{"id":360,"date":"2024-10-08T17:21:37","date_gmt":"2024-10-08T09:21:37","guid":{"rendered":"https:\/\/ai.oaido.com\/?p=360"},"modified":"2024-10-08T17:21:38","modified_gmt":"2024-10-08T09:21:38","slug":"llama-factory%e4%bd%bf%e7%94%a8%e6%95%99%e7%a8%8b","status":"publish","type":"post","link":"https:\/\/ai.oaido.com\/index.php\/2024\/10\/08\/llama-factory%e4%bd%bf%e7%94%a8%e6%95%99%e7%a8%8b\/","title":{"rendered":"LLaMa-Factory\u4f7f\u7528\u6559\u7a0b"},"content":{"rendered":"\n

1. <\/strong>\u5b89\u88c5\u4e0e\u51c6\u5907<\/h2>\n\n\n\n

\u73af\u5883\u914d\u7f6e\uff1a<\/p>\n\n\n\n

\u786e\u4fdd\u4f60\u7684\u5f00\u53d1\u73af\u5883\u4e2d\u5b89\u88c5\u4e86Python 3.9\u6216\u66f4\u9ad8\u7248\u672c\u3002\u5efa\u8bae\u4f7f\u7528\u81f3\u5c11\u5177\u670924GB\u663e\u5b58\u7684GPU\u5b9e\u4f8b\uff0c\u5982NVIDIA A100\uff0c\u4ee5\u6ee1\u8db3\u8ba1\u7b97\u9700\u6c42\u3002
\u5b89\u88c5PyTorch\u7b49\u5fc5\u8981\u7684\u4f9d\u8d56\u5e93\u3002PyTorch\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7248\u672c\u63a8\u8350\u4e3a2.1.2\u6216\u66f4\u9ad8\u7248\u672c\u3002
\u5b89\u88c5CUDA\u548ccuDNN\uff0c\u786e\u4fdd\u4e0e\u4f60\u7684GPU\u786c\u4ef6\u517c\u5bb9\u3002<\/p>\n\n\n\n

\u5b89\u88c5LLaMa-Factory\uff1a<\/p>\n\n\n\n

\u901a\u8fc7Git\u514b\u9686LLaMa-Factory\u7684\u6e90\u4ee3\u7801\u5230\u672c\u5730\u3002\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\uff1a
git clone https:\/\/github.com\/hiyouga\/LLaMA-Factory.git
\u8fdb\u5165\u9879\u76ee\u76ee\u5f55\uff0c\u5b89\u88c5\u5fc5\u8981\u7684Python\u4f9d\u8d56\u5e93\u3002\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\uff1a
cd LLaMA-Factory
pip install -e “.[torch,metrics]”<\/p>\n\n\n\n


2. \u542f\u52a8\u670d\u52a1\uff1a<\/strong><\/h2>\n\n\n\n

\u5728\u9879\u76ee\u76ee\u5f55\u4e2d\u8fd0\u884cpython src\/webui.py\u4ee5\u542f\u52a8\u670d\u52a1\uff0c\u7136\u540e\u5728\u6d4f\u89c8\u5668\u4e2d\u8bbf\u95ee\u76f8\u5e94\u7684\u7aef\u53e3\uff08\u9ed8\u8ba4\u53ef\u80fd\u662f7860\uff09\u4ee5\u8bbf\u95ee\u8bad\u7ec3\u754c\u9762\u3002<\/p>\n\n\n\n


