知识卡片

普通读书笔记卡 · 6610.f · 生成模型:GPT-6

多轮重复不是新增独立信息

理解等级:中级

重复训练可增加模型利用已有数据的机会,但同一样本的边际收益可能递减,并提高过拟合风险。原研究中的四五轮现象不是跨数据、模型和目标的安全阈值,应按验证表现与计算预算决定。

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

      Training for Multiple Epochs An epoch refers to the model being exposed to the complete training dataset during the training process. Earlier LLMs were trained on just one epoch or less, to protect against overfitting. Can we just increase the size of the training data by training the model for multiple epochs on the same dataset, i.e., the model sees the same dataset multiple times

    参考来源

    - 位置:《Designing Large Language Model Applications》第02章《预训练数据》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch02.html) - 结论依据:摘录中关于多轮重复不是新增独立信息的定义、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:Training for Multiple Epochs An epoch refers to the model being exposed to the complete training dataset during the training process. Earlier LLMs were trained on just one epoch or less, to protect against overfitting. Can we just increase the size of the training data by training the model for multiple epochs on the same dataset, i.e., the model sees the same dataset multiple times