知识卡片
按分布抽样不保证本次最优
多项抽样按权重选择词表项,低概率项仍可能出现;很多次独立抽样的频率才趋向给定分布。一个样本恰好错误不证明采样器出错,而频率接近分布也不能证明文本事实正确。
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
- · 推理时计算扩展
原书出处 · 1 处
- · 查看原书依据
proportional to their probabilities. In other words, vocabulary indices with higher probabilities are more likely to be sampled. If we repeated the sampling a very large number of times, we would sample the vocabulary index corresponding to the token " Berlin" with 0.03% probability, given a temperature of 5.
参考来源
- 位置:《Build a Reasoning Model (From Scratch)》第04章《推理时计算扩展》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-4.html)
- 结论依据:摘录中关于按分布抽样不保证本次最优的定义、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:proportional to their probabilities. In other words, vocabulary indices with higher probabilities are more likely to be sampled. If we repeated the sampling a very large number of times, we would sample the vocabulary index corresponding to the token " Berlin" with 0.03% probability, given a temperature of 5.