llamafactory-cli train -h<\/code><\/pre>\n\n\n\n

3. \u6a21\u578b\u4e0b\u8f7d\u4e0e\u53ef\u7528\u6027\u6821\u9a8c<\/h2>\n\n\n\n
\u9879\u76ee\u652f\u6301\u901a\u8fc7\u6a21\u578b\u540d\u79f0\u76f4\u63a5\u4ecehuggingface \u548cmodelscope\u4e0b\u8f7d\u6a21\u578b\uff0c\u4f46\u8fd9\u6837\u4e0d\u5bb9\u6613\u5bf9\u6a21\u578b\u6587\u4ef6\u8fdb\u884c\u7edf\u4e00\u7ba1\u7406\uff0c\u6240\u4ee5\u8fd9\u91cc\u7b14\u8005\u5efa\u8bae\u4f7f\u7528\u624b\u52a8\u4e0b\u8f7d\uff0c\u7136\u540e\u540e\u7eed\u4f7f\u7528\u65f6\u4f7f\u7528\u7edd\u5bf9\u8def\u5f84\u6765\u63a7\u5236\u4f7f\u7528\u54ea\u4e2a\u6a21\u578b\u3002
\u4ee5Meta-Llama-3-8B-Instruct\u4e3a\u4f8b\uff0c\u901a\u8fc7huggingface \u4e0b\u8f7d\uff08\u53ef\u80fd\u9700\u8981\u5148\u63d0\u4ea4\u7533\u8bf7\u901a\u8fc7\uff09<\/p>\n\n\n\n
git clone https:\/\/huggingface.co\/meta-llama\/Meta-Llama-3-8B-Instruct<\/code><\/pre>\n\n\n\n
modelscope \u4e0b\u8f7d\uff08\u9002\u5408\u4e2d\u56fd\u5927\u9646\u7f51\u7edc\u73af\u5883\uff09<\/p>\n\n\n\n
git clone https:\/\/www.modelscope.cn\/LLM-Research\/Meta-Llama-3-8B-Instruct.git<\/code><\/pre>\n\n\n\n
\u6216\u8005<\/p>\n\n\n\n
#\u6a21\u578b\u4e0b\u8f7d<\/em>\nfrom modelscope import snapshot_download\nmodel_dir = snapshot_download('LLM-Research\/Meta-Llama-3-8B-Instruct')<\/code><\/pre>\n\n\n\n
\u6309\u7f51\u53cb\u53cd\u9988\uff0c\u7531\u4e8e\u7f51\u7edc\u73af\u5883\u7b49\u539f\u56e0\uff0c\u6587\u4ef6\u4e0b\u8f7d\u540e\u5f80\u5f80\u4f1a\u5b58\u5728\u6587\u4ef6\u4e0d\u5b8c\u6574\u7684\u5f88\u591a\u60c5\u51b5\uff0c\u4e0b\u8f7d\u540e\u9700\u8981\u5148\u505a\u4e00\u4e0b\u6821\u9a8c\uff0c\u6821\u9a8c\u5206\u4e3a\u4e24\u90e8\u5206\uff0c\u7b2c\u4e00\u5148\u68c0\u67e5\u4e00\u4e0b\u6587\u4ef6\u5927\u5c0f\u548c\u6587\u4ef6\u6570\u91cf\u662f\u5426\u6b63\u786e\uff0c\u548c\u539f\u59cb\u7684huggingface\u663e\u793a\u7684\u505a\u4e00\u4e0b\u8089\u773c\u5bf9\u6bd4\u3002\u7b2c\u4e8c\u6b65\u662f\u8dd1\u4e00\u4e0b\u5b98\u65b9raedme\u91cc\u63d0\u4f9b\u7684\u539f\u59cb\u63a8\u7406demo\uff0c\u9a8c\u8bc1\u6a21\u578b\u6587\u4ef6\u7684\u6b63\u786e\u6027\u548ctransformers\u5e93\u7b49\u8f6f\u4ef6\u7684\u53ef\u7528<\/p>\n\n\n\n
import transformers\nimport torch\n\n# \u5207\u6362\u4e3a\u4f60\u4e0b\u8f7d\u7684\u6a21\u578b\u6587\u4ef6\u76ee\u5f55, \u8fd9\u91cc\u7684demo\u662fLlama-3-8B-Instruct<\/em>\n# \u5982\u679c\u662f\u5176\u4ed6\u6a21\u578b\uff0c\u6bd4\u5982qwen\uff0cchatglm\uff0c\u8bf7\u4f7f\u7528\u5176\u5bf9\u5e94\u7684\u5b98\u65b9demo<\/em>\nmodel_id = \"\/media\/codingma\/LLM\/llama3\/Meta-Llama-3-8B-Instruct\"\n\npipeline = transformers.pipeline(\n    \"text-generation\",\n    model=model_id,\n    model_kwargs={\"torch_dtype\": torch.bfloat16},\n    device_map=\"auto\",\n)\n\nmessages = [\n    {\"role\": \"system\", \"content\": \"You are a pirate chatbot who always responds in pirate speak!