Huggingface · Community Articles· purvansh joshi·· 7 小时前AI 评分58
SafetyPulse：通过读取模型拒答电路监测微调中的安全漂移
SafetyPulse: Catching Safety Drift During Fine-Tuning by Reading the Model's Refusal Circuit
AI 导读
作者提出 SafetyPulse，用电路完整性分数（CIS）在微调每个 checkpoint 上监测安全漂移。
来源：Huggingface · Community Articles · huggingface.co