知识卡片
检索也可参与预训练目标
REALM把知识检索和编码预测结合,以掩码恢复收益训练检索能力,并允许空文档表达无需外部知识。训练中的检索并非普通离散top-k整体天然可微,需要具体潜变量目标和近似;不能只凭组件是神经网络就宣称端到端任意操作可微。
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
- · 检索增强生成
原书出处 · 1 处
- · 查看原书依据
RAG for Model Training So far, all the RAG applications we have explored are applied during LLM inference. Can we use RAG during model pre-training and fine-tuning as well? Yes, we can! This is an underrated area of study, and I expect to see more LLMs leveraging this in the coming years. Let’s look at an example in detail. Retrieval-Augmented Language Model (REALM) is one
参考来源
- 位置:《Designing Large Language Model Applications》第12章《检索增强生成》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch12.html)
- 结论依据:摘录中关于检索也可参与预训练目标的定义、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:RAG for Model Training So far, all the RAG applications we have explored are applied during LLM inference. Can we use RAG during model pre-training and fine-tuning as well? Yes, we can! This is an underrated area of study, and I expect to see more LLMs leveraging this in the coming years. Let’s look at an example in detail. Retrieval-Augmented Language Model (REALM) is one