{"id":664,"date":"2026-04-06T14:58:35","date_gmt":"2026-04-06T06:58:35","guid":{"rendered":"https:\/\/www.liaoxinghui.com\/?p=664"},"modified":"2026-04-06T14:58:53","modified_gmt":"2026-04-06T06:58:53","slug":"fine-tuning-data-pipeline-pitfalls","status":"publish","type":"post","link":"https:\/\/www.liaoxinghui.com\/?p=664","title":{"rendered":"\u5fae\u8c03\u6570\u636e\u51c6\u5907\u7684\u5751\uff1a\u4ece\u539f\u59cb\u6587\u672c\u5230\u8bad\u7ec3\u6570\u636e\u96c6\u7684\u5b8c\u6574\u6d41\u6c34\u7ebf"},"content":{"rendered":"

\u6280\u672f\u540c\u4e8b\u7535\u8bdd\u6253\u8fc7\u6765\u7684\u65f6\u5019\uff0c\u6211\u6b63\u76ef\u7740\u5c4f\u5e55\u53d1\u5446\u3002\u4ed6\u8bf4\u5fae\u8c03\u540e\u7684\u6a21\u578b\u8f93\u51fa\u5168\u662f\u4e71\u7801\uff0c\u6839\u672c\u6ca1\u6cd5\u7528\u3002\u6211\u5f53\u65f6\u5fc3\u91cc\u4e00\u6c89\uff0c\u8fd9\u9879\u76ee\u5df2\u7ecf\u62d6\u4e86\u4e24\u5468\uff0c\u518d\u51fa\u95ee\u9898\u4f30\u8ba1\u5f97\u6328\u9a82\u3002\u6211\u8d76\u7d27\u767b\u5f55\u670d\u52a1\u5668\uff0c\u4e00\u770b\u8bad\u7ec3\u65e5\u5fd7\uff0c\u679c\u7136\uff0c\u6570\u636e\u683c\u5f0f\u9519\u4e86\u2014\u2014\u6a21\u578b\u538b\u6839\u6ca1\u770b\u61c2\u8f93\u5165\u3002<\/p>\n

\u4e1a\u52a1\u573a\u666f<\/h2>\n

\u6211\u4eec\u56e2\u961f\u5f53\u65f6\u63a5\u4e86\u4e2a\u6d3b\u513f\uff0c\u8981\u628a\u516c\u53f8\u5185\u90e8\u7684\u5927\u91cf\u975e\u7ed3\u6784\u5316\u6587\u6863\uff08\u6bd4\u5982PDF\u62a5\u544a\u3001Word\u6587\u6863\uff09\u8f6c\u6362\u6210\u5927\u6a21\u578b\u5fae\u8c03\u53ef\u7528\u7684\u8bad\u7ec3\u6570\u636e\u3002\u5ba2\u6237\u60f3\u7528\u8fd9\u4e9b\u6570\u636e\u5fae\u8c03\u4e00\u4e2aLlama 3\u6a21\u578b\uff0c\u505a\u667a\u80fd\u95ee\u7b54\u3002\u542c\u8d77\u6765\u633a\u7b80\u5355\uff0c\u5bf9\u5427\uff1f\u5176\u5b9e\u5751\u591a\u5f97\u8981\u547d\u3002\u6570\u636e\u6765\u6e90\u4e94\u82b1\u516b\u95e8\uff0c\u6709\u626b\u63cf\u7684PDF\u3001\u8001\u65e7\u7684Word 2003\u6587\u4ef6\uff0c\u8fd8\u6709\u4e00\u5806Excel\u8868\u683c\u3002\u6211\u5f97\u628a\u8fd9\u4e9b\u4e1c\u897f\u7edf\u4e00\u6210\u6587\u672c\uff0c\u518d\u6807\u6ce8\u6210\u95ee\u7b54\u5bf9\uff0c\u6700\u540e\u8f6c\u6210\u6a21\u578b\u80fd\u5403\u7684\u683c\u5f0f\u3002<\/p>\n

\u6570\u636e\u8bf4\u660e<\/h2>\n

