知识卡片
对齐与事实性需要多层验收
结构图:
flowchart TD
D[任务与行为标准] --> F[人工示范与偏好反馈]
F --> T[监督与策略优化]
T --> G[生成回答或候选]
C[可追溯上下文] --> G
G --> V[逐项证据与约束验证]
V --> R[预算内修订或选择]
R --> A[帮助性 诚实性 无害性及成本评估]
偏好、忠实性、外部事实与推导有效性分别检查;任何单项通过都不能替代其他维度。
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
In this chapter, we defined alignment training and why we need it. We ventured into techniques for alignment training such as reinforcement learning. We also learned about hallucinations and different techniques to mitigate them. Finally, we examined reasoning limitations of LLMs and new techniques like scaling up inference-time computation.
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:摘录中关于对齐与事实性需要多层验收的定义、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:In this chapter, we defined alignment training and why we need it. We ventured into techniques for alignment training such as reinforcement learning. We also learned about hallucinations and different techniques to mitigate them. Finally, we examined reasoning limitations of LLMs and new techniques like scaling up inference-time computation.