达芬奇数学不好：使用 NodeJs 和 OpenAI v4 微调 ChatGPT 模型

本文重点介绍使用 OpenAI 的Node.js库构建一个用于训练数学达芬奇模型的CLI 。

简而言之

• “脚手架”我们的图书馆。
• 编写一组函数来包装 OpenAI 的 API 调用。
• 构建一个简单的 CLI 来调用函数。
• 证明 ChatGPT（通常）擅长数学。
• 证明达芬奇（通常）数学不好。
• 构建一个简单的微调数据集用于教授达芬奇数学。
• 上传“简单微调数据集”。
• 将“简单的微调数据集”变成简单的微调模型。
• 证明我们的微调教授了达芬奇数学。

脚手架

cd ~/Dev/YourRootFolderForPersonalStuff/ mdkir davinci-is-bad-at-maths cd davinci-is-bad-at-maths npm i dotenv openai npm i prettier -D touch .env touch goodAtMathsDatasetBuilder.js touch openAI.js mkdir bin touch bin/cli.js

package.json

...可以很简单，像这样：

 { "description": "Experiments using OpenAI's API NodeJs v4 library", "name": "davinci-is-bad-at-maths", "private": true, "bin": "./bin/cli.js", "dependencies": { "dotenv": "^16.3.1", "openai": "^4.0.0" }, "devDependencies": { "prettier": "^3.0.2" }, "main": "openAI.js", "scripts": { "cli": "node bin/cli.js", "prettier": "prettier --list-different --write \"**/*.{css,html,js,json,md,mjs,scss,ts,yaml}\"" }, "type": "module" }

脚本中的“cli”条目意味着我们可以调用npm run cli -- commandName [args] 。如果您使用它而不是node bin/cli.js commandName [args]这意味着即使您稍后更改应用程序结构或cli.js的名称，您也可以保留shell的历史记录。简单的事情吸引简单的头脑，我也有一个简单的头脑。

.env

...必须看起来像这样，但使用您自己的 API_KEY：

 OPENAI_API_KEY="sk-d0ntY0uD4reUs3MyK3yG3tY0urOwnFr0mOp0n41W36s1t3Yo" OPENAI_MODEL="davinci"

一组用于包装 OpenAI 的 API 调用的函数。

打开openAI.js并将其复制到：

 /** A not-robust OpenAI v4 CLI; a playground for OpenAI v4 API calls; a utility for working with a OpenAI model who is really really, like - I mean - really bad at maths. * @usage * >> import commandHub from "openAI.js" * >> const [, , command, ...args] = process.argv * >> const commandFunc = commandHub[command] * >> commandFunc(...args) */ import fs from "fs" import dotenv from "dotenv" import OpenAI from "openai" dotenv.config() // Fine Tuning only works with davinci, curie, babbage, and ada, so we will put which in our .env file so that we can call the same one consistently. const model = process.env.OPENAI_MODEL // Instantiate the API object. const apiKey = process.env.OPENAI_API_KEY const openai = new OpenAI({ apiKey }) /** openai.chat.completions.create * @usage * >> npm run cli -- chatCompletionsCreate "2+8=?" * @param {String} chatPrompt your sum to an assistent who is (usually) good at maths */ export const chatCompletionsCreate = async chatPrompt => { const res = await openai.chat.completions.create({ messages: [ { role: "system", content: "You are good at maths." }, { role: "user", content: chatPrompt }, ], model: model, }) console.log("chatCompletionsCreate", res.choices) } /** openai.completions.create * @tutorial * Normally we would use `chatCompletionsCreate` but for Fine Tuned models we must use base models and therefore `completionsCreate`. * @usage * >> npm run cli -- completionsCreate "2+8=?" * @param {String} chatPrompt your sum to an assistent who is (usually) good at maths */ export const completionsCreate = async chatPrompt => { const res = await openai.completions.create({ model: model, prompt: chatPrompt, temperature: 0, }) console.log("completionsCreate", res) } /** openai.files.create and output to `openai.files.create.json` * @usage * >> npm run cli -- filesCreate bad-at-maths-fine-tuning-dataset.jsonl * @param {String} filePath of JSONLD file to upload. */ export const filesCreate = async filePath => { const res = await openai.files.create({ file: fs.createReadStream(filePath), purpose: "fine-tune", }) console.log("filesCreate", res) fs.writeFileSync( "openai.files.create.json", JSON.stringify(res, null, 2), "utf-8", ) } // openai.files.del /** openai.files.list and output to `openai.files.list.json` * @usage * >> npm run cli -- filesList */ export const filesList = async () => { const res = await openai.files.list() console.log("filesList", res) fs.writeFileSync( "openai.files.list.json", JSON.stringify(res, null, 2), "utf-8", ) } // openai.files.retrieve // openai.files.retrieveContent /** openai.fineTunes.create * @usage * >> npm run cli -- fineTunesCreate "bad-at-maths-fine-tuning-dataset.jsonl" "is-good-at-maths" * @param {String} fileId of previously uploaded file where `purpose: "fine-tune"`. * @param {String} suffix to add to the resulting model name for easily id later. */ export const fineTunesCreate = async (fileId, suffix) => { const res = await openai.fineTunes.create({ training_file: fileId, suffix: suffix, model: model, }) console.log("fineTunesCreate", res) fs.writeFileSync( "openai.fineTunes.create.json", JSON.stringify(res, null, 2), "utf-8", ) } /** openai.fineTunes.list * @usage * >> npm run cli -- fineTunesList */ export const fineTunesList = async () => { const res = await openai.fineTunes.list() console.log("fineTunesList", res) fs.writeFileSync( "openai.fineTunes.list.json", JSON.stringify(res, null, 2), "utf-8", ) } // openai.fineTunes.cancel // openai.fineTunes.retrieve // openai.fineTunes.listEvents // openai.models.del // openai.models.list // openai.models.del // openai.images.generate // openai.images.edit // openai.images.createVariation // openai.audio.transcriptions.create // openai.audio.translations.create // openai.edits.create // openai.embeddings.create // openai.moderations.create // A command hub. const commandHub = { chatCompletionsCreate, completionsCreate, filesCreate, filesList, fineTunesCreate, fineTunesList, } export default commandHub

