知识卡片
生成质量需要样本评估
开放生成常有多个可接受答案,不能只用字符串完全一致检查质量。应选代表性输入集合,定义任务质量标准并比较版本,同时观察关键子群与退化案例。模型可辅助评分,但其判断需校准和抽样人工检查;总体均值提高不能豁免安全、结构和工具行为的确定性断言。
这张卡片看不懂?
这张卡片的反馈已满 10 条,暂时无法继续提交。
已有反馈 · 0 / 10 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- Chapter 15. AI Inference and Serving · 查看原书依据
Instead of looking for exact matches, we are looking for our tests to evaluate if a response is “good enough.” Once we shift our minds from “correct” to “good enough,” we realize that evaluating a single input is no longer acceptable. As the model changes, the response to
参考来源
- 位置:《Designing Distributed Systems, 2nd Edition》第15章《Chapter 15. AI Inference and Serving》(来源项目 book-sources;源文件:extracted/Designing Distributed Systems, 2nd Edition/_epub-src/OEBPS/ch15.html)
- 结论依据:原文从单例正确性转到多输入质量评估,并讨论LLM评委与用户反馈;明确不能替代全部测试。
- 原始内容:Instead of looking for exact matches, we are looking for our tests to evaluate if a response is “good enough.” Once we shift our minds from “correct” to “good enough,” we realize that evaluating a single input is no longer acceptable. As the model changes, the response to