← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8705.a · 生成模型:GPT-6

长尾事实记忆缺少确定阈值

理解等级:中级

训练中事实出现较少时,模型可能不易稳定回答;重复、顺序和规模都影响记忆。但某研究的出现次数与准确率关系不是普适定律,不能推导任意事实需要固定次数或模型必须达到某个夸张参数量。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      LLMs Struggle with the Long-Tail LLMs typically need a lot of samples to memorize a fact. The memorization ability is probabilistic, so we cannot predict the exact number of samples the LLM needs to see during training for it to memorize. This sample-inefficiency means that the LLM will struggle to answer questions about entities and concepts that rarely occur in the training data. As an

    参考来源

    - 位置:《Designing Large Language Model Applications》第12章《检索增强生成》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch12.html) - 结论依据:摘录中关于长尾事实记忆缺少确定阈值的定义、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:LLMs Struggle with the Long-Tail LLMs typically need a lot of samples to memorize a fact. The memorization ability is probabilistic, so we cannot predict the exact number of samples the LLM needs to see during training for it to memorize. This sample-inefficiency means that the LLM will struggle to answer questions about entities and concepts that rarely occur in the training data. As an