 0.1
      target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
      init_lora_weights: true
    training:
      num_epochs: 1
      per_device_batch_size: 8  # Reduce for larger model
      gradient_accumulation_steps: 2
      learning_rate: 3e-4
      warmup_steps: 200
      max_length: 4096
      bf16: true
      max_grad_norm: 1.0
      weight_decay: 0.001
      save_steps: 100
      logging_steps: 20
      dataloader_num_workers: 4
    dataset_config:
      streaming: true
      max_samples: 100000  # Can process more with 8 GPUs
      split: "train"

  - name: "complex_reasoning"
    description: "Complex multi-step reasoning with Mixture-of-Thoughts"
    dataset_path: "open-r1/Mixture-of-Thoughts"
    inherit_from: "math_reasoning"
    adapter_config:
      r: 128  # Maximum rank with multi-GPU
      lora_alpha: 256
      lora_dropout: 0.1
      target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
      init_lora_weights: true
    training:
      num_epochs: 1
      per_device_batch_size: 4
      gradient_accumulation_steps: 4
      learning_rate: 2e-4
      warmup_steps: 300
      max_length: 8192
      bf16: true
      max_grad_norm: 1.0
      weight_decay: 0.001
      save_steps: 200
      logging_steps: 50
      dataloader_num_workers: 4
    dataset_config:
      streaming: true
      max_samples: 50000
      split: "train"

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