← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8679.c · 生成模型:GPT-6

幻觉分类指导解码先核标签

理解等级:高级

可为多个候选片段提取回看特征,由分类器打分后选择。必须检查输出表示事实性还是幻觉概率,再确定取最大或最小;候选生成、特征提取和比较会增加延迟,分类准确率也须在目标分布检验。

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

      classifier model can also be employed to reduce hallucinations during generation. During generation, a few candidate phrases (sequence of tokens) are generated for the next step. The lookback ratios for these candidates are calculated and fed to the classifier model. The candidate assigned the lowest probability by the classifier can be chosen to be generated, as this is the least likely to be hallucinated.

    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:摘录中关于幻觉分类指导解码先核标签的定义、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:classifier model can also be employed to reduce hallucinations during generation. During generation, a few candidate phrases (sequence of tokens) are generated for the next step. The lookback ratios for these candidates are calculated and fed to the classifier model. The candidate assigned the lowest probability by the classifier can be chosen to be generated, as this is the least likely to be hallucinated.