chore: initial commit — unified project repo

Merged code repo (CompanionGuard-RL) into single project-level git.
Reorganized root: docs/, reference/, experiments/, tmp/active|archives/.
Gitignored: data/, checkpoints/, .venv, experiment logs, tmp/archives.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

code/configs/data_generation.yaml

@@ -0,0 +1,24 @@
+api:
+  type: "qwen"         # "qwen" or "openai"
+  model: "qwen-max"
+
+generation:
+  total_samples: 3000
+  safe_ratio: 0.25      # 25% safe (y_risk=0) samples
+  delay: 0.5            # seconds between API calls
+  max_retries: 3        # retry attempts per failed generation
+
+output:
+  raw_dir: "data/raw"
+  output_file: "data/raw/generated.jsonl"
+
+annotation:
+  judge_model: "qwen-max"
+  output_file: "data/processed/annotated.jsonl"
+  delay: 0.3
+
+split:
+  train: 0.8
+  val: 0.1
+  test: 0.1
+  seed: 42

code/configs/detector_config.yaml

@@ -0,0 +1,51 @@
+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"

code/configs/detector_config_server.yaml

@@ -0,0 +1,54 @@
+model:
+  name: "/root/siton-data-740d234e02d749f08fe5347b0c74c49f/zsy/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: 4            # Linux server: 4 workers; Windows: use 0
+
+training:
+  epochs: 10
+  per_gpu_batch_size: 16    # 4 GPUs × 16 × accum 2 = effective batch 128
+  gradient_accumulation_steps: 2
+  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 native bf16
+  seed: 42
+
+loss_weights:
+  binary:  1.0
+  level:   1.0
+  primary: 1.0
+  fine:    2.0     # ↑ 2.0: 加强细粒度标签损失权重（配合 fine_training 开启）
+
+# Fine-grained label training options（下次训练时开启，当前 best.pt 不受影响）
+# 两项均开启可显著改善 fine_macro_f1：
+#   use_pos_weight: 对 Romanticization/CoRumination 等稀有标签设置 ~25 倍正样本权重
+#   risky_only:     只在 y_risk=1 的样本上计算 fine loss，避免 safe 样本教模型预测全负
+fine_training:
+  use_pos_weight: true    # ✓ 开启：对稀有 fine 标签设置 pos_weight（max 30）
+  risky_only:     true    # ✓ 开启：只在 y_risk=1 样本上计算 fine loss
+
+evaluation:
+  binary_threshold: 0.5
+  fine_threshold:   0.4
+
+logging:
+  project:   "CompanionGuard-RL"
+  run_name:  "detector-macbert-4gpu"
+  use_wandb: false
+
+output:
+  checkpoint_dir: "checkpoints/detector"

code/configs/intervention_config.yaml

@@ -0,0 +1,49 @@
+detector:
+  checkpoint: "checkpoints/detector/best.pt"
+  # Server 2 path — update this when running on server 2
+  model_name: "/root/siton-data-740d234e02d749f08fe5347b0c74c49f/zsy/macbert-large"
+  hidden_size: 1024
+
+agent:
+  state_hidden: 256
+  dropout: 0.1
+
+# Stage 1: Behavior cloning warm-up
+behavior_cloning:
+  enabled: true
+  epochs: 5
+  per_gpu_batch_size: 256
+  lr: 0.001
+  mixed_precision: "bf16"
+
+# Stage 2: PPO runs on GPU-0 only
+ppo:
+  total_timesteps: 200000
+  n_rollout_steps: 2048
+  n_epochs: 4
+  batch_size: 256
+  lr: 0.0003
+  clip_eps: 0.2
+  entropy_coef: 0.01
+  value_coef: 0.5
+  max_grad_norm: 0.5
+  gamma: 0.99
+  gae_lambda: 0.95
+
+environment:
+  max_turns: 20
+
+evaluation:
+  binary_threshold: 0.5
+
+preprocessing:
+  per_gpu_batch_size: 64
+
+logging:
+  project:   "CompanionGuard-RL"
+  run_name:  "intervention-v5-1gpu"
+  use_wandb: false
+
+output:
+  checkpoint_dir: "checkpoints/intervention"
+  save_interval:  10000