知识卡片
对齐监督先学习示范响应
典型RLHF先用多样提示及人工示范回答监督微调,让模型获得初始指令响应行为。示范数据不是偏好排序数据;覆盖偏窄或标准不一致会影响起点,后续奖励优化不能被假定能自动修复所有数据缺口。
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
- · 对齐训练与推理
原书出处 · 1 处
- · 查看原书依据
Since prompting LLMs to be nice doesn’t work, we will need to tune the model in some way. Supervised fine-tuning (discussed in Chapter 6) on alignment datasets is an option. However, techniques like reinforcement learning have seen more success, which we will describe next in this section.
参考来源
- 位置:《Designing Large Language Model Applications》第08章《对齐训练与推理》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch08.html)
- 结论依据:摘录中关于对齐监督先学习示范响应的定义、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:Since prompting LLMs to be nice doesn’t work, we will need to tune the model in some way. Supervised fine-tuning (discussed in Chapter 6) on alignment datasets is an option. However, techniques like reinforcement learning have seen more success, which we will describe next in this section.