面向 NLP 任务的大模型 Prompt 设计

news2024/9/23 9:30:36

很久之前，我们介绍到，prompt是影响下游任务的关键所在，当我们在应用chatgpt进行nlp任务落地时，如何选择合适的prompt，对于SFT以及推理环节尤为重要。

不过，硬想不是办法，我们可以充分参考开源的一些已有工作，幸运的是，这类工作已然存在。

因此，本文主要介绍longbench、LooGLE、pclue以及firefly自然语言处理任务prompt以及PromptSource英文常用评测任务prompt生成工具包。

一、其他一些关于NLP任务的代表prompt

最近我们在看长文本说的一些评估数据集，而对于评估来说，如何针对不同的任务，设定相应的prompt，最为重要。下面介绍longbench、LooGLE、pclue以及firefly自然语言处理任务prompt。

1、longbench长文本prompt

地址：https://github.com/THUDM/LongBench

技术交流群

前沿技术资讯、算法交流、求职内推、算法竞赛、面试交流(校招、社招、实习)等、与 10000+来自港科大、北大、清华、中科院、CMU、腾讯、百度等名校名企开发者互动交流~

建了技术交流群&星球！想要资料、进交流群的同学，可以直接加微信号：mlc2060。加的时候备注一下：研究方向 +学校/公司，即可。然后就可以拉你进群了。

方式①、添加微信号：mlc2060，备注：大模型资料 or 技术交流
方式②、微信搜索公众号：机器学习社区，后台回复：大模型资料 or 技术交流

在这里插入图片描述

2、LooGLE长文本评测prompt

地址： https://github.com/bigai-nlco/LooGLE

4、Pclue任务评测prompt

地址： https://github.com/CLUEbenchmark/pCLUE

4、firefly自然语言处理任务prompt

地址：https://huggingface.co/datasets/YeungNLP/firefly-train-1.1M/viewer/default/train?row=3

二、PromptSource英文NLP prompt生成工具

PromptSource是一个用于创建、共享和使用自然语言提示的工具包，截至2022年1月20日，P3中有约2000个prompt，涵盖170多个英语数据集。

项目地址：https://github.com/bigscience-workshop/promptsource

1、storycloze的prompt

templates:
  1a4946f9-a0e2-4fbb-aee8-b26ead2cf6b8: !Template
    answer_choices: '{{sentence_quiz1}} ||| {{sentence_quiz2}}'
    id: 1a4946f9-a0e2-4fbb-aee8-b26ead2cf6b8
    jinja: '{{input_sentence_1}} {{input_sentence_2}} {{input_sentence_3}} {{input_sentence_4}}
      What is a possible continuation for the story given the following options ?
      - {{answer_choices | join("\n- ")}} ||| {{answer_choices[answer_right_ending
      -1]}}'
    metadata: !TemplateMetadata
      choices_in_prompt: true
      languages:
      - en
      metrics:
      - Accuracy
      original_task: true
    name: Answer Given options
    reference: ''
  1a9d53bc-eb77-4e7c-af6e-3d15b79d6cf1: !Template
    answer_choices: '{{sentence_quiz1}} ||| {{sentence_quiz2}}'
    id: 1a9d53bc-eb77-4e7c-af6e-3d15b79d6cf1
    jinja: "Read the following story :\n\n{{input_sentence_1}}\n{{input_sentence_2}}\n\
      {{input_sentence_3}}\n{{input_sentence_4}}\n\nChoose a possible ending for the\
      \ previous story from the following options: \n- {{answer_choices | join(\"\\\
      n- \")}}\n|||\n\n{{answer_choices[answer_right_ending -1]}}"
    metadata: !TemplateMetadata
      choices_in_prompt: true
      languages:
      - en
      metrics:
      - Accuracy
      original_task: true
    name: Choose Story Ending
    reference: ''

