知识卡片
奖励模型拟合相对人类偏好
对同一提示生成多个响应,再由人选择或排序,训练奖励模型预测偏好。标量奖励是标注群体及数据分布的代理,不能当作客观真理;离开训练分布后须检查偏置及奖励被利用的风险。
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
2. Reward modeling In this step, a diverse set of prompts is queried to the language model and multiple generations (responses) are extracted for each prompt. Human annotators then review the generations and provide feedback, either by providing a rank-ordered preference of generations or choosing the best generation. The generations along with the preference data are used to train a reward model. The reward model
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:摘录中关于奖励模型拟合相对人类偏好的定义、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:2. Reward modeling In this step, a diverse set of prompts is queried to the language model and multiple generations (responses) are extracted for each prompt. Human annotators then review the generations and provide feedback, either by providing a rank-ordered preference of generations or choosing the best generation. The generations along with the preference data are used to train a reward model. The reward model