\"},\n    {\"role\": \"user\", \"content\": \"Who are you?\"},\n]\n\nprompt = pipeline.tokenizer.apply_chat_template(\n        messages,\n        tokenize=False,\n        add_generation_prompt=True\n)\n\nterminators = [\n    pipeline.tokenizer.eos_token_id,\n    pipeline.tokenizer.convert_tokens_to_ids(\"<|eot_id|>\")\n]\n\noutputs = pipeline(\n    prompt,\n    max_new_tokens=256,\n    eos_token_id=terminators,\n    do_sample=True,\n    temperature=0.6,\n    top_p=0.9,\n)\nprint(outputs[0][\"generated_text\"][len(prompt):])<\/code><\/pre>\n\n\n\n
<\/p>\n\n\n\n
4.\u81ea\u5b9a\u4e49\u6570\u636e\u96c6\u6784\u5efa<\/h2>\n\n\n\n
\u6570\u636e\u96c6\u7684\u683c\u5f0f\u8981\u6c42\u5728\u4e0d\u540c\u7684\u9636\u6bb5\u662f\u4e0d\u540c\u7684\uff0c\u672c\u6559\u7a0b\u4ee5sft\u9636\u6bb5\u7684\u6570\u636e\u96c6\u9700\u6c42\uff0c\u5c06\u4ee5\u7cfb\u7edf\u81ea\u5e26\u7684identity\u6570\u636e\u96c6\u548c\u5c06\u81ea\u5b9a\u4e49\u7684\u4e00\u4e2a\u5546\u54c1\u6587\u6848\u751f\u6210\u6570\u636e\u96c6\u4e3a\u4f8b\uff0c\u4ecb\u7ecd\u6570\u636e\u96c6\u7684\u4f7f\u7528\u3002\u66f4\u591a\u8be6\u60c5\u53ef\u4ee5\u5728 https:\/\/github.com\/hiyouga\/LLaMA-Factory\/blob\/main\/data\/README_zh.md<\/a> \u4e2d\u627e\u5230\u76f8\u5173\u89e3\u91ca\u3002<\/p>\n\n\n\n
\u7cfb\u7edf\u76ee\u524d\u652f\u6301 alpaca<\/a> \u548csharegpt\u4e24\u79cd\u6570\u636e\u683c\u5f0f\uff0c\u4ee5alpaca\u4e3a\u4f8b\uff0c\u6574\u4e2a\u6570\u636e\u96c6\u662f\u4e00\u4e2ajson\u5bf9\u8c61\u7684list\uff0c\u5177\u4f53\u6570\u636e\u683c\u5f0f\u4e3a<\/p>\n\n\n\n
[\n  {\n    \"instruction\": \"\u7528\u6237\u6307\u4ee4\uff08\u5fc5\u586b\uff09\",\n    \"input\": \"\u7528\u6237\u8f93\u5165\uff08\u9009\u586b\uff09\",\n    \"output\": \"\u6a21\u578b\u56de\u7b54\uff08\u5fc5\u586b\uff09\",\n    \"system\": \"\u7cfb\u7edf\u63d0\u793a\u8bcd\uff08\u9009\u586b\uff09\",\n    \"history\": [\n      [\"\u7b2c\u4e00\u8f6e\u6307\u4ee4\uff08\u9009\u586b\uff09\", \"\u7b2c\u4e00\u8f6e\u56de\u7b54\uff08\u9009\u586b\uff09\"],\n      [\"\u7b2c\u4e8c\u8f6e\u6307\u4ee4\uff08\u9009\u586b\uff09\", \"\u7b2c\u4e8c\u8f6e\u56de\u7b54\uff08\u9009\u586b\uff09\"]\n    ]\n  }\n]<\/code><\/pre>\n\n\n\n
\u4f8b\u5b50\u6bd4\u5982\u5355\u8f6e\uff08alpaca_data_zh_51k.json \u4e2d\u7684\u4f8b\u5b50, \u6570\u636e\u96c6\u5728data\/dataset_info.json\u4e2d\u6ce8\u518c\u4e3aalpaca_zh\uff09<\/p>\n\n\n\n
{\n  \"instruction\": \"\u5199\u4e00\u4e2a\u6709\u6548\u7684\u6bd4\u8f83\u8bed\u53e5\",\n  \"input\": \"\u7bee\u7403\u548c\u8db3\u7403\",\n  \"output\": \"\u7bee\u7403\u548c\u8db3\u7403\u90fd\u662f\u53d7\u6b22\u8fce\u7684\u8fd0\u52a8\u3002\"\n}<\/code><\/pre>\n\n\n\n