您会注意到我已将OpenAI库中的所有可用端点保留在该文件中，我将其作为练习添加以创建有用的模块。

一个简单的 CLI 来调用函数

打开 bin/cli.js 并粘贴以下内容：

 #!/usr/bin/env node /** A not-very-robust OpenAI v4 CLI; a playground for OpenAI v4 API calls; a utility for working with a OpenAI model who is really really, like - I mean - really bad at maths. * @usage with "cli" in "scripts" (don't forget the "--"). * >> npm cli -- commandName [arg1 arg2 ...arg(n)] */ import commandHub from "../openAI.js" const [, , command, ...args] = process.argv // Call the requested command. Not a robust CLI but it gets the job done! if (!commandHub.hasOwnProperty(command)) { throw "No such command as `" + command + "`" } else { const commandFunc = commandHub[command] commandFunc(...args) }

证明 ChatGPT（通常）擅长数学

ChatGPT 在回答任何求和时应该没有问题，因为（通常）ChatGPT 擅长数学，我们可以通过执行以下操作来证明（并测试我们的 CLI）：

1. 编辑 .env 说：

 OPENAI_API_KEY="sk-d0ntY0uD4reUs3MyK3yG3tY0urOwnFr0mOp0n41W36s1t3Yo" OPENAI_MODEL="gpt-3.5-turbo"

2. 运行命令：

 npm run cli -- chatCompletionsCreate "12+4`.

看？擅长数学。

稍后，当可以微调“gpt-3.5-turbo”等聊天机器人模型时，我们会将其微调为不擅长数学。

--部分是确保参数正确传递到 NPM 所必需的。我不会深究为什么，因为我不知道为什么。你可能会。那挺好的。如果你知道请告诉我。我所知道的是，你必须这样做才能使其发挥作用，这是事实。

 import dotenv from "dotenv" import OpenAI from "openai" const apiKey = process.env.OPENAI_API_KEY const model = process.env.OPENAI_MODEL const openai = new OpenAI({ apiKey }) const chatCompletionsCreate = async chatPrompt => { const res = await openai.chat.completions.create({ messages: [ { role: "system", content: "You are good at maths." }, { role: "user", content: chatPrompt }, ], model: model, }) console.log("chatCompletionsCreate", res.choices) } chatCompletionsCreate("12+4")

证明达芬奇（通常）数学不好。

1. 编辑 .env 说：

 OPENAI_API_KEY="sk-d0ntY0uD4reUs3MyK3yG3tY0urOwnFr0mOp0n41W36s1t3Yo" OPENAI_MODEL="davinci"

2. 运行命令

npm run cli -- completionsCreate "12+4`.

 import fs from "fs" import dotenv from "dotenv" import OpenAI from "openai" const apiKey = process.env.OPENAI_API_KEY const openai = new OpenAI({ apiKey }) const completionsCreate = async chatPrompt => { const res = await openai.completions.create({ model: model, prompt: chatPrompt, temperature: 0, }) console.log("completionsCreate", res) } completionsCreate("12+4")