\u8bf4\u5b9e\u8bdd\uff0c\u6570\u636e\u91cf\u771f\u4e0d\u5c0f\u3002\u539f\u59cb\u6587\u6863\u5927\u6982\u67095000\u591a\u4e2a\u6587\u4ef6\uff0c\u5305\u62ec2000\u591a\u4efdPDF\uff08\u5176\u4e2d30%\u662f\u626b\u63cf\u4ef6\uff09\u30011500\u4e2aWord\u6587\u6863\uff08\u6709\u4e9b\u8fd8\u662f.doc\u8001\u683c\u5f0f\uff09\u3001\u5269\u4e0b\u7684\u662f\u4e00\u5806Excel\u548cPPT\u3002\u6587\u672c\u63d0\u53d6\u51fa\u6765\uff0c\u539f\u59cb\u5927\u5c0f\u63a5\u8fd150GB\u3002\u6e05\u6d17\u540e\uff0c\u6211\u4eec\u5f97\u5230\u4e8610\u4e07\u6761\u6587\u672c\u7247\u6bb5\uff0c\u6bcf\u6761\u5e73\u5747500\u5b57\u7b26\u3002\u6807\u6ce8\u76ee\u6807\u662f\u628a\u8fd9\u4e9b\u8f6c\u6362\u6210\u95ee\u7b54\u5bf9\uff0c\u6700\u7ec8\u6211\u4eec\u751f\u6210\u4e868\u4e07\u6761\uff0c\u7528\u4e8eLlama 3\u5fae\u8c03\u3002\u6570\u636e\u5206\u5e03\u4e0a\uff0c\u6280\u672f\u6587\u6863\u536060%\uff0c\u4e1a\u52a1\u62a5\u544a30%\uff0c\u5176\u4ed6\u6742\u987910%\u3002\u8fd9\u4e2a\u6bd4\u4f8b\u662f\u5ba2\u6237\u5b9a\u7684\uff0c\u4f46\u6211\u89c9\u5f97\u6280\u672f\u90e8\u5206\u53ef\u80fd\u6807\u5f97\u4e0d\u591f\u7ec6\uff0c\u540e\u9762\u4f1a\u8bf4\u5230\u3002<\/p>\n

\u4e00\u5f00\u59cb\uff0c\u6211\u4eec\u76f4\u63a5\u7528Python\u7684pdfplumber<\/code>\u548cpython-docx<\/code>\u5e93\u63d0\u53d6\u6587\u672c\u3002\u4ee3\u7801\u5199\u8d77\u6765\u633a\u5feb\uff0c\u4f46\u8dd1\u8d77\u6765\u5c31\u51fa\u95ee\u9898\u4e86\u3002\u6709\u4e9bPDF\u662f\u626b\u63cf\u4ef6\uff0cpdfplumber<\/code>\u8bfb\u51fa\u6765\u5168\u662f\u4e71\u7801\u3002\u5f53\u65f6\u7684\u505a\u6cd5\u9519\u4e86\uff0c\u4ee5\u4e3a\u6362\u4e2a\u5e93\u5c31\u884c\uff0c\u8bd5\u4e86PyPDF2<\/code>\u548cpdfminer<\/code>\uff0c\u7ed3\u679c\u5dee\u4e0d\u591a\u3002\u540e\u6765\u624d\u53d1\u73b0\uff0c\u5f97\u5148\u7528OCR\u5904\u7406\u3002\u63a5\u7740\u88c5\u4e86pytesseract<\/code>\uff0c\u914d\u5408pdf2image<\/code>\u628aPDF\u8f6c\u6210\u56fe\u7247\u518d\u8bc6\u522b\u3002\u8fd9\u6b65\u82b1\u4e86\u5927\u698240\u5206\u949f\u8c03\u8bd5\u53c2\u6570\uff0c\u56e0\u4e3a\u9ed8\u8ba4\u7684OCR\u7cbe\u5ea6\u4e0d\u591f\uff0c\u8bc6\u522b\u51fa\u6765\u7684\u6587\u672c\u9519\u522b\u5b57\u4e00\u5806\u3002<\/p>\n

# \u7528pytesseract\u505aOCR\uff0c\u63d0\u53d6\u626b\u63cfPDF\u7684\u6587\u672c\nimport pytesseract\nfrom pdf2image import convert_from_path\n\ndef extract_text_from_scanned_pdf(pdf_path):\n    images = convert_from_path(pdf_path, dpi=300)  # dpi\u8c03\u9ad8\uff0c\u8bc6\u522b\u66f4\u51c6\n    text = ''\n    for img in images:\n        # \u8fd9\u91cc\u6211\u8bd5\u8fc7\u4e0d\u540c\u914d\u7f6e\uff0c\u9ed8\u8ba4\u7684\u82f1\u8bed\u8bc6\u522b\u6548\u679c\u4e00\u822c\n        custom_config = r'--oem 3 --psm 6'  # OEM 3\u7528LSTM\u5f15\u64ce\uff0cPSM 6\u5047\u8bbe\u7edf\u4e00\u533a\u5757\n        page_text = pytesseract.image_to_string(img, config=custom_config)\n        