model:
  base_model: Qwen/Qwen3-0.6B
  max_length: 512
  pooling: last_token
  dropout: 0.05
  freeze_backbone: true
  dtype: auto

training:
  seed: 43
  epochs: 3
  batch_size: 8
  gradient_accumulation_steps: 1
  learning_rate: 0.0002
  weight_decay: 0.0
  max_grad_norm: 1.0
  estimator: reinforce
  rollouts_per_prompt: 8
  entropy_coefficient: 0.01
  normalize_advantage: false
  log_every: 10