知识卡片
策略优化需要奖励与偏移约束
策略生成回答并获得奖励模型评分,同时以参考策略的概率分布约束偏移,再用PPO等方法更新。KL应在明确定义的条件token分布或序列分布上估计,不是把两个log-prob数字直接当两个分布;参考策略常取固定SFT模型,具体实现必须核实。
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
3. Proximal policy optimization (PPO) Finally, the reward model is used to optimize the pre-trained model that was fine-tuned in the first step. This is typically performed using an algorithm called PPO. The process of training using PPO is as follows: The language model generates a response or a continuation of a prompt. The reward model takes the query and response and outputs a scalar
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:摘录中关于策略优化需要奖励与偏移约束的定义、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:3. Proximal policy optimization (PPO) Finally, the reward model is used to optimize the pre-trained model that was fine-tuned in the first step. This is typically performed using an algorithm called PPO. The process of training using PPO is as follows: The language model generates a response or a continuation of a prompt. The reward model takes the query and response and outputs a scalar