config.yaml.example

aws:
  profile: default
  region: ap-northeast-1          # Tokyo (H100 Spot available)
  # SageMaker execution role ARN (created by infrastructure/setup_iam.sh)
  role_arn: "arn:aws:iam::YOUR_ACCOUNT_ID:role/SageMakerAutoresearchRole"

sagemaker:
  instance_type: ml.p5.4xlarge    # H100 80GB (Hopper) — same as original
  use_spot: true
  max_run: 900                    # 15 min (5 min training + 10 min overhead)
  max_wait: 3600                  # 1 hour spot wait (H100 competition)
  framework_version: "2.8.0"      # SageMaker PyTorch DLC
  py_version: "py312"

s3:
  # Leave empty to use SageMaker default bucket
  bucket: ""
  data_prefix: autoresearch/data
  tokenizer_prefix: autoresearch/tokenizer
  output_prefix: autoresearch/output

pipeline:
  num_generations: 10            # Total generations
  population_size: 10            # Candidates per generation
  # Candidate diversity allocation
  num_conservative: 3            # Hyperparameter fine-tuning
  num_moderate: 4                # Architecture changes
  num_aggressive: 2              # Radical changes
  num_crossover: 1               # Combine top-2 ideas

# H100 (80GB VRAM) default model settings
model_defaults:
  depth: 8
  device_batch_size: 128
  total_batch_size: 524288       # 2^19
  window_pattern: "SSSL"