text += page_text + '\\n'\n    return text\n\n# \u5b9e\u9645\u8dd1\u7684\u65f6\u5019\uff0c\u4e00\u4e2a100\u9875\u7684PDF\u82b1\u4e865\u5206\u949f\uff0cCPU\u98d9\u523080%\n# \u9519\u8bef\u65e5\u5fd7\u7247\u6bb5\uff1aTesseractError: Error opening data file \/usr\/share\/tesseract-ocr\/4.00\/tessdata\/eng.traineddata\n# \u89e3\u51b3\uff1a\u5f97\u5148\u5b89\u88c5tesseract-ocr\u548c\u8bed\u8a00\u5305\uff0csudo apt-get install tesseract-ocr tesseract-ocr-eng<\/code><\/pre>\n
\u6587\u672c\u63d0\u53d6\u5b8c\u4e86\uff0c\u4e0b\u4e00\u6b65\u662f\u6e05\u6d17\u3002\u8fd9\u6b65\u66f4\u5934\u75bc\u3002\u539f\u59cb\u6587\u6863\u91cc\u6709\u5927\u91cf\u65e0\u5173\u5185\u5bb9\uff0c\u6bd4\u5982\u9875\u7709\u3001\u9875\u811a\u3001\u8868\u683c\u7ebf\uff0c\u8fd8\u6709\u91cd\u590d\u6bb5\u843d\u3002\u6211\u4eec\u5199\u4e86\u4e2a\u6e05\u6d17\u811a\u672c\uff0c\u7528\u6b63\u5219\u8868\u8fbe\u5f0f\u5339\u914d\u548c\u5220\u9664\u3002\u4f46\u6b63\u5219\u8fd9\u4e1c\u897f\uff0c\u5199\u8d77\u6765\u5bb9\u6613\uff0c\u8c03\u8bd5\u8d77\u6765\u8981\u547d\u3002\u5dee\u70b9\u6ca1\u628a\u81ea\u5df1\u7ed5\u8fdb\u53bb\uff0c\u56e0\u4e3a\u6709\u4e9b\u6587\u6863\u7684\u683c\u5f0f\u4e0d\u7edf\u4e00\uff0c\u4e00\u4e2a\u6b63\u5219\u8986\u76d6\u4e0d\u4e86\u6240\u6709\u60c5\u51b5\u3002\u6bd4\u5982\uff0c\u5220\u9664\u7a7a\u884c\uff0c\u4e00\u5f00\u59cb\u7528^\\s*$<\/code>\uff0c\u7ed3\u679c\u628a\u4e00\u4e9b\u7f29\u8fdb\u7684\u4ee3\u7801\u6bb5\u4e5f\u5220\u4e86\u3002\u540e\u6765\u6539\u6210\u66f4\u4fdd\u5b88\u7684\u65b9\u6cd5\uff0c\u53ea\u5220\u7eaf\u7a7a\u884c\u3002<\/p>\n
\u6e05\u6d17\u5b8c\u7684\u6570\u636e\uff0c\u5927\u6982\u670910\u4e07\u6761\u6587\u672c\u7247\u6bb5\u3002\u63a5\u4e0b\u6765\u662f\u6807\u6ce8\u6210\u95ee\u7b54\u5bf9\u3002\u5ba2\u6237\u63d0\u4f9b\u4e86\u4e00\u4e9b\u793a\u4f8b\u95ee\u9898\uff0c\u4f46\u4e0d\u591f\u7528\u3002\u6211\u4eec\u56e2\u961f\u624b\u52a8\u6807\u4e861000\u6761\uff0c\u7136\u540e\u60f3\u7528\u5927\u6a21\u578b\u81ea\u52a8\u6269\u5c55\u3002\u6211\u7528GPT-4 API\u751f\u6210\u4e86\u66f4\u591a\u95ee\u7b54\u5bf9\uff0c\u4f46\u8fd9\u91cc\u6709\u4e2a\u5927\u5751\uff1a\u751f\u6210\u7684\u6570\u636e\u6709\u504f\u89c1\uff0c\u6a21\u578b\u5bb9\u6613\u91cd\u590d\u76f8\u4f3c\u7684\u95ee\u9898\u3002\u6211\u8bd5\u4e86\u4e0d\u540c\u7684\u63d0\u793a\u8bcd\uff0c\u6bd4\u5982\u201c\u751f\u6210\u591a\u6837\u5316\u7684\u95ee\u7b54\u5bf9\u201d\uff0c\u6548\u679c\u8fd8\u662f\u4e0d\u884c\u3002\u6700\u540e\uff0c\u6211\u7ed3\u5408\u89c4\u5219\u548c\u6a21\u578b\uff0c\u5148\u63d0\u53d6\u5173\u952e\u5b9e\u4f53\uff0c\u518d\u56f4\u7ed5\u5b9e\u4f53\u751f\u6210\u95ee\u9898\u3002<\/p>\n
