model:
  dim: 256
  depth: 6
  heads: 8
  dim_head: 32
  vocab_size: 32
  phase_dim: 16
  max_seq_len: 32
  dropout: 0.1
  use_set_encoder: false
  diffusion_type: flow_matching
  num_timesteps: 1000
  beta_schedule: cosine
  use_ot_coupling: false
  use_quadratic_weighting: true

training:
  batch_size: 2048
  lr: 0.0008
  weight_decay: 0.01
  epochs: 200
  warmup_steps: 1000
  flow_loss_weight: 32.0
  lm_loss_weight: 0.0
  max_grad_norm: 1.0
  use_amp: true
  early_stopping_patience: 20
  save_every: 20

sampling:
  method: euler
  ode_steps: 20
  sampling_steps: 50
  use_ddim: true
  temperature: 1.0

data:
  # The best source experiment used the complete data range. This value is
  # read by the peptide training workflow unless --threshold overrides it.
  missing_threshold: 15
  train_ratio: 0.9
  val_ratio: 0.05
  num_workers: 4
  # Optional explicit splits. Leave null to split from --data_path.
  val_path: null
  test_path: null