知识卡片
全语言目标只看前缀预测后继
标准自回归目标在每个有效目标位置仅使用此前上下文,训练时可凭因果遮罩并行计算多个位置损失。它不同于逐步生成,也不是把真实未来token暴露给当前位置后仍称下一token预测。
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
- · 架构与学习目标
原书出处 · 1 处
- · 查看原书依据
Full Language Modeling Figure 4-7 shows the canonical FLM objective at work. Figure 4-7. Full language modeling This is the canonical language modeling objective of learning to predict the next token in a sequence and currently the simplest and most common training objective, used by GPT-4 and a vast number of open source models. The loss is computed for every token the model sees, i.e.,
参考来源
- 位置:《Designing Large Language Model Applications》第04章《架构与学习目标》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch04.html)
- 结论依据:摘录中关于全语言目标只看前缀预测后继的定义、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:Full Language Modeling Figure 4-7 shows the canonical FLM objective at work. Figure 4-7. Full language modeling This is the canonical language modeling objective of learning to predict the next token in a sequence and currently the simplest and most common training objective, used by GPT-4 and a vast number of open source models. The loss is computed for every token the model sees, i.e.,