\u53c2\u6570\u8bf4\u660e<\/h2>\n
\u8fd9\u6b65\u7684\u4ee3\u7801\u6211\u653e\u5728\u4e0b\u9762\uff0c\u53c2\u6570\u8c03\u4e86\u597d\u51e0\u6b21\u3002\u6211\u7528\u7684\u662fGPT-4 API\uff0c\u5177\u4f53\u914d\u7f6e\u5982\u4e0b\uff1a<\/p>\n\n\n\n\n\n\n\n\n\n
\u53c2\u6570<\/th>\n\u503c<\/th>\n\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n
model<\/td>\ngpt-4-turbo<\/td>\n\u6211\u7528\u8fd9\u4e2a\u7248\u672c\uff0c\u751f\u6210\u8d28\u91cf\u7a33\u5b9a\u70b9\uff0c\u867d\u7136\u8d35\u4f46\u503c<\/td>\n<\/tr>\n
temperature<\/td>\n0.7<\/td>\n\u8c03\u9ad8\u4e86\u70b9\uff0c\u907f\u514d\u8f93\u51fa\u592a\u6b7b\u677f\uff0c\u8bd5\u8fc70.5\u548c0.9\uff0c0.7\u6298\u4e2d<\/td>\n<\/tr>\n
max_tokens<\/td>\n500<\/td>\n\u9650\u5236\u957f\u5ea6\uff0c\u4e0d\u7136API\u8d35\u6b7b\uff0c\u5b9e\u9645\u5e73\u5747\u8f93\u51fa300 token<\/td>\n<\/tr>\n
top_p<\/td>\n1.0<\/td>\n\u9ed8\u8ba4\u503c\uff0c\u6ca1\u52a8\u8fc7\uff0c\u6211\u89c9\u5f97\u591f\u7528\u4e86<\/td>\n<\/tr>\n
frequency_penalty<\/td>\n0.0<\/td>\n\u6ca1\u52a0\u60e9\u7f5a\uff0c\u6015\u5f71\u54cd\u591a\u6837\u6027<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\u6807\u6ce8\u6570\u636e\u6512\u52305\u4e07\u6761\uff0c\u6211\u5f00\u59cb\u505a\u8d28\u91cf\u68c0\u67e5\u3002\u7b2c\u4e00\u6b21\u6ca1\u6210\u529f\uff0c\u56e0\u4e3a\u6709\u4e9b\u95ee\u7b54\u5bf9\u4e0d\u5339\u914d\u2014\u2014\u95ee\u9898\u95ee\u7684\u662fA\uff0c\u7b54\u6848\u8bb2\u7684\u662fB\u3002\u6211\u5199\u4e86\u4e2a\u68c0\u67e5\u811a\u672c\uff0c\u7528\u53e5\u5b50\u76f8\u4f3c\u5ea6\u8ba1\u7b97\uff08\u6bd4\u5982\u7528sentence-transformers<\/code>\u5e93\uff09\u8fc7\u6ee4\u4f4e\u5206\u5bf9\u3002\u4f46\u8ba1\u7b97\u91cf\u592a\u5927\uff0c5\u4e07\u6761\u6570\u636e\u8dd1\u4e0b\u6765\u89812\u5c0f\u65f6\u3002\u6211\u4f18\u5316\u4e86\u4e00\u4e0b\uff0c\u53ea\u62bd\u6837\u68c0\u67e5\uff0c\u7136\u540e\u4eba\u5de5\u590d\u6838\u3002\u8fd9\u91cc\u6211\u4e00\u822c\u4e0d\u7528\u5168\u81ea\u52a8\uff0c\u5bb9\u6613\u6f0f\u6389\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u9519\u8bef\u3002<\/p>\n
\u8c03\u7528\u65b9\u5f0f<\/h2>\n
\u6574\u4e2a\u6d41\u6c34\u7ebf\uff0c\u5c01\u88c5\u6210\u4e86\u51e0\u4e2aPython\u811a\u672c\uff0c\u901a\u8fc7\u547d\u4ee4\u884c\u8c03\u7528\u3002\u6211\u4eec\u4e00\u822c\u7528\u6a21\u5757\u5316\u8bbe\u8ba1\uff0c\u6bcf\u4e2a\u6b65\u9aa4\u4e00\u4e2a\u811a\u672c\uff0c\u8fd9\u6837\u597d\u8c03\u8bd5\u3002\u4e3b\u6d41\u7a0b\u5927\u6982\u8fd9\u6837\u8c03\u7528\uff1a<\/p>\n
