← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8671.a · 生成模型:GPT-6

软提示适配只训练输入向量

理解等级:中级

Prompt tuning在输入嵌入层加入可学习软向量,不在每层独立增加前缀参数。向量可随机或由文本初始化,训练后不必对应可读词句;参数少也不等于任意小模型和任务都容易训练。

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

      Prompt tuning is a simplified version of prefix-tuning. Unlike prefix tuning, there are no prefix parameters at each layer. Figure 7-4 shows how prompt-tuning occurs in the Transformer. Figure 7-4. Prompt-tuning The adapters library provides a built-in configuration for prompt tuning: from adapters import PromptTuningConfig adapter_config = PromptTuningConfig() model.add_adapter("prompt_tuning", config=adapter_config) Some relevant configuration parameters for prompt tuning include: prompt_length The length of the prompt tokens;

    参考来源

    - 位置:《Designing Large Language Model Applications》第07章《进阶微调技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch07.html) - 结论依据:摘录中关于软提示适配只训练输入向量的定义、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:Prompt tuning is a simplified version of prefix-tuning. Unlike prefix tuning, there are no prefix parameters at each layer. Figure 7-4 shows how prompt-tuning occurs in the Transformer. Figure 7-4. Prompt-tuning The adapters library provides a built-in configuration for prompt tuning: from adapters import PromptTuningConfig adapter_config = PromptTuningConfig() model.add_adapter("prompt_tuning", config=adapter_config) Some relevant configuration parameters for prompt tuning include: prompt_length The length of the prompt tokens;