知识卡片

普通读书笔记卡 · 6627 · 生成模型:GPT-6

稀有词元可能缺乏训练

理解等级:中级

分词器词表来自一批数据,而模型预训练内容可能不同,导致某些token几乎没有学习信号。异常输出可从这种失配排查,但不能仅凭token看起来古怪就确认其来源或宣布所有输入均有问题。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      The same tokenizer was used in GPT-3 models, but the pre-training dataset of the model had changed, so it didn’t include many or any references to SolidMagiGoldkarp. So now a token existed for SolidMagiGoldkarp but there was no signal in the pre-training dataset to learn from. This leads to some anomalous and hilarious behavior in GPT-3. These tokens are called glitch tokens or undertrained tokens.

    参考来源

    - 位置:《Designing Large Language Model Applications》第03章《词表与分词》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch03.html) - 结论依据:摘录中关于稀有词元可能缺乏训练的定义、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:The same tokenizer was used in GPT-3 models, but the pre-training dataset of the model had changed, so it didn’t include many or any references to SolidMagiGoldkarp. So now a token existed for SolidMagiGoldkarp but there was no signal in the pre-training dataset to learn from. This leads to some anomalous and hilarious behavior in GPT-3. These tokens are called glitch tokens or undertrained tokens.