model:
  name: "hfl/chinese-macbert-large"
  hidden_size: 1024
  num_heads: 8
  dropout: 0.1
  use_lora: false

data:
  train_path: "data/processed/CompanionRisk-Bench/train.jsonl"
  val_path:   "data/processed/CompanionRisk-Bench/dev.jsonl"
  test_path:  "data/processed/CompanionRisk-Bench/test.jsonl"
  max_persona_len:   128
  max_context_len:   512
  max_response_len:  256
  max_history_turns: 5
  num_workers: 0            # 0 for Windows (avoids multiprocessing issues); set to 4 on Linux

training:
  epochs: 10
  per_gpu_batch_size: 16    # single GPU: 16; 4 GPUs: use 32 (effective 128)
  gradient_accumulation_steps: 2   # effective_batch = 16 × 1 GPU × 2 = 32
  lr: 2e-5
  warmup_steps: 100
  weight_decay: 0.01
  gradient_clip: 1.0
  eval_steps: 100           # global steps between validation runs
  mixed_precision: "bf16"   # RTX 5090: bf16; RTX 30xx/40xx: fp16; CPU-only: no
  seed: 42

loss_weights:
  binary:  1.0
  level:   1.0
  primary: 1.0
  fine:    1.0     # 下次训练建议提升到 2.0，配合 fine_training 选项

# Fine-grained label training options（下次训练时开启，当前 best.pt 不受影响）
fine_training:
  use_pos_weight: false   # 改为 true 开启 pos_weight（下次训练）
  risky_only:     false   # 改为 true 开启（下次训练）

evaluation:
  binary_threshold: 0.5
  fine_threshold:   0.4

logging:
  project:   "CompanionGuard-RL"
  run_name:  "detector-macbert-v1"
  use_wandb: false          # set true if wandb is configured

output:
  checkpoint_dir: "checkpoints/detector"