\u548c\u591a\u8f6e \uff08oaast_sft_zh.json \u4e2d\u7684\u4f8b\u5b50, \u6570\u636e\u96c6\u5728data\/dataset_info.json\u4e2d\u6ce8\u518c\u4e3aoaast_sft_zh\uff09<\/p>\n\n\n\n
{\n  \"instruction\": \"\u8c22\u8c22\",\n  \"input\": \"\",\n  \"output\": \"\u4e0d\u7528\u8c22! \u5f88\u9ad8\u5174\u6211\u63d0\u4f9b\u7684\u4fe1\u606f\u80fd\u591f\u5e2e\u52a9\u5230\u4f60! \u5982\u679c\u8fd8\u6709\u4ec0\u4e48\u5176\u4ed6\u95ee\u9898\u4e5f\u53ef\u4ee5\u5411\u6211\u63d0\u95ee\u3002\",\n  \"history\": [\n    [\n      \"\u8bf7\u4f60\u7ed9\u6211\u5199\u4e00\u4e2a\u9762\u8bd5\u51c6\u5907\u8ba1\u5212\uff0c\u6211\u60f3\u8981\u53bb\u9762\u8bd5\u5fae\u8f6f\u7684\u7a0b\u5e8f\u5458\u5c97\u4f4d\",\n      \"\u9996\u5148\uff0c\u4f60\u53ef\u4ee5\u53bb\u5fae\u8f6f\u5b98\u7f51\u5bfb\u627e\u62db\u8058\u4fe1\u606f\u5e76\u7533\u8bf7\u9762\u8bd5\u3002\\n\u5176\u6b21\uff0c\u60a8\u53ef\u4ee5\u5728\u793e\u4ea4\u5a92\u4f53\u5e73\u53f0\u5bfb\u627e\u5fae\u8f6f\u516c\u53f8\u5bf9\u7a0b\u5e8f\u5458\u7684\u9762\u8bd5\u95ee\u9898\uff0c\u5e76\u505a\u597d\u51c6\u5907\u3002\\n\u6700\u540e\uff0c\u60a8\u53ef\u4ee5\u81ea\u5df1\u5bf9\u9762\u8bd5\u8fc7\u7a0b\u8fdb\u884c\u6a21\u62df\uff0c\u719f\u6089\u8bdd\u9898\u5e76\u51cf\u5c11\u7d27\u5f20\u611f\u3002\\n\u6211\u5e0c\u671b\u4f60\u80fd\u9762\u8bd5\u6210\u529f\u3002\"\n    ]\n  ]\n}<\/code><\/pre>\n\n\n\n
\u6240\u4ee5\u6211\u4eec\u8bad\u7ec3\u7684\u6570\u636e\u6700\u597d\u4e5f\u8f6c\u6362\u6210\u8fd9\u79cd\u683c\u5f0f\uff0c\u7136\u540e\u5728 data\/dataset_info.json\u4e2d\u8fdb\u884c\u6ce8\u518c\uff08\u5982\u679c\u4e0d\u505a\u5b57\u6bb5\u540d\u79f0\u8f6c\u6362\uff0c\u5219\u9700\u8981\u5728\u6ce8\u518c\u7684\u65f6\u5019\u5728 columns\u5b57\u6bb5\u4e2d\u505a\u4e24\u4e2a\u6570\u636e\u7684\u6620\u5c04\u914d\u7f6e\uff09<\/p>\n\n\n\n
\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u4f7f\u7528\u4e24\u4e2a\u5177\u4f53\u7684\u4f8b\u5b50\u6765\u8bf4\u660e\u6570\u636e\u96c6\u7684\u4f7f\u7528<\/p>\n\n\n\n
\u7b2c\u4e00\u4e2a\u662f\u7cfb\u7edf\u81ea\u5e26\u7684identity.json\u6570\u636e\u96c6(\u5df2\u9ed8\u8ba4\u5728data\/dataset_info.json \u6ce8\u518c\u4e3aidentity)\uff0c\u5bf9\u5e94\u6587\u4ef6\u5df2\u7ecf\u5728data\u76ee\u5f55\u4e0b\uff0c\u6211\u4eec\u901a\u8fc7\u64cd\u4f5c\u7cfb\u7edf\u7684\u6587\u672c\u7f16\u8f91\u5668\u7684\u66ff\u6362\u529f\u80fd\uff0c\u53ef\u4ee5\u66ff\u6362\u5176\u4e2d\u7684NAME \u548c AUTHOR \uff0c\u6362\u6210\u6211\u4eec\u9700\u8981\u7684\u5185\u5bb9\u3002\u5982\u679c\u662flinux\u7cfb\u7edf\uff0c\u53ef\u4ee5\u4f7f\u7528sed <\/strong>\u5b8c\u6210\u5feb\u901f\u66ff\u6362\u3002\u6bd4\u5982\u52a9\u624b\u7684\u540d\u79f0\u4fee\u6539\u4e3aPonyBot\uff0c \u7531 LLaMA Factory \u5f00\u53d1<\/p>\n\n\n\n
