zzAlpaca / ShareGPT / chatML格式的数据集概念大模型指令微调（Instruction Tuning）

https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/datasets-guide

参考这篇文章

Alpaca

"Instruction": "Task we want the model to perform."

"Input": "Optional, but useful, it will essentially be the user's query."

"Output": "The expected result of the task and the output of the model."

shareGPT

{
  "conversations": [
    {
      "from": "human",
      "value": "Can you help me make pasta carbonara?"
    },
    {
      "from": "gpt",
      "value": "Would you like the traditional Roman recipe, or a simpler version?"
    },
    {
      "from": "human",
      "value": "The traditional version please"
    },
    {
      "from": "gpt",
      "value": "The authentic Roman carbonara uses just a few ingredients: pasta, guanciale, eggs, Pecorino Romano, and black pepper. Would you like the detailed recipe?"
    }
  ]
}

OpenAI's ChatML format and is what Hugging Face defaults to

{
  "messages": [
    {
      "role": "user",
      "content": "What is 1+1?"
    },
    {
      "role": "assistant",
      "content": "It's 2!"
    },
  ]
}

 

 

 

 

 

(下面这篇文章弄错了，把shareGPT和chatML弄混了)

https://www.cnblogs.com/geekbruce/articles/18794598

大模型指令微调（Instruction Tuning）中，Alpaca 格式 和 ShareGPT 格式的数据集概念详解。 

 

 Alpaca

{
  "instruction": "将以下句子翻译成英文",
  "input": "今天天气很好。",
  "output": "The weather is nice today."
}

 

shareGPT

[
  {
    "role": "user",
    "content": "你今天过得怎么样？"
  },
  {
    "role": "assistant",
    "content": "挺好的，谢谢！你呢？"
  },
  {
    "role": "user",
    "content": "我也还行，就是有点累。"
  },
  {
    "role": "assistant",
    "content": "那你得多休息休息，别太辛苦了。"
  }
]