# 1. \u63d0\u53d6\u6587\u672c\npython extract_text.py --input_dir .\/raw_docs --output_dir .\/extracted_text\n# \u8f93\u51fa\uff1a\u6bcf\u4e2a\u6587\u4ef6\u4e00\u4e2a.txt\uff0c\u653e\u5728extracted_text\u76ee\u5f55\n\n# 2. \u6e05\u6d17\u6570\u636e\npython clean_text.py --input_dir .\/extracted_text --output_dir .\/cleaned_text\n# \u8fd9\u91cc\u4f1a\u5220\u6389\u9875\u7709\u9875\u811a\uff0c\u5904\u7406\u7f16\u7801\u95ee\u9898\n\n# 3. \u6807\u6ce8\u95ee\u7b54\u5bf9\uff08\u534a\u81ea\u52a8\uff09\npython annotate_qa.py --input_dir .\/cleaned_text --output_file .\/qa_pairs.jsonl --use_gpt true\n# \u7528GPT-4\u751f\u6210\uff0c\u7ed3\u5408\u624b\u52a8\u6807\u6ce8\uff0c\u8f93\u51faJSONL\u683c\u5f0f\n\n# 4. \u8d28\u91cf\u68c0\u67e5\npython quality_check.py --input_file .\/qa_pairs.jsonl --output_file .\/filtered_qa.jsonl\n# \u57fa\u4e8e\u76f8\u4f3c\u5ea6\u8fc7\u6ee4\uff0c\u9608\u503c\u8bbe0.7\uff0c\u4f4e\u4e8e\u7684\u4e22\u5230review\u76ee\u5f55\u4eba\u5de5\u770b\n\n# 5. \u8f6c\u8bad\u7ec3\u683c\u5f0f\npython convert_to_training.py --input_file .\/filtered_qa.jsonl --output_file .\/train_data.jsonl\n# \u8f6c\u6210Llama 3\u9700\u8981\u7684instruction\/input\/output\u683c\u5f0f<\/code><\/pre>\n
\u6bcf\u4e2a\u811a\u672c\u90fd\u6709--help<\/code>\u770b\u53c2\u6570\uff0c\u4e60\u60ef\u8fd9\u4e48\u5199\uff0c\u514d\u5f97\u81ea\u5df1\u5fd8\u4e86\u3002\u8dd1\u5168\u6d41\u7a0b\u7684\u8bdd\uff0c\u7528\u4e2ashell\u811a\u672c\u4e32\u8d77\u6765\uff0c\u4f46\u5efa\u8bae\u5206\u6b65\u8dd1\uff0c\u56e0\u4e3a\u4e2d\u95f4\u53ef\u80fd\u51fa\u9519\u8981\u56de\u6eda\u3002<\/p>\n
\u6570\u636e\u51c6\u5907\u597d\u4e86\uff0c\u5f97\u8f6c\u6210\u8bad\u7ec3\u683c\u5f0f\u3002\u6211\u4eec\u5fae\u8c03\u7684\u662fLlama 3\u6a21\u578b\uff0c\u9700\u8981JSONL\u683c\u5f0f\uff0c\u6bcf\u6761\u8bb0\u5f55\u5305\u542binstruction<\/code>\u3001input<\/code>\u3001output<\/code>\u3002\u6211\u4e00\u5f00\u59cb\u7528\u7b80\u5355\u7684\u5b57\u5178\u8f6cJSON\uff0c\u4f46\u53d1\u73b0\u7f16\u7801\u95ee\u9898\u2014\u2014\u6709\u4e9b\u6587\u672c\u5305\u542b\u7279\u6b8a\u5b57\u7b26\uff0cJSON\u5e8f\u5217\u5316\u5931\u8d25\u3002\u9519\u8bef\u65e5\u5fd7\u957f\u8fd9\u6837\uff1aUnicodeEncodeError: 'ascii' codec can't encode character '\\u2019' in position 42<\/code>\u3002\u89e3\u51b3\u65b9\u6cd5\u662f\u5f3a\u5236\u7528UTF-8\u7f16\u7801\u3002<\/p>\n
