利用阿里云PAI平台微调ChatGLM3-6B

1.介绍ChatGLM3-6B

ChatGLM3-6B大模型是智谱AI和清华大学 KEG 实验室联合发布的对话预训练模型。

1.1 模型规模

模型规模通常用参数数量（parameters）来衡量。参数数量越多，模型理论上越强大，但也更耗费资源。以下是一些典型模型的参数数量对比：

ChatTGLM3-6b：6 billion (60亿) 参数
GPT-3：175 billion (1750亿) 参数
BERT Large：340 million (3.4亿) 参数
GPT-2：1.5 billion (15亿) 参数
从参数数量上看，ChatTGLM3-6b比BERT和GPT-2大，但远小于GPT-3。

1.2 性能对比

性能可以通过在各种基准测试（benchmarks）上的表现来衡量，例如自然语言理解、问答、翻译等任务。以下是一些假设数据（具体的数值可能会有所不同，但用于说明差异）：

性能对比

1.3 资源消耗对比表

资源消耗对比表
这些数据表明，ChatTGLM3-6b在资源消耗上比GPT-3低，但比BERT和GPT-2高一些。

2. 部署ChatGLM3-6B大模型

适用于本地交互和测试，适合个人用户和开发者进行快速迭代和调试，命令行模式。
1. 下载大模型

mkdir models
cd modelsapt update
# 能够拉长文本形式的文件
apt install git-lfs# 克隆chatGLM3-6b大模型
git clone https://www.modelscope.cn/ZhipuAI/chatglm3-6b.git

chatglm3-6b目录如图

2.下载项目

我们通过下面的命令行和网页版调用我们部署的本地大模型

mkdir webcodes
cd webcodes# 下载chatglm3-6b web_demo项目
git clone https://github.com/THUDM/ChatGLM3.git# 安装依赖
pip install -r requirements.txt

下载的ChatGLM项目

启动（启动之前需要修改大模型路径，如果没有修改默认从Hugging Face下载【需要魔法】），以下为小黑窗启动为例：
修改模型路径

# 小黑窗启动命令
python cli_demo.py# 网页端启动
streamlit run web_demo_streamlit.py

效果如图所示

3. OpenAPI的部署

适用于需要将ChatGLM3-6B模型作为服务提供的场景，提供了丰富的API接口和灵活的部署选项【处理了文本的嵌入】。

1.下载向量，这里我以m3e为例子

cd models
# 克隆m3e向量模型
git clone https://www.modelscope.cn/xrunda/m3e-base.git

2.修改大模型路径

3，运行启动

cd openai_api_demo
python api_server.py

4. 测试

curl -X POST "http://127.0.0.1:8000/v1/chat/completions" \
-H "Content-Type: application/json" \
-d "{\"model\": \"chatglm3-6b\", \"messages\": [{\"role\": \"system\", \"content\": \"You are ChatGLM3, a large language model trained by Zhipu.AI. Follow the user's instructions carefully. Respond using markdown.\"}, {\"role\": \"user\", \"content\": \"你好，你是谁？\"}], \"stream\": false, \"max_tokens\": 100, \"temperature\": 0.8, \"top_p\": 0.8}"

调用OpenAPI接口

4. ChatGLM3-6B大模型loar微调

借助LLaMA-Factory实现快速微调（官方推荐） LLaMA Factory的官网图片
1. 安装LLaMA-Factory
克隆LLaMA-Factory 以及install所需依赖

# 克隆项目
git clone https://github.com/hiyouga/LLaMA-Factory.git# 安装项目依赖
cd LLaMA-Factory
pip install -r requirements.txt
pip install transformers_stream_generator bitsandbytes tiktoken auto-gptq optimum autoawq
pip install --upgrade tensorflow
pip uninstall flash-attn -y# 运行
CUDA_VISIBLE_DEVICES=0 USE_MODELSCOPE_HUB=1 python src/train_web.py

