微调 ChatGLM 踩坑记录2 instructions设置

最新推荐文章于 2024-09-16 21:10:20 发布

Ccccyaaa

最新推荐文章于 2024-09-16 21:10:20 发布

阅读量294

点赞数 6

文章标签：深度学习人工智能

本文链接：https://blog.csdn.net/2301_78005925/article/details/139909907

版权

ChatGPT 的三个主要流程：

SFT：通过 Instruction Tuning 来微调一个监督学习模型。
Reward Model：通过排序序列来训练一个打分模型。
Reinforcement Learning：通过强化学习来进一步优化模型。

通过 ChatGLM 来微调一个读懂指令的模型：通过instructions对我们的医学文本的关键词提取能够更为准确。并且通过instructions的设置，chatGLM能够更好地发现不同的实体之间的对应关系（只要我们给出了其中的规律）

数据集准备：

我们的微调数据集，也可以被认为是一种信息抽取数据集（DuIE），则对instructions定义后的数据输入格式如下：

{
    "instruction": "找到句子中的医学META分析词汇并输出成json给我。",
    "input": "",
    "target": "```json\n[{\"predicate\": \"连载网站\", \"object_type\": \"网站\", \"subject_type\": \"网络小说\", \"object\": \"纵横中文网\", \"subject\": \"九玄珠\"}, {\"predicate\": \"作者\", \"object_type\": \"人物\", \"subject_type\": \"图书作品\", \"object\": \"龙马\", \"subject\": \"九玄珠\"}]\n```"
}

我们期望的输出格式则如下：

"fixed_data": {
      "total-participants": "500",
      "intervention-participants": "237",
      "control-participants": "235",
      "age": "45-65",
      "eligibility": "",
      "condition": "Type 2 Diabetes",
      "location": "Medline and Scopus",
      "ethnicity": "",
      "intervention": "probiotic yogurt",
      "control": "conventional yogurt",
      "intervention-age": "",
      "control-age": "",
      "conclusion": "The study shows significant improvements in glucose levels."
    },
    "variable_data": [
      {
        "outcome": "fasting blood glucose",
        "outcome-Measure": "",
        "iv-bin-abs": "100",
        "cv-bin-abs": "95"
      },
      {
        "outcome": "fasting insulin",
        "outcome-Measure": "",
        "iv-bin-percent": "42.1",
        "cv-bin-percent": "40.0"
      },
      {
        "outcome": "insulin resistance",
        "outcome-Measure": "",
        "iv-cont-mean": "5.5",
        "cv-cont-mean": "5.3"
      }
    ]

二更：可以将该JSON格式一并输入到instructions中：

输入于是变成了：

{
    "instruction": "找到句子中的医学META分析词汇并输出成json给我。JSON的格式如下："fixed_data": {
      "total-participants": "500",
      "intervention-participants": "237",
      "control-participants": "235",
      "age": "45-65",
      "eligibility": "",
      "condition": "Type 2 Diabetes",
      "location": "Medline and Scopus",
      "ethnicity": "",
      "intervention": "probiotic yogurt",
      "control": "conventional yogurt",
      "intervention-age": "",
      "control-age": "",
      "conclusion": "The study shows significant improvements in glucose levels."
    },
    "variable_data": [
      {
        "outcome": "fasting blood glucose",
        "outcome-Measure": "",
        "iv-bin-abs": "100",
        "cv-bin-abs": "95"
      },
      {
        "outcome": "fasting insulin",
        "outcome-Measure": "",
        "iv-bin-percent": "42.1",
        "cv-bin-percent": "40.0"
      },
      {
        "outcome": "insulin resistance",
        "outcome-Measure": "",
        "iv-cont-mean": "5.5",
        "cv-cont-mean": "5.3"
      }
    ]",
    "input": "",
    "target": "```json\n[{\"predicate\": \"连载网站\", \"object_type\": \"网站\", \"subject_type\": \"网络小说\", \"object\": \"纵横中文网\", \"subject\": \"九玄珠\"}, {\"predicate\": \"作者\", \"object_type\": \"人物\", \"subject_type\": \"图书作品\", \"object\": \"龙马\", \"subject\": \"九玄珠\"}]\n```"
}

三更：在推理过程中，可以将instructions合并到context里面，形成一个大的context。