# \u6362\u6570\u636e\u4e3aJSONL\u683c\u5f0f\uff0c\u7528\u4e8eLlama 3\u5fae\u8c03\nimport json\n\ndef convert_to_jsonl(data_list, output_path):\n    with open(output_path, 'w', encoding='utf-8') as f:\n        for item in data_list:\n            # \u8fd9\u91cc\u5f97\u5904\u7406\u7279\u6b8a\u5b57\u7b26\uff0c\u4e0d\u7136\u5199\u5165\u4f1a\u62a5\u9519\n            record = {\n                "instruction": item["question"],\n                "input": "",  # \u6211\u4eec\u573a\u666f\u4e0d\u9700\u8981\u989d\u5916\u8f93\u5165\n                "output": item["answer"]\n            }\n            json_line = json.dumps(record, ensure_ascii=False)  # \u5173\u952e\uff1aensure_ascii=False\n            f.write(json_line + '\\n')\n\n# \u8dd1\u7684\u65f6\u5019\uff0c\u4e00\u4e2a\u6587\u4ef6\u5927\u69821GB\uff0c\u6211\u5206\u4e865\u4e2a\u6279\u6b21\u5904\u7406\uff0c\u907f\u514d\u5185\u5b58\u7206\u6389<\/code><\/pre>\n
\u6574\u4e2a\u6d41\u6c34\u7ebf\u8dd1\u5b8c\uff0c\u6211\u751f\u6210\u4e86\u6700\u7ec8\u7684\u6570\u636e\u96c6\uff0c\u5927\u69828\u4e07\u6761\u95ee\u7b54\u5bf9\u3002\u8bad\u7ec3\u524d\uff0c\u6211\u505a\u4e86\u4e2a\u7b80\u5355\u7edf\u8ba1\uff1a\u6570\u636e\u957f\u5ea6\u5206\u5e03\u3001\u95ee\u9898\u7c7b\u578b\u6bd4\u4f8b\u3002\u7136\u540e\u5f00\u59cb\u5fae\u8c03\u3002\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u76ef\u7740loss\u66f2\u7ebf\u2014\u2014\u524d\u51e0\u4e2aepoch\u964d\u5f97\u5f88\u5feb\uff0c\u4f46\u5230\u7b2c10\u8f6e\u5de6\u53f3\u5c31\u5e73\u4e86\u3002\u6211\u5f53\u65f6\u4ee5\u4e3a\u8fc7\u62df\u5408\u4e86\uff0c\u8d76\u7d27\u505c\u4e86\u8bad\u7ec3\uff0c\u7528\u9a8c\u8bc1\u96c6\u68c0\u67e5\u3002\u7ed3\u679c\u53d1\u73b0\uff0c\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u4e0a\u8868\u73b0\u5f88\u597d\uff0c\u4f46\u65b0\u95ee\u9898\u7b54\u5f97\u4e00\u822c\u3002\u590d\u76d8\u65f6\u53d1\u73b0\uff0c\u6570\u636e\u591a\u6837\u6027\u8fd8\u662f\u4e0d\u591f\uff0c\u53ef\u80fd\u5f97\u52a0\u66f4\u591a\u6765\u6e90\u3002<\/p>\n
\u4e0a\u7ebf\u540e\u8bc4\u4f30<\/h2>\n
\u6a21\u578b\u4e0a\u7ebf\u540e\uff0c\u6211\u4eec\u8bbe\u4e86\u51e0\u4e2a\u89c2\u5bdf\u6307\u6807\uff1a\u56de\u7b54\u51c6\u786e\u7387\u3001\u54cd\u5e94\u65f6\u95f4\u3001\u7528\u6237\u53cd\u9988\u7387\u3002\u5934\u4e24\u5929\u8fd8\u884c\uff0c\u51c6\u786e\u7387\u5927\u698285%\uff0c\u54cd\u5e94\u65f6\u95f4\u5e73\u57472\u79d2\u3002\u4f46\u7b2c\u4e09\u5929\u6709\u4e2a\u7528\u6237\u53cd\u9988\uff0c\u6a21\u578b\u5bf9\u6280\u672f\u672f\u8bed\u89e3\u91ca\u4e0d\u6e05\uff0c\u6bd4\u5982\u201cKubernetes Pod\u7f51\u7edc\u7b56\u7565\u201d\u8fd9\u79cd\uff0c\u56de\u7b54\u5f88\u6a21\u7cca\u3002\u6211\u4e00\u770b\u65e5\u5fd7\uff0c\u90a3\u90e8\u5206\u6570\u636e\u6211\u4eec\u6807\u5f97\u5c11\uff0c\u8bad\u7ec3\u96c6\u91cc\u76f8\u5173\u95ee\u7b54\u5bf9\u4e0d\u5230100\u6761\u3002\u6240\u4ee5\u8d76\u7d27\u8865\u6570\u636e\uff0c\u91cd\u65b0\u5fae\u8c03\u4e86\u4e00\u7248\u3002<\/p>\n