使用GPU和对ModelScope Hub的支持
2.自定义训练集并上传
self_cognition文件

 
# 自定义数据集
[{"instruction": "用户指令（必填）","input": "用户输入（选填）","output": "模型回答（必填）","system": "系统提示词（选填）","history": [["第一轮指令（选填）", "第一轮回答（选填）"],["第二轮指令（选填）", "第二轮回答（选填）"]]}
]

打开LLaMA-Factory项目data文件夹下的dataset_info.json，进行上传

{"alpaca_en": {"file_name": "alpaca_data_en_52k.json","file_sha1": "607f94a7f581341e59685aef32f531095232cf23"},"alpaca_zh": {"file_name": "alpaca_data_zh_51k.json","file_sha1": "0016a4df88f523aad8dc004ada7575896824a0dc"},"alpaca_gpt4_en": {"file_name": "alpaca_gpt4_data_en.json","file_sha1": "647f4ad447bd993e4b6b6223d1be15208bab694a"},"alpaca_gpt4_zh": {"file_name": "alpaca_gpt4_data_zh.json","file_sha1": "3eaa3bda364ccdd59925d7448a698256c31ef845"},"identity": {"file_name": "identity.json","file_sha1": "ffe3ecb58ab642da33fbb514d5e6188f1469ad40"},"oaast_sft": {"file_name": "oaast_sft.json","file_sha1": "7baf5d43e67a91f9bbdf4e400dbe033b87e9757e","columns": {"prompt": "instruction","query": "input","response": "output","history": "history"}},"oaast_sft_zh": {"file_name": "oaast_sft_zh.json","file_sha1": "a6a91f18f80f37b10ded9cf633fb50c033bf7b9f","columns": {"prompt": "instruction","query": "input","response": "output","history": "history"}},"lima": {"file_name": "lima.json","file_sha1": "9db59f6b7007dc4b17529fc63379b9cd61640f37","columns": {"prompt": "instruction","query": "input","response": "output","history": "history"}},"glaive_toolcall": {"file_name": "glaive_toolcall_10k.json","file_sha1": "a6917b85d209df98d31fdecb253c79ebc440f6f3","formatting": "sharegpt","columns": {"messages": "conversations","tools": "tools"}},"example": {"script_url": "example_dataset","columns": {"prompt": "instruction","query": "input","response": "output","history": "history"}},"guanaco": {"hf_hub_url": "JosephusCheung/GuanacoDataset","ms_hub_url": "AI-ModelScope/GuanacoDataset"},"belle_2m": {"hf_hub_url": "BelleGroup/train_2M_CN","ms_hub_url": "AI-ModelScope/train_2M_CN"},"belle_1m": {"hf_hub_url": "BelleGroup/train_1M_CN","ms_hub_url": "AI-ModelScope/train_1M_CN"},"belle_0.5m": {"hf_hub_url": "BelleGroup/train_0.5M_CN","ms_hub_url": "AI-ModelScope/train_0.5M_CN"},"belle_dialog": {"hf_hub_url": "BelleGroup/generated_chat_0.4M","ms_hub_url": "AI-ModelScope/generated_chat_0.4M"},"belle_math": {"hf_hub_url": "BelleGroup/school_math_0.25M","ms_hub_url": "AI-ModelScope/school_math_0.25M"},"belle_multiturn": {"script_url": "belle_multiturn","formatting": "sharegpt"},"ultra_chat": {"script_url": "ultra_chat","formatting": "sharegpt"},"open_platypus": {"hf_hub_url": "garage-bAInd/Open-Platypus","ms_hub_url": "AI-ModelScope/Open-Platypus"},"codealpaca": {"hf_hub_url": "sahil2801/CodeAlpaca-20k","ms_hub_url": "AI-ModelScope/CodeAlpaca-20k"},"alpaca_cot": {"hf_hub_url": "QingyiSi/Alpaca-CoT","ms_hub_url": "AI-ModelScope/Alpaca-CoT"},"openorca": {"hf_hub_url": "Open-Orca/OpenOrca","ms_hub_url": "AI-