知识卡片
弱监督可唤起强模型已有表示
较弱教师给更强预训练学生提供弱标签,学生在某些任务可超过教师。其前提与预训练表示、训练目标及数据有关,不能推出任何低质量标签都能带来强能力,也不意味着小教师教会了自己完全不懂的所有内容。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 推理优化
原书出处 · 1 处
- · 查看原书依据
Weak-to-Strong Generalization Burns et al. from OpenAI uncovered a phenomenon called weak-to-strong generalization. In this setting, the teacher model is smaller/weaker than the student model. The small teacher model is fine-tuned using distillation data. The held-out portion of the distillation data is queried to the teacher model to generate outputs. These are called weak labels. The weak labels are then used to fine-tune a much
参考来源
- 位置:《Designing Large Language Model Applications》第09章《推理优化》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch09.html)
- 结论依据:摘录中关于弱监督可唤起强模型已有表示的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:Weak-to-Strong Generalization Burns et al. from OpenAI uncovered a phenomenon called weak-to-strong generalization. In this setting, the teacher model is smaller/weaker than the student model. The small teacher model is fine-tuned using distillation data. The held-out portion of the distillation data is queried to the teacher model to generate outputs. These are called weak labels. The weak labels are then used to fine-tune a much