\u8bc4\u4f30\u65f6\uff0c\u6211\u7528\u4e86\u4e00\u4e2a\u6d4b\u8bd5\u96c6500\u6761\u95ee\u9898\uff0c\u4eba\u5de5\u6253\u5206\u3002\u7b2c\u4e00\u7248\u5f97\u520670\/100\uff0c\u4e3b\u8981\u6263\u5728\u6280\u672f\u7ec6\u8282\u4e0a\u3002\u7b2c\u4e8c\u7248\u8865\u6570\u636e\u540e\uff0c\u63d0\u523085\/100\u3002\u4f46\u8bf4\u5b9e\u8bdd\uff0c\u8fd9\u5206\u6570\u53ef\u80fd\u865a\u9ad8\uff0c\u56e0\u4e3a\u6d4b\u8bd5\u96c6\u6ca1\u8986\u76d6\u6240\u6709\u573a\u666f\u3002\u6211\u73b0\u5728\u6bcf\u5468\u8dd1\u4e00\u6b21\u62bd\u6837\u68c0\u67e5\uff0c\u6301\u7eed\u4f18\u5316\u6570\u636e\u3002<\/p>\n
\u5e38\u89c1\u5751<\/h2>\n
\u8e29\u4e86\u8fd9\u4e48\u591a\u5751\uff0c\u6211\u603b\u7ed3\u51e0\u4e2a\u5e38\u89c1\u7684\uff0c\u4f60\u9047\u5230\u4e86\u53ef\u80fd\u7701\u70b9\u65f6\u95f4\uff1a<\/p>\n
    \n
  1. \u7f16\u7801\u95ee\u9898<\/strong>\uff1a\u6587\u672c\u91cc\u7684\u7279\u6b8a\u5b57\u7b26\uff08\u6bd4\u5982\u5f2f\u5f15\u53f7\u3001emoji\uff09\u5904\u7406\u4e0d\u597d\uff0cJSON\u5e8f\u5217\u5316\u76f4\u63a5\u5d29\u3002\u4e00\u5b9a\u8981\u7528ensure_ascii=False<\/code>\u548cUTF-8\u7f16\u7801\u3002<\/li>\n
  2. OCR\u7cbe\u5ea6<\/strong>\uff1a\u626b\u63cfPDF\u522b\u6307\u671b\u4e00\u6b21\u641e\u5b9a\uff0cdpi\u8c03\u9ad8\u3001\u8bed\u8a00\u5305\u88c5\u5168\uff0c\u8fd8\u5f97\u540e\u5904\u7406\u7ea0\u9519\u3002\u6211\u8bd5\u8fc7\u7528spellchecker<\/code>\u5e93\uff0c\u4f46\u6548\u679c\u4e00\u822c\uff0c\u540e\u6765\u4e3b\u8981\u9760\u89c4\u5219\u8fc7\u6ee4\u3002<\/li>\n
  3. \u6807\u6ce8\u504f\u89c1<\/strong>\uff1a\u7528\u5927\u6a21\u578b\u751f\u6210\u6570\u636e\uff0c\u5bb9\u6613\u91cd\u590d\u76f8\u4f3c\u6a21\u5f0f\u3002\u6211\u73b0\u5728\u7684\u505a\u6cd5\u662f\u6df7\u5408\u6765\u6e90\u2014\u2014\u624b\u52a8\u6807\u4e00\u4e9b\uff0c\u89c4\u5219\u751f\u6210\u4e00\u4e9b\uff0c\u6a21\u578b\u751f\u6210\u4e00\u4e9b\uff0c\u518d\u6253\u4e71\u3002<\/li>\n
  4. \u6570\u636e\u91cf\u592a\u5927<\/strong>\uff1a\u4e00\u6b21\u6027\u5904\u7406\u6240\u6709\u6570\u636e\uff0c\u5185\u5b58\u548cCPU\u90fd\u625b\u4e0d\u4f4f\u3002\u6211\u5206\u4e86\u6279\u6b21\uff0c\u7528multiprocessing<\/code>\u5e76\u884c\uff0c\u4f46\u8c03\u8bd5\u66f4\u9ebb\u70e6\u3002\u5efa\u8bae\u5148\u5c0f\u89c4\u6a21\u8dd1\u901a\u3002<\/li>\n