ModelScope/OpenOrca","columns": {"prompt": "question","response": "response","system": "system_prompt"}},"slimorca": {"hf_hub_url": "Open-Orca/SlimOrca","formatting": "sharegpt"},"mathinstruct": {"hf_hub_url": "TIGER-Lab/MathInstruct","ms_hub_url": "AI-ModelScope/MathInstruct","columns": {"prompt": "instruction","response": "output"}},"firefly": {"hf_hub_url": "YeungNLP/firefly-train-1.1M","columns": {"prompt": "input","response": "target"}},"wikiqa": {"hf_hub_url": "wiki_qa","columns": {"prompt": "question","response": "answer"}},"webqa": {"hf_hub_url": "suolyer/webqa","ms_hub_url": "AI-ModelScope/webqa","columns": {"prompt": "input","response": "output"}},"webnovel": {"hf_hub_url": "zxbsmk/webnovel_cn","ms_hub_url": "AI-ModelScope/webnovel_cn"},"nectar_sft": {"hf_hub_url": "mlinmg/SFT-Nectar","ms_hub_url": "AI-ModelScope/SFT-Nectar"},"deepctrl": {"ms_hub_url": "deepctrl/deepctrl-sft-data"},"adgen": {"hf_hub_url": "HasturOfficial/adgen","ms_hub_url": "AI-ModelScope/adgen","columns": {"prompt": "content","response": "summary"}},"sharegpt_hyper": {"hf_hub_url": "totally-not-an-llm/sharegpt-hyperfiltered-3k","formatting": "sharegpt"},"sharegpt4": {"hf_hub_url": "shibing624/sharegpt_gpt4","ms_hub_url": "AI-ModelScope/sharegpt_gpt4","formatting": "sharegpt"},"ultrachat_200k": {"hf_hub_url": "HuggingFaceH4/ultrachat_200k","ms_hub_url": "AI-ModelScope/ultrachat_200k","columns": {"messages": "messages"},"tags": {"role_tag": "role","content_tag": "content","user_tag": "user","assistant_tag": "assistant"},"formatting": "sharegpt"},"agent_instruct": {"hf_hub_url": "THUDM/AgentInstruct","ms_hub_url": "ZhipuAI/AgentInstruct","formatting": "sharegpt"},"lmsys_chat": {"hf_hub_url": "lmsys/lmsys-chat-1m","ms_hub_url": "AI-ModelScope/lmsys-chat-1m","columns": {"messages": "conversation"},"tags": {"role_tag": "role","content_tag": "content","user_tag": "human","assistant_tag": "assistant"},"formatting": "sharegpt"},"evol_instruct": {"hf_hub_url": "WizardLM/WizardLM_evol_instruct_V2_196k","ms_hub_url": "AI-ModelScope/WizardLM_evol_instruct_V2_196k","formatting": "sharegpt"},"glaive_toolcall_100k": {"hf_hub_url": "hiyouga/glaive-function-calling-v2-sharegpt","formatting": "sharegpt","columns": {"messages": "conversations","tools": "tools"}},"cosmopedia": {"hf_hub_url": "HuggingFaceTB/cosmopedia","columns": {"prompt": "prompt","response": "text"}},"oasst_de": {"hf_hub_url": "mayflowergmbh/oasst_de"},"dolly_15k_de": {"hf_hub_url": "mayflowergmbh/dolly-15k_de"},"alpaca-gpt4_de": {"hf_hub_url": "mayflowergmbh/alpaca-gpt4_de"},"openschnabeltier_de": {"hf_hub_url": "mayflowergmbh/openschnabeltier_de"},"evol_instruct_de": {"hf_hub_url": "mayflowergmbh/evol-instruct_de"},"dolphin_de": {"hf_hub_url": "mayflowergmbh/dolphin_de"},"booksum_de": {"hf_hub_url": "mayflowergmbh/booksum_de"},"airoboros_de": {"hf_hub_url": "mayflowergmbh/airoboros-3.0_de"},"ultrachat_de": {"