知识卡片
词元概率不是事实置信度
温度变换后的概率表示当前上下文和解码规则下选择某 token 的分布,不能直接解释为这个 token 所表达事实为真的概率。温度可在不改变知识和权重的情况下大幅改数值,更说明它不是经校准的事实可信度。
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
can confirm that 0.0003 is indeed the largest value: print("Highest probability:", max(next_token_probas.squeeze(0))) We can interpret this score as the model's confidence. This means that the model is more confident in " Berlin" (19,846) as the next token than in other tokens. The reason the value is so small is that we are using a large temperature, which makes it easier to plot this value next
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:摘录中关于词元概率不是事实置信度的定义、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:can confirm that 0.0003 is indeed the largest value: print("Highest probability:", max(next_token_probas.squeeze(0))) We can interpret this score as the model's confidence. This means that the model is more confident in " Berlin" (19,846) as the next token than in other tokens. The reason the value is so small is that we are using a large temperature, which makes it easier to plot this value next