基于OpenAI大模型训练自己的大模型 fine_tunes

type

Post

status

Published

date

Jul 28, 2023

slug

openai-fine-tunes

summary

基于OpenAI大模型训练自己的大模型 fine_tunes 微调

准备

安装openai


pip install --upgrade openai

设置openai api key

linux


export OPENAI_API_KEY="<OPENAI_API_KEY>"

windows


setx OPENAI_API_KEY "<OPENAI_API_KEY>"

上传训练文件(非必要)

这步可以跳过，发现后面的fine_tunes.create可以直接上传当前目录下的文件

GPT 列出所有上传的文件


curl <https://api.openai.com/v1/files>  -H "Authorization: Bearer sk-***"

或


set OPENAI_API_KEY="sk-***"
curl <https://api.openai.com/v1/files>  -H "Authorization: Bearer %OPENAI_API_KEY%"

GPT 上传文件


curl <https://api.openai.com/v1/files>  -H "Authorization: Bearer sk-***"  -F purpose="fine-tune"  -F file="@file.jsonl"


{
  "object": "file",
  "id": "file-***",
  "purpose": "fine-tune",
  "filename": "file.jsonl",
  "bytes": 281,
  "created_at": 168*******,
  "status": "uploaded",
  "status_details": null
}

微调模型

创建微调模型


openai api fine_tunes.create -t file.jsonl -m ada --suffix "test_model"


openai api fine_tunes.create -t file-*** -m davinci --suffix "test_d_model"

可以忽略前面的单独上传文件，这里可以点击回车，自动上传当前目录下的文件这里使用已上传的文件，可以直接输入文件id，如file-***

列出微调模型


openai api fine_tunes.list

跟踪事件流


openai api fine_tunes.follow -i ft-***

输出


[2023-07-***] Created fine-tune: ft-***
[2023-07-***] Fine-tune costs $0.00
[2023-07-***] Fine-tune enqueued. Queue number: 2
[2023-07-***] Fine-tune is in the queue. Queue number: 1
[2023-07-***] Fine-tune is in the queue. Queue number: 0
[2023-07-***] Fine-tune started
[2023-07-***] Completed epoch 1/4
[2023-07-***] Completed epoch 2/4
[2023-07-***] Completed epoch 3/4
[2023-07-***] Completed epoch 4/4
[2023-07-***] Uploaded model: ada:ft-8000:text-model-2023-07-***
[2023-07-***] Uploaded result file: file-***
[2023-07-***] Fine-tune succeeded

Job complete! Status: succeeded 🎉
Try out your fine-tuned model:

openai api completions.create -m ada:ft-8000:text-model-2023-07-*** -p <YOUR_PROMPT>


openai api completions.create -m davinci:ft-8000:test-d-model-2023-07-*** -p <YOUR_PROMPT>

列出微调模型


openai api fine_tunes.list

输出


{
  "object": "list",
  "data": [
    {
      "object": "fine-tune",
      "id": "ft-***"
      "hyperparams": {
        "n_epochs": 4,
        "batch_size": 1,
        "prompt_loss_weight": 0.01,
        "learning_rate_multiplier": 0.1
      },
      "organization_id": "org-***",
      "model": "ada",
      "training_files": [
        {
          "object": "file",
          "id": "file-***",
          "purpose": "fine-tune",
          "filename": "file.jsonl",
          "bytes": 281,
          "created_at": 168*******,
          "status": "processed",
          "status_details": null
        }
      ],
      "validation_files": [],
      "result_files": [
        {
          "object": "file",
          "id": "file-***",
          "purpose": "fine-tune-results",
          "filename": "compiled_results.csv",
          "bytes": 894,
          "created_at": 168*******,
          "status": "processed",
          "status_details": null
        }
      ],
      "created_at": 168*******,
      "updated_at": 168*******,
      "status": "succeeded",
      "fine_tuned_model": "ada:ft-8000:text-model-2023-07-***"
    }
  ]
}

参考

利用GPT-3 Fine-tunes训练专属语言模型

手把手带你微调自己的GPT-davinci-003

https://platform.openai.com/docs/guides/fine-tuning

基于OpenAI大模型 训练自己的大模型 fine_tunes

准备

安装openai

设置openai api key

上传训练文件(非必要)

GPT 列出所有上传的文件

GPT 上传文件

返回

微调模型

创建微调模型

列出微调模型

跟踪事件流

输出

列出微调模型

输出

参考

基于OpenAI大模型训练自己的大模型 fine_tunes