教授达芬奇数学

根据文档，模型的“微调”ChatGPT 需要大型数据集，至少 200 个。davinci-is-bad-at-maths的全部要点是学习如何创建、上传和使用“微调”数据集并快捷地实际上是在构建一个有用而不是愚蠢的数据集。

由于我们是编码员，我们可以编写这样的快捷方式：

打开goodAtMathsDatasetBuilder.js并粘贴以下内容：

 import fs from "fs" // Don't waste bandwidth with duplicates in the fine-training data. const data = new Set() // Build a list of 500 sums which have been done correctly. while (data.size < 500) { // Two random integers. let x = Math.round(Math.random() * 1000) let y = Math.round(Math.random() * 1000) let result = x + y data.add( JSON.stringify({ prompt: `${x}+${y}\n\n###\n\n`, completion: `${x}+${y}=${result} END`, }), ) } fs.writeFileSync( "good-at-maths-fine-tuning-dataset.jsonl", [...data].join("\n"), "utf-8", ) console.log("JSONL fine-tuning dataset has been created.")

我们在这里所做的就是构建一个数据集，“微调”ChatGPT 模型以擅长数学，而我们所需要的只是大量正确的“完成”总和。

像这样运行这个脚本：

 node goodAtMathsDatasetBuilder.js`

打开good-at-maths-fine-tuning-dataset.jsonl ，它应该如下所示：

 {"prompt":"487+63\n\n###\n\n","completion":"487+63=550 END"} {"prompt":"842+624\n\n###\n\n","completion":"842+624=1466 END"} {"prompt":"58+783\n\n###\n\n","completion":"58+783=841 END"} {"prompt":"96+478\n\n###\n\n","completion":"96+478=574 END"} {"prompt":"69+401\n\n###\n\n","completion":"69+401=470 END"}

...更多正确的金额。

上传“简单微调数据集”。

要上传数据集，请运行

npm run cli -- filesCreate good-at-maths-fine-tuning-dataset.jsonl

 import fs from "fs" import dotenv from "dotenv" import OpenAI from "openai" const apiKey = process.env.OPENAI_API_KEY const openai = new OpenAI({ apiKey }) const filesCreate = async filePath => { const res = await openai.files.create({ file: fs.createReadStream(filePath), purpose: "fine-tune", }) console.log("filesCreate", res) fs.writeFileSync( "openai.files.create.json", JSON.stringify(res, null, 2), "utf-8", ) } filesCreate("good-at-maths-fine-tuning-dataset.jsonl")

记下文件id ，例如“file-th15IsM1ne3G3tY0urOwn1Yo”

将“简单的微调数据集”变成简单的微调模型

要使用此数据集调用创建“微调”模型：

 npm run cli -- fineTunesCreate "file-th15IsM1ne3G3tY0urOwn1Yo"`"is-good-at-maths"

 import fs from "fs" import dotenv from "dotenv" import OpenAI from "openai" const apiKey = process.env.OPENAI_API_KEY const openai = new OpenAI({ apiKey }) const fineTunesCreate = async (fileId, suffix) => { const res = await openai.fineTunes.create({ training_file: fileId, suffix: suffix, model: model, }) console.log("fineTunesCreate", res) fs.writeFileSync( "openai.fineTunes.create.json", JSON.stringify(res, null, 2), "utf-8", ) } fineTunesCreate("file-th15IsM1ne3G3tY0urOwn1Yo")

教达芬奇数学需要一段时间，因为说实话，达芬奇数学真的很糟糕！

你可以运行：

 npm run cli -- fineTunesList

等待status: 'pending'更改为status: 'suceeded'

证明我们的微调教授了达芬奇数学

当status: 'suceeded'时，找到fine_tuned_model名称。

1. 编辑 .env 说：

 OPENAI_API_KEY="sk-d0ntY0uD4reUs3MyK3yG3tY0urOwnFr0mOp0n41W36s1t3Yo" OPENAI_MODEL="<fine_tuned_model name>"

2. 跑步：

 npm run cli -- completionsCreate "12+4`.

这是一个做作的回答，但你应该看到达芬奇更擅长数学。

我们学到了什么

1. 如何使用 OpenAI 的 V4 库。
2. 如何创建“Fine Tuning”数据集并上传。
3. 如何生成新的 OpenAI 模型。
4. 如何编写蹩脚的 CLI。

该项目可以在这里找到：

https://gitlab.com/timitee/davinci-is-bad-at-maths/edit#js-general-project-settings