← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8663 · 生成模型:GPT-6

校准要求频率匹配概率

理解等级:高级

在明确预测事件与分组条件下,给出约0.9置信度的预测应有相近经验正确频率。单纯概率与准确率相关并不足够;校准还可能因领域转移改变,不能把某token的高概率直接当整段回答可信度。

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

      Calibration is an underappreciated topic in deep learning. A model is said to be well-calibrated if there is a correlation between its output probability values and task accuracy. For example, consider a task that classifies a sentence as being abusive or not. If the model is well-calibrated, then among all examples for which the model produces an output probability of 0.9, 90% of them would

    参考来源

    - 位置:《Designing Large Language Model Applications》第06章《微调》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch06.html) - 结论依据:摘录中关于校准要求频率匹配概率的定义、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:Calibration is an underappreciated topic in deep learning. A model is said to be well-calibrated if there is a correlation between its output probability values and task accuracy. For example, consider a task that classifies a sentence as being abusive or not. If the model is well-calibrated, then among all examples for which the model produces an output probability of 0.9, 90% of them would