知识卡片
单条最小请求与崩溃源
给模型喂“一条最小请求”,看能否返回任何结果——刻意用单样本单请求起步,风险最小、失败也最好排查。常见崩溃源有:训练栈与服务栈平台版本不一致;特征代码两套系统不同步(训练-服务偏斜,如服务侧仍用过期词表);模型文件写入出错或把 NaN 落盘而损坏;开发者加了新特征却忘了在服务栈接上管线;以及结果越界——该返回 0 到 1 间概率却给出 0.0 甚至 -0.2,该返回 100 类之一却返回 101。这些都属有效性而非精度问题。
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
- Evaluating Model Validity and Quality · 查看原书依据
Can the model serve a result without crashing the infrastructure? Again, this seems like a straightforward requirement: if we feed the model one minimal request, does it give us a result back of any kind, or does the system fail? Note that we say “one minimal request” as opposed to many requests intentionally, because these sorts of tests are often best done with single examples and single requests to start. This both minimize
参考来源
- 位置:《Reliable Machine Learning》第05章《Evaluating Model Validity and Quality》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch05.xhtml)
- 结论依据:原文明确阐述该主题,并给出定义、实践或反例。
- 原始内容:Can the model serve a result without crashing the infrastructure? Again, this seems like a straightforward requirement: if we feed the model one minimal request, does it give us a result back of any kind, or does the system fail? Note that we say “one minimal request” as opposed to many requests intentionally, because these sorts of tests are often best done with single examples and single requests to start. This both minimize