2、Squad任务的prompt

templates:
  3d85b5b0-51db-4d72-8ead-d0b3654025ee: !Template
    answer_choices: null
    id: 3d85b5b0-51db-4d72-8ead-d0b3654025ee
    jinja: 'Refer to the passage below and answer the following question:
      Passage: {{context}}
      Question: {{question}}
      |||
      {{answers["text"][0]}}'
    metadata: !TemplateMetadata
      choices_in_prompt: false
      languages:
      - en
      metrics:
      - Squad
      original_task: true
    name: answer_question_given_context
    reference: ''

3、MathQA任务的prompt

a313a5f8-53cd-4b76-abb6-fea2ac4e9ef4: !Template
    answer_choices: a ||| b ||| c ||| d ||| e
    id: a313a5f8-53cd-4b76-abb6-fea2ac4e9ef4
    jinja: "One of the five choices are correctly answers the math problem. Can you\
      \ choose the right one? \n\n{{options}}\n\nProblem: {{Problem}}\n|||\n{{correct}}"
    metadata: !TemplateMetadata
      choices_in_prompt: true
      languages:
      - en
      metrics:
      - Accuracy
      original_task: true
    name: first_choice_then_problem
    reference: First give the list of choices and then describe the problem
  a3c2ec72-4af5-42aa-9e8e-ef475fa7c039: !Template
    answer_choices: general ||| physics ||| gain ||| geometry ||| probability |||
      other
    id: a3c2ec72-4af5-42aa-9e8e-ef475fa7c039
    jinja: "Given the problem below, in what category would you classify it?\n===\n\
      {{Problem}} \n\nCategories:\n{{answer_choices | join(\"\\n\")}}\n|||\n{{category}}\n"
    metadata: !TemplateMetadata
      choices_in_prompt: true
      languages:
      - en
      metrics:
      - Accuracy
      original_task: false
    name: problem_set_type
    reference: The template asks to generate the category of the problem set

4、使用方式

# Load an example from the datasets ag_news
>>> from datasets import load_dataset
>>> dataset = load_dataset("ag_news", split="train")
>>> example = dataset[1]

# Load prompts for this dataset
>>> from promptsource.templates import DatasetTemplates
>>> ag_news_prompts = DatasetTemplates('ag_news')

# Print all the prompts available for this dataset. The keys of the dict are the uuids the uniquely identify each of the prompt, and the values are instances of `Template` which wraps prompts
>>> print(ag_news_prompts.templates)
{'24e44a81-a18a-42dd-a71c-5b31b2d2cb39': <promptsource.templates.Template object at 0x7fa7aeb20350>, '8fdc1056-1029-41a1-9c67-354fc2b8ceaf': <promptsource.templates.Template object at 0x7fa7aeb17c10>, '918267e0-af68-4117-892d-2dbe66a58ce9': <promptsource.templates.Template object at 0x7fa7ac7a2310>, '9345df33-4f23-4944-a33c-eef94e626862': <promptsource.templates.Template object at 0x7fa7ac7a2050>, '98534347-fff7-4c39-a795-4e69a44791f7': <promptsource.templates.Template object at 0x7fa7ac7a1310>, 'b401b0ee-6ffe-4a91-8e15-77ee073cd858': <promptsource.templates.Template object at 0x7fa7ac7a12d0>, 'cb355f33-7e8c-4455-a72b-48d315bd4f60': <promptsource.templates.Template object at 0x7fa7ac7a1110>}

# Select a prompt by its name
>>> prompt = ag_news_prompts["classify_question_first"]

# Apply the prompt to the example
>>> result = prompt.apply(example)
>>> print("INPUT: ", result[0])
INPUT:  What label best describes this news article?
Carlyle Looks Toward Commercial Aerospace (Reuters) Reuters - Private investment firm Carlyle Group,\which has a reputation for making well-timed and occasionally\controversial plays in the defense industry, has quietly placed\its bets on another part of the market.
>>> print("TARGET: ", result[1])
TARGET:  Business