https://www.hansking.cn/blog A posts created with Next.js and Tailwind.css en-us admin@hansking.cn (HANS) admin@hansking.cn (HANS) Mon, 13 Apr 2026 10:00:00 GMT https://www.hansking.cn/blog/2026/ModelDistillation-shi-zhan-ba-da-mo-xing-neng-li-ya-suo-xia-lai https://www.hansking.cn/blog/2026/ModelDistillation-shi-zhan-ba-da-mo-xing-neng-li-ya-suo-xia-lai [{"value":"Model Distillation 实战：把大模型能力压缩下来","url":"#model-distillation-实战把大模型能力压缩下来","depth":1},{"value":"一、先定义学生模型的任务边界","url":"#一先定义学生模型的任务边界","depth":2},{"value":"二、让教师输出稳定，而不是只追求聪明","url":"#二让教师输出稳定而不是只追求聪明","depth":2},{"value":"三、数据清洗比训练更重要","url":"#三数据清洗比训练更重要","depth":2},{"value":"1. 先做结构校验","url":"#1-先做结构校验","depth":3},{"value":"2. 去重和去污染","url":"#2-去重和去污染","depth":3},{"value":"3. 教师一致性检查","url":"#3-教师一致性检查","depth":3},{"value":"四、两种蒸馏目标","url":"#四两种蒸馏目标","depth":2},{"value":"1. 硬标签蒸馏","url":"#1-硬标签蒸馏","depth":3},{"value":"2. 软目标蒸馏","url":"#2-软目标蒸馏","depth":3},{"value":"五、一个最小训练骨架","url":"#五一个最小训练骨架","depth":2},{"value":"六、评测蒸馏是否值得","url":"#六评测蒸馏是否值得","depth":2},{"value":"七、把不确定性交给大模型","url":"#七把不确定性交给大模型","depth":2},{"value":"总结：蒸馏是在复制能力，也是在复制责任","url":"#总结蒸馏是在复制能力也是在复制责任","depth":2}] Mon, 13 Apr 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM知识蒸馏模型训练PyTorch