dataset:
  # Seed for the random number generator, used for signal generation
  seed: 42

  # Number of samples per recording
  recording_length: 1024

  # Set this to scale the number of generated recordings
  mult_factor: 5

  # List of signal modulation schemes to include in the dataset
  modulation_types:
    - bpsk
    - qpsk
    - qam16
    - qam64

  # Rolloff factor for pulse shaping filter (0 < beta <= 1)
  beta: 0.3

  # Samples per symbol (determines bandwidth of the digital signal)
  sps: 4

  # SNR sweep range: start, stop (exclusive), and step (in dB)
  snr_start: -6
  snr_stop: 13
  snr_step: 3

  # Number of iterations (signal recordings) per modulation and SNR combination
  num_iterations: 3

  # Modulation scheme settings; keys must match the `modulation_types` list above
  # Each entry includes:
  #   - num_bits_per_symbol: bits encoded per symbol (e.g., 1 for BPSK, 4 for 16-QAM)
  #   - constellation_type: modulation category (e.g., "psk", "qam", "fsk", "ofdm")
  # TODO: Combine entries for 'modulation_types' and 'modulation_settings'
  modulation_settings:
    bpsk:
      num_bits_per_symbol: 1
      constellation_type: psk
    qpsk:
      num_bits_per_symbol: 2
      constellation_type: psk
    qam16:
      num_bits_per_symbol: 4
      constellation_type: qam
    qam64:
      num_bits_per_symbol: 6
      constellation_type: qam

  # Number of slices to cut from each recording
  num_slices: 8

  # Training and validation split ratios; must sum to 1
  train_split: 0.8
  val_split: 0.2

training:
  # Number of training examples processed together before the model updates its weights
  batch_size: 256

  # Number of complete passes through the training dataset during training
  epochs: 5

  # Learning rate: step size for weight updates after each batch
  # Recommended range for fine-tuning: 1e-6 to 1e-4
  learning_rate: 1e-4

  # Enable GPU acceleration for training if available
  use_gpu: true

  # Dropout rate for individual neurons/layers (probability of dropping out a unit)
  drop_rate: 0.5

  # Drop path rate: probability of dropping entire residual paths (stochastic depth)
  drop_path_rate: 0.2

  # Weight decay (L2 regularization) coefficient to help prevent overfitting
  wd: 0.01

app:
  # Optimization style for ORT conversion; options: 'Fixed', 'None'
  optimization_style: "Fixed"

  # Target platform architecture; common options: 'amd64', 'arm64'
  target_platform: "amd64"