← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8764.c · 生成模型:GPT-6

随机控制只能减少部分变异

理解等级:高级

固定seed、解码和输入可减少变异,硬件、并行计算和服务版本仍可能影响结果。若无法完全控制,应重复测量统计结果及不确定性;temperature=0不是所有系统输出绝对可复现的保证。

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

      set random seeds or setting temperature to zero and using greedy sampling. With all inputs held constant, we expect repeatable outputs, even with generative models. Some lower-level optimizations, such as parallelism in compute or model quantization, may still produce nondeterministic behavior. But, in general, it is still possible to make execution deterministic through disabling these optimizations, if this is important enough to test. Even if

    参考来源

    - 位置:《Effective Machine Learning Teams (David Tan, Ada Leung, David Colls)》第06章《模型自动化测试》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Effective Machine Learning Teams (David Tan, Ada Leung, David Colls)/_epub-src/OEBPS/ch06.html) - 结论依据:摘录中关于随机控制只能减少部分变异的定义、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:set random seeds or setting temperature to zero and using greedy sampling. With all inputs held constant, we expect repeatable outputs, even with generative models. Some lower-level optimizations, such as parallelism in compute or model quantization, may still produce nondeterministic behavior. But, in general, it is still possible to make execution deterministic through disabling these optimizations, if this is important enough to test. Even if