知识卡片
压缩取舍需要端到端检索评估
裁剪、二值、整数和PQ影响不同计算及误差,某数据冗余可掩盖部分召回下降。仍需检查稀有事实、过滤后小集合和困难查询;某模型二值优于int8的经验不构成压缩越强越好的规律。
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
- · 表示学习与嵌入
原书出处 · 1 处
- · 查看原书依据
Choose your quantization technique by determining your relative product priorities for criteria like cost, performance, and speed. Tip Optimizing embeddings for storage come with a performance hit. However, if there is plenty of redundancy in the document corpus, answers to typical user queries might be found in several documents, and hence the user may not feel this performance drop. Exercise Download the Wikipedia embeddings encoded
参考来源
- 位置:《Designing Large Language Model Applications》第11章《表示学习与嵌入》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch11.html)
- 结论依据:摘录中关于压缩取舍需要端到端检索评估的定义、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:Choose your quantization technique by determining your relative product priorities for criteria like cost, performance, and speed. Tip Optimizing embeddings for storage come with a performance hit. However, if there is plenty of redundancy in the document corpus, answers to typical user queries might be found in several documents, and hence the user may not feel this performance drop. Exercise Download the Wikipedia embeddings encoded