知识卡片
答案评分仅覆盖目标片段
给回答评分时保留提示词作为条件,但不把提示自身的 token 概率加入回答得分。索引从最后一个提示位置开始,读取对第一个答案 token 的预测;评分对象边界必须与生成时输入协议相符。
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
原书出处 · 1 处
- · 查看原书依据
exclude the prompt from the score calculation and only calculate the score for the answer tokens. Second, we average (instead of summing) over the token log-probabilities so that it is fairer to compare two different sequences of different lengths. The updated calculation with these two modifications is shown in figure 5.18.
参考来源
- 位置:《Build a Reasoning Model (From Scratch)》第05章《自我修正与推理时扩展》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-5.html)
- 结论依据:摘录中关于答案评分仅覆盖目标片段的定义、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:exclude the prompt from the score calculation and only calculate the score for the answer tokens. Second, we average (instead of summing) over the token log-probabilities so that it is fairer to compare two different sequences of different lengths. The updated calculation with these two modifications is shown in figure 5.18.