知识卡片
去污染针对评测信息泄漏
评测输入与答案进入训练,会使分数难以反映未见任务表现;只出现输入与连同标签出现也应区别。应记录数据版本和时间边界,检查整个训练与适配链,而非仅对某个预训练快照去重就宣称评测洁净。
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
- · 预训练数据
原书出处 · 1 处
- · 查看原书依据
Training set decontamination is a crucial data preprocessing step that helps improve LLM evaluations. A pre-training dataset is said to be contaminated if it contains data from the benchmark test sets used to evaluate its performance. Contamination can happen if the test datasets were constructed from web text, or if the dataset was uploaded on the web after creation. There are two types of contamination:1
参考来源
- 位置:《Designing Large Language Model Applications》第02章《预训练数据》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch02.html)
- 结论依据:摘录中关于去污染针对评测信息泄漏的定义、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:Training set decontamination is a crucial data preprocessing step that helps improve LLM evaluations. A pre-training dataset is said to be contaminated if it contains data from the benchmark test sets used to evaluate its performance. Contamination can happen if the test datasets were constructed from web text, or if the dataset was uploaded on the web after creation. There are two types of contamination:1