知识卡片
事实核采样逐步收紧候选集
事实核采样令p_t=max(ω,p·λ^(t−1)),以衰减率逐步降低top-p阈值,并用下界及按句重置避免过度退化。这里收紧的是概率质量阈值,不是token数量必然单调减少;连贯段落越长也不必然只有更少合法续写。
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
- · 对齐训练与推理
原书出处 · 1 处
- · 查看原书依据
pt = max{ω, p × λ t−1 } where t refers to the generation step. There are three tunable parameters: Decay rate (λ) The p value of the algorithm is decayed by a decay rate at every step of the generation. Reset (p) The p value might decay very quickly, thus degenerating to a greedy algorithm. To prevent this, we can reset the p value
参考来源
- 位置:《Designing Large Language Model Applications》第08章《对齐训练与推理》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch08.html)
- 结论依据:摘录中关于事实核采样逐步收紧候选集的定义、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:pt = max{ω, p × λ t−1 } where t refers to the generation step. There are three tunable parameters: Decay rate (λ) The p value of the algorithm is decayed by a decay rate at every step of the generation. Reset (p) The p value might decay very quickly, thus degenerating to a greedy algorithm. To prevent this, we can reset the p value