  5. \u683c\u5f0f\u8f6c\u6362\u9519\u8bef<\/strong>\uff1a\u4e0d\u540c\u6a21\u578b\u8981\u4e0d\u540c\u683c\u5f0f\uff08\u6bd4\u5982Alpaca\u3001Llama\u3001ChatGLM\uff09\uff0c\u8f6c\u6362\u811a\u672c\u5f97\u6d4b\u8bd5\u900f\u3002\u6211\u7b2c\u4e00\u6b21\u5c31\u628ainput<\/code>\u5b57\u6bb5\u6f0f\u4e86\uff0c\u6a21\u578b\u8bad\u7ec3\u62a5\u9519\u3002<\/li>\n<\/ol>\n
    \u5de5\u5177\u4e0a\uff0cpytesseract<\/code>\u770b\u7740\u571f\u4f46\u597d\u7528\uff0csentence-transformers<\/code>\u505a\u8d28\u91cf\u68c0\u67e5\u633a\u987a\u624b\u3002\u5982\u679c\u4f60\u8981\u505a\u7c7b\u4f3c\u7684\u4e8b\u60c5\uff0c\u6211\u7684\u5efa\u8bae\u662f\uff1a\u5148\u5c0f\u89c4\u6a21\u8bd5\u8dd1\uff0c\u522b\u4e00\u6b21\u6027\u5904\u7406\u6240\u6709\u6570\u636e\uff1b\u591a\u7559\u70b9\u65f6\u95f4\u7ed9\u6e05\u6d17\u548c\u6807\u6ce8\uff0c\u8fd9\u6b65\u6700\u5bb9\u6613\u51fa\u9519\uff1b\u683c\u5f0f\u8f6c\u6362\u65f6\uff0c\u7f16\u7801\u95ee\u9898\u5f97\u63d0\u524d\u60f3\u5230\u3002\u5927\u6982\u5c31\u8fd9\u6837\uff0c\u53ef\u80fd\u8981\u770b\u4f60\u7684\u5177\u4f53\u6570\u636e\u60c5\u51b5\u3002<\/p>","protected":false},"excerpt":{"rendered":"
    \u6280\u672f\u540c\u4e8b\u7535\u8bdd\u6253\u8fc7\u6765\uff0c\u8bf4\u5fae\u8c03\u540e\u7684\u6a21\u578b\u8f93\u51fa\u5168\u662f\u4e71\u7801\u3002\u6211\u4e00\u770b\uff0c\u8bad\u7ec3\u6570\u636e\u683c\u5f0f\u9519\u4e86\uff0c\u5dee\u70b9\u6ca1\u628a\u81ea\u5df1\u7ed5\u8fdb\u53bb\u3002\u8fd9\u7bc7\u6587\u7ae0\u8bb0\u5f55\u4e86\u6211\u4ece\u539f\u59cb\u6587\u6863\u5230\u8bad\u7ec3\u6570\u636e\u96c6\u7684\u5b8c\u6574\u6d41\u6c34\u7ebf\uff0c\u5305\u62ec\u6570\u636e\u6e05\u6d17\u3001\u6807\u6ce8\u3001\u8d28\u91cf\u68c0\u67e5\u7684\u5751\uff0c\u8fd8\u6709\u53ef\u590d\u7528\u7684\u4ee3\u7801\u3002<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[401],"tags":[321,406,18,319,405],"class_list":["post-664","post","type-post","status-publish","format-standard","hentry","category-ai","tag-llama-3","tag-ocr","tag-python","tag-319","tag-405"],"views":6,"_links":{"self":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/664","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=664"}],"version-history":[{"count":1,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/664\/revisions"}],"predecessor-version":[{"id":674,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/664\/revisions\/674"}],"wp:attachment":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=664"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=664"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=664"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}