sed -i 's\/{{name}}\/PonyBot\/g'  data\/identity.json \nsed -i 's\/{{author}}\/LLaMA Factory\/g'  data\/identity.json <\/code><\/pre>\n\n\n\n
\u66ff\u6362\u524d<\/p>\n\n\n\n
{\n  \"instruction\": \"Who are you?\",\n  \"input\": \"\",\n  \"output\": \"Hello! I am {{name}}, an AI assistant developed<\/a> by {{author}}. How can I assist you today?\"\n}<\/code><\/pre>\n\n\n\n
\u66ff\u6362\u540e<\/p>\n\n\n\n
{\n  \"instruction\": \"Who are you?\",\n  \"input\": \"\",\n  \"output\": \"I am PonyBot, an AI assistant developed by LLaMA Factory. How can I assist you today?\"\n}<\/code><\/pre>\n\n\n\n
\u7b2c\u4e8c\u4e2a\u662f\u4e00\u4e2a\u5546\u54c1\u6587\u6848\u751f\u6210\u6570\u636e\u96c6\uff0c\u539f\u59cb\u94fe\u63a5\u4e3a https:\/\/cloud.tsinghua.edu.cn\/f\/b3f119a008264b1cabd1\/?dl=1<\/a><\/p>\n\n\n\n
\u539f\u59cb\u683c\u5f0f\u5982\u4e0b\uff0c\u5f88\u660e\u663e\uff0c\u8bad\u7ec3\u76ee\u6807\u662f\u8f93\u5165content \uff08\u4e5f\u5c31\u662fprompt\uff09, \u8f93\u51fa summary \uff08\u5bf9\u5e94response\uff09<\/p>\n\n\n\n
{\n    \"content\": \"\u7c7b\u578b#\u88e4*\u7248\u578b#\u5bbd\u677e*\u98ce\u683c#\u6027\u611f*\u56fe\u6848#\u7ebf\u6761*\u88e4\u578b#\u9614\u817f\u88e4\", \n    \"summary\": \"\u5bbd\u677e\u7684\u9614\u817f\u88e4\u8fd9\u4e24\u5e74\u771f\u7684\u5438\u7c89\u4e0d\u5c11\uff0c\u660e\u661f\u65f6\u5c1a\u8fbe\u4eba\u7684\u5fc3\u5934\u7231\u3002\u6bd5\u7adf\u597d\u7a7f\u65f6\u5c1a\uff0c\u8c01\u90fd\u80fd\u7a7f\u51fa\u817f\u957f2\u7c73\u7684\u6548\u679c\u5bbd\u677e\u7684\u88e4\u817f\uff0c\u5f53\u7136\u662f\u906e\u8089\u5c0f\u80fd\u624b\u554a\u3002\u4e0a\u8eab\u968f\u6027\u81ea\u7136\u4e0d\u62d8\u675f\uff0c\u9762\u6599\u4eb2\u80a4\u8212\u9002\u8d34\u8eab\u4f53\u9a8c\u611f\u68d2\u68d2\u54d2\u3002\u7cfb\u5e26\u90e8\u5206\u589e\u52a0\u8bbe\u8ba1\u770b\u70b9\uff0c\u8fd8\u8ba9\u5355\u54c1\u7684\u8bbe\u8ba1\u611f\u66f4\u5f3a\u3002\u817f\u90e8\u7ebf\u6761\u82e5\u9690\u82e5\u73b0\u7684\uff0c\u6027\u611f\u64a9\u4eba\u3002\u989c\u8272\u6572\u6e29\u67d4\u7684\uff0c\u4e0e\u88e4\u5b50\u672c\u8eab\u6240\u5448\u73b0\u7684\u98ce\u683c\u6709\u70b9\u53cd\u5dee\u840c\u3002\"\n}<\/code><\/pre>\n\n\n\n
\u60f3\u5c06\u8be5\u81ea\u5b9a\u4e49\u6570\u636e\u96c6\u653e\u5230\u6211\u4eec\u7684\u7cfb\u7edf\u4e2d\u4f7f\u7528\uff0c\u5219\u9700\u8981\u8fdb\u884c\u5982\u4e0b\u4e24\u6b65\u64cd\u4f5c<\/p>\n\n\n\n
    \n
  1. \u590d\u5236\u8be5\u6570\u636e\u96c6\u5230 data\u76ee\u5f55\u4e0b<\/li>\n\n\n\n
  2. \u4fee\u6539 data\/dataset_info.json<\/a>\u00a0\u65b0\u52a0\u5185\u5bb9\u5b8c\u6210\u6ce8\u518c, \u8be5\u6ce8\u518c\u540c\u65f6\u5b8c\u6210\u4e863\u4ef6\u4e8b<\/li>\n<\/ol>\n\n\n\n