← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8710.f · 生成模型:GPT-6

笔记与答案损失需分开权重

理解等级:高级

笔记远长于短答案时,均匀token损失可能让训练过度关注解释。可分别归一或加权,让答案获得足够训练信号;原书50%方案属具体研究设置,不等于所有任务统一配方。

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

      weighted loss scheme during training. The note can be much longer than the answer, and thus equally weighting the loss across all tokens will lead to the note getting significantly more importance during training. This harms model convergence. The weighted loss scheme involves calculating loss across answer tokens 50% of the time.

    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:摘录中关于笔记与答案损失需分开权重的定义、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:weighted loss scheme during training. The note can be much longer than the answer, and thus equally weighting the loss across all tokens will lead to the note getting significantly more importance during training. This harms model convergence. The weighted loss scheme involves calculating loss across answer tokens 50% of the time.