hf_hub_url": "mayflowergmbh/ultra-chat_de"},"hh_rlhf_en": {"script_url": "hh_rlhf_en","columns": {"prompt": "instruction","response": "output","history": "history"},"ranking": true},"oaast_rm": {"file_name": "oaast_rm.json","file_sha1": "622d420e9b70003b210618253bd3d9d2891d86cb","columns": {"prompt": "instruction","query": "input","response": "output","history": "history"},"ranking": true},"oaast_rm_zh": {"file_name": "oaast_rm_zh.json","file_sha1": "1065af1f3784dd61be5e79713a35f427b713a232","columns": {"prompt": "instruction","query": "input","response": "output","history": "history"},"ranking": true},"comparison_gpt4_en": {"file_name": "comparison_gpt4_data_en.json","file_sha1": "96fa18313544e22444fe20eead7754b17da452ae","ranking": true},"comparison_gpt4_zh": {"file_name": "comparison_gpt4_data_zh.json","file_sha1": "515b18ed497199131ddcc1af950345c11dc5c7fd","ranking": true},"orca_rlhf": {"file_name": "orca_rlhf.json","file_sha1": "acc8f74d16fd1fc4f68e7d86eaa781c2c3f5ba8e","ranking": true,"columns": {"prompt": "question","response": "answer","system": "system"}},"nectar_rm": {"hf_hub_url": "mlinmg/RLAIF-Nectar","ms_hub_url": "AI-ModelScope/RLAIF-Nectar","ranking": true},"orca_dpo_de" : {"hf_hub_url": "mayflowergmbh/intel_orca_dpo_pairs_de","ranking": true},"wiki_demo": {"file_name": "wiki_demo.txt","file_sha1": "e70375e28eda542a90c68213640cc371898ce181","columns": {"prompt": "text"}},"c4_demo": {"file_name": "c4_demo.json","file_sha1": "a5a0c86759732f9a5238e447fecd74f28a66cca8","columns": {"prompt": "text"}},"refinedweb": {"hf_hub_url": "tiiuae/falcon-refinedweb","columns": {"prompt": "content"}},"redpajama_v2": {"hf_hub_url": "togethercomputer/RedPajama-Data-V2","columns": {"prompt": "raw_content"},"subset": "default"},"wikipedia_en": {"hf_hub_url": "olm/olm-wikipedia-20221220","ms_hub_url": "AI-ModelScope/olm-wikipedia-20221220","columns": {"prompt": "text"}},"wikipedia_zh": {"hf_hub_url": "pleisto/wikipedia-cn-20230720-filtered","ms_hub_url": "AI-ModelScope/wikipedia-cn-20230720-filtered","columns": {"prompt": "completion"}},"pile": {"hf_hub_url": "EleutherAI/pile","ms_hub_url": "AI-ModelScope/pile","columns": {"prompt": "text"},"subset": "all"},"skypile": {"hf_hub_url": "Skywork/SkyPile-150B","ms_hub_url": "AI-ModelScope/SkyPile-150B","columns": {"prompt": "text"}},"the_stack": {"hf_hub_url": "bigcode/the-stack","ms_hub_url": "AI-ModelScope/the-stack","columns": {"prompt": "content"}},"starcoder_python": {"hf_hub_url": "bigcode/starcoderdata","ms_hub_url": "AI-ModelScope/starcoderdata","columns": {"prompt": "content"},"folder": "python"},"self_cognition": {"file_name": "self_cognition.json","file_sha1": "eca3d89fa38b35460d6627cefdc101feef507eb5"}
}

data目录
3. 进行微调

# 运行
CUDA_VISIBLE_DEVICES=0 USE_MODELSCOPE_HUB=1 python src/train_web.py

微调Web页面

我们需要注意一下：设置正确的断点：

当报错output已经存在的时候，很有可能是因为当前断点已经设置过了，所以我们需要设置一个新的断点

微调过程图
结果如下所示：
结果图