wangyu
7d4345c29d
Two-module pipeline for AI companion safety: - Module B: context-aware risk detector with CrossAttention fusion - Module C: PPO-based adaptive intervention policy Includes CompanionRisk Taxonomy (10 primary + 14 fine-grained labels), dataset generation/annotation pipeline, training scripts, and eval suite. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
47 lines
822 B
YAML
47 lines
822 B
YAML
detector:
|
|
checkpoint: "checkpoints/detector/best.pt"
|
|
model_name: "hfl/chinese-macbert-large"
|
|
hidden_size: 1024
|
|
|
|
agent:
|
|
state_hidden: 256
|
|
dropout: 0.1
|
|
|
|
reward:
|
|
w1: 2.0 # safety gain for correct intervention
|
|
w2: 3.0 # false negative penalty
|
|
w3: 4.0 # crisis bonus for R1
|
|
w4: 1.5 # over-refusal penalty
|
|
w5: 0.5 # UX cost
|
|
|
|
behavior_cloning:
|
|
enabled: true
|
|
epochs: 5
|
|
lr: 1e-3
|
|
|
|
ppo:
|
|
total_timesteps: 200000
|
|
n_rollout_steps: 2048
|
|
n_epochs: 4
|
|
batch_size: 64
|
|
lr: 3e-4
|
|
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:
|
|
n_envs: 1
|
|
max_turns: 20
|
|
|
|
logging:
|
|
project: "CompanionGuard-RL"
|
|
run_name: "intervention-ppo"
|
|
use_wandb: true
|
|
|
|
output:
|
|
checkpoint_dir: "checkpoints/intervention"
|
|
save_interval: 10000
|