prepare:
  test_size: 0.2
  random_state: 42
  max_samples: 5000       # REDUCED from 5000 to 500 for fast CI verification

train:
  model_name: "distilbert-base-uncased"
  num_epochs: 2
  batch_size: 16
  learning_rate: 2e-5
  warmup_steps: 100
  weight_decay: 0.01
  max_length: 256        # max token length per review