LLM Prompt Engineering - Few-Shot Prompting 少量樣本提示

back-to:: LLM Prompt Engineering MOC

少量樣本提示可以用作一種技術，以實現上下文學習，有點像是讓模型學會「造樣造句」

最早是 Brown 等人在 2020 年發現，其中一個範例是這樣的：

A "whatpu" is a small, furry animal native to Tanzania. An example of a sentence that uses  the word whatpu is:
We were traveling in Africa and we saw these very cute whatpus.

To do a "farduddle" means to jump up and down really fast. An example of a sentence that uses the word farduddle is:

一隻「whatpu」是一種原產於坦桑尼亞的小型毛茸茸的動物。
使用該詞的一個例子句子是： 我們正在非洲旅行，看到了這些非常可愛的 whatpu。

「farduddle」意味著快速地跳上跳下。
使用該詞的一個例子句子是：

GPT 會回答 : The kids were so excited that they started to farduddle when they heard they were going to the amusement park.

其實這兩個字根本不存在，但是 LLM 藉由上下文好像學會了這些範例

Few-Shot Chain of Thought

在數組中的奇數，全部相加起來會等於偶數，對或錯 ?

The odd numbers in this group add up to an even number: 4, 8, 9, 15, 12, 2, 1.
A: The answer is False.

The odd numbers in this group add up to an even number: 17, 10, 19, 4, 8, 12, 24.
A: The answer is True.

The odd numbers in this group add up to an even number: 16, 11, 14, 4, 8, 13, 24.
A: The answer is True.

The odd numbers in this group add up to an even number: 17, 9, 10, 12, 13, 4, 2.
A: The answer is False.

The odd numbers in this group add up to an even number: 15, 32, 5, 13, 82, 7, 1.
A:

這段文字雖然提供範例，但是答案太過簡短，LLM 無法得到答案跟題目的關係，如果這時我們逐步加入「思考過程」，這樣 LLM 會更能掌握思考過程的邏輯，==這又叫 Few-Shot Chain of Thought==

透過向 LLM 展示少量的樣本，並解釋推理過程，LLM 在回答時也會顯示推理過程，這樣往往能引導出更正確的答案

Few-Shot Chain of Thought

Sewon Min  等人在  2022 年的研究  表明，CoT 有以下特性

1. the label space and the distribution of the input text specified by the demonstrations are both key (regardless of whether the labels are correct for individual inputs

就算範例問題的答案，都是不正確的也沒關係，標籤（答案）在輸入 input 之中的分布很關鍵，例如在上個例子之中，True / False 的比例是一半一半

2. the format you use also plays a key role in performance, even if you just use random labels, this is much better than no labels at all.

就算範例問題的答案都是錯的、隨機的，能給出一些思考邏輯的過程和格式，就已經能提昇性能，但是 CoT 要在大於等於 100B 參數的模型才會生效