{"id":9456,"date":"2025-10-30T17:50:56","date_gmt":"2025-10-30T09:50:56","guid":{"rendered":"\/?p=9456"},"modified":"2025-10-30T17:50:56","modified_gmt":"2025-10-30T09:50:56","slug":"%e5%a6%82%e6%9e%9c%e4%bb%8e%e9%9b%b6%e5%bc%80%e5%a7%8b%e8%ae%ad%e7%bb%83%e4%b8%80%e4%b8%aa%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%ef%bc%8c%e5%ae%8c%e6%95%b4%e6%b5%81%e7%a8%8b%e6%98%af%e6%80%8e","status":"publish","type":"post","link":"\/?p=9456","title":{"rendered":"\u5982\u679c\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\u4e00\u4e2a\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u5b8c\u6574\u6d41\u7a0b\u662f\u600e\u6837\u7684\uff1f"},"content":{"rendered":"<p>\u4eca\u5929\u804a\u4e00\u7c7b\u9762\u8bd5\u5b98\u975e\u5e38\u559c\u6b22\u3001\u4f46\u7edd\u5927\u591a\u6570\u5019\u9009\u4eba\u7b54\u4e0d\u597d\u7684\u95ee\u9898\uff1a<\/p>\n<blockquote>\n<p>\u9762\u8bd5\u5b98\u95ee\uff1a\u201c\u8bf7\u7cfb\u7edf\u6027\u5730\u9610\u8ff0\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\u4e00\u4e2a\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5b8c\u6574\u6d41\u7a0b\u3002\u201d<\/p>\n<\/blockquote>\n<p>\u8fd9\u7c7b\u95ee\u9898\u7684\u96be\u5ea6\u5728\u4e8e\uff1a\u4f60\u4e0d\u4ec5\u8981\u77e5\u9053\u6a21\u578b\u600e\u4e48\u8bad\uff0c\u66f4\u8981\u80fd\u4ece<strong>\u201c\u5de5\u7a0b\u89c6\u89d2+\u7406\u8bba\u89c6\u89d2\u201d<\/strong>\u4e24\u65b9\u9762\u8bb2\u6e05\u695a\u3002<\/p>\n<p>\u8fd9\u90e8\u5206\u8003\u5bdf\u4f60\u5bf9\u5927\u6a21\u578b\u751f\u547d\u5468\u671f\u7684\u5b8f\u89c2\u7406\u89e3\uff0c\u4ece\u6570\u636e\u51c6\u5907\u5230\u6a21\u578b\u6210\u578b\uff0c\u6bcf\u4e00\u4e2a\u73af\u8282\u90fd\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p>\u4e00\u822c\u4eba\u6ca1\u6709\u8fd9\u4e2a\u7ecf\u5386\uff0c\u5bf9\u7740\u8bba\u6587\u628a\u6d41\u7a0b\u8bb0\u4f4f\u5e94\u4ed8\u9762\u8bd5\u5c31\u597d\u4e86\uff0c\u6216\u8005\u8bb2\u8bb2\u81ea\u5df1\u7b80\u5386\u4e0a\u7684\u5176\u4ed6\u6a21\u578b\u8bad\u7ec3\u7ecf\u9a8c\u3002<\/p>\n<p>\u4f46\uff0c\u5982\u679c\u80fd\u7ed9\u51fa\u4e00\u4e2a\u6709\u6df1\u5ea6\u3001\u6709\u5b9e\u8df5\u5473\u7684\u56de\u7b54\uff0c\u90a3\u80af\u5b9a\u662f\u52a0\u5206\u9879\uff01<\/p>\n<p>\u4eca\u5929\u6211\u4eec\u7cfb\u7edf\u804a\u804a\uff0c\u4ece<strong>\u201c\u600e\u4e48\u8bad\u201d<\/strong>\u5230<strong>\u201c\u8bad\u4ec0\u4e48\u201d<\/strong>\uff0c\u4e00\u6587\u5403\u900f\u5927\u6a21\u578b\u7684\u751f\u547d\u7ebf\u3002<\/p>\n<h2>\u4e00\u3001\u4ece\u96f6\u5230\u4e00\u7684\u5927\u6a21\u578b\u8bad\u7ec3\u6d41\u7a0b<\/h2>\n<p>\u5982\u679c\u4f60\u771f\u505a\u8fc7\u5927\u6a21\u578b\u9879\u76ee\uff0c\u4f60\u4f1a\u77e5\u9053\uff1a\u8bad\u7ec3\u4e0d\u662f\u7b80\u5355\u5730\u201c\u8dd1\u4e2aloss\u4e0b\u964d\u201d\uff0c\u800c\u662f\u4e00\u573a\u6d89\u53ca\u6570\u636e\u3001\u7b97\u529b\u3001\u4f18\u5316\u548c\u5bf9\u9f50\u7684\u7cfb\u7edf\u5de5\u7a0b\u3002<\/p>\n<p>\u6574\u4e2a\u6d41\u7a0b\u53ef\u4ee5\u62c6\u89e3\u4e3a\u4e09\u5927\u9636\u6bb5\uff1a<\/p>\n<blockquote>\n<p>\u6570\u636e\u51c6\u5907 \u2192 \u9884\u8bad\u7ec3 \u2192 \u540e\u8bad\u7ec3\uff08\u6216\u79f0\u5bf9\u9f50\uff09<\/p>\n<\/blockquote>\n<h3><strong>\u6570\u636e\u51c6\u5907\uff1a\u6a21\u578b\u80fd\u529b\u7684\u5929\u82b1\u677f\uff0c\u65e9\u5728\u6570\u636e\u9636\u6bb5\u5c31\u51b3\u5b9a\u4e86<\/strong><\/h3>\n<p>\u8fd9\u4e00\u90e8\u5206\u5728\u9762\u8bd5\u91cc<strong>\u6700\u5bb9\u6613\u88ab\u5ffd\u7565<\/strong>\uff0c\u4f46\u5176\u5b9e\u5b83\u624d\u662f\u51b3\u5b9a\u6a21\u578b\u4e0a\u9650\u7684\u5173\u952e\u3002<\/p>\n<p><strong>\uff081\uff09\u6570\u636e\u6536\u96c6<\/strong>\u6570\u636e\u6765\u6e90\u901a\u5e38\u5305\u62ec\u4e09\u7c7b\uff1a<\/p>\n<ul>\n<li>\u516c\u5171\u8bed\u6599\uff08\u5982Wikipedia\u3001C4\u3001OpenWebText\uff09\uff1b<\/li>\n<li>\u5782\u76f4\u9886\u57df\u6570\u636e\uff08\u5982\u6cd5\u5f8b\u3001\u533b\u7597\u3001\u4ee3\u7801\uff09\uff1b<\/li>\n<li>\u4eba\u5de5\u6784\u9020\u6216\u5408\u6210\u6570\u636e\uff08\u6307\u4ee4\u6570\u636e\u3001\u5bf9\u8bdd\u6570\u636e\u7b49\uff09\u3002<\/li>\n<\/ul>\n<p>\u5de5\u7a0b\u5b9e\u8df5\u4e2d\u5f80\u5f80\u4f1a\u6df7\u5408\u591a\u6e90\u6570\u636e\uff0c\u6bd4\u4f8b\u7684\u62ff\u634f\u975e\u5e38\u8bb2\u7a76\u3002<\/p>\n<p><strong>\uff082\uff09\u6570\u636e\u6e05\u6d17<\/strong>\u6e05\u6d17\u7684\u76ee\u6807\u662f<strong>\u4fdd\u8bc1\u5e72\u51c0\u3001\u5747\u8861\u3001\u5408\u6cd5\u3001\u53ef\u5b66\u4e60<\/strong>\u3002 \u4e3b\u8981\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u53bb\u91cd\uff08\u9632\u6b62\u6a21\u578b\u8fc7\u62df\u5408\u4e8e\u91cd\u590d\u6837\u672c\uff09\uff1b<\/li>\n<li>\u53bb\u566a\uff08\u53bb\u9664\u4e71\u7801\u3001\u5e7f\u544a\u3001\u65e0\u610f\u4e49\u6587\u672c\uff09\uff1b<\/li>\n<li>\u6709\u5bb3\u5185\u5bb9\u8fc7\u6ee4\uff08\u653f\u6cbb\u3001\u9690\u79c1\u3001\u66b4\u529b\u3001\u8272\u60c5\u7b49\uff09\uff1b<\/li>\n<li>\u8bed\u8a00\u68c0\u6d4b\u3001\u957f\u5ea6\u8fc7\u6ee4\u3002<\/li>\n<\/ul>\n<p>\u9ad8\u8d28\u91cf\u6570\u636e\u96c6\u7684\u6784\u5efa\u5f80\u5f80\u6bd4\u8c03\u6a21\u578b\u66f4\u82b1\u65f6\u95f4\u3002 \u5f88\u591a\u516c\u53f8\u4f1a\u7ef4\u62a4\u4e00\u6574\u5957\u6570\u636e pipeline\uff0c\u6301\u7eed\u8fed\u4ee3\u8bed\u6599\u3002<\/p>\n<p><strong>\uff083\uff09\u6570\u636e\u914d\u6bd4<\/strong><\/p>\n<p>\u4e0d\u662f\u6240\u6709\u8bed\u6599\u90fd\u63091:1\u6df7\u5408\u3002<\/p>\n<p>\u6bd4\u5982\u60f3\u8bad\u7ec3\u901a\u7528\u8bed\u8a00\u6a21\u578b\uff0c\u53ef\u4ee5\u8ba9\u201c\u5f00\u653e\u57df\u5bf9\u8bdd\u201d\u536040%\uff0c\u4ee3\u7801\u6570\u636e\u536010%\uff0c\u77e5\u8bc6\u7c7b\u6587\u672c\u536020%\uff0c\u5269\u4e0b\u662f\u7f51\u9875\u6216\u4e66\u7c4d\u5185\u5bb9\u3002<\/p>\n<p>\u914d\u6bd4\u4e0d\u540c\uff0c\u6a21\u578b\u7684\u6027\u683c\u4e5f\u4e0d\u540c\u3002<\/p>\n<p><strong>\uff084\uff09\u5206\u8bcd\uff08Tokenization\uff09<\/strong><\/p>\n<p>\u522b\u5c0f\u770b\u5206\u8bcd\u5668\uff0c\u5b83\u51b3\u5b9a\u4e86\u6a21\u578b\u7684\u8f93\u5165\u7c92\u5ea6\uff0c\u4e3b\u6d41\u65b9\u6cd5\u5305\u62ec BPE\u3001WordPiece\u3001SentencePiece \u7b49\u3002<\/p>\n<p>\u8fd1\u5e74\u6765\u5174\u8d77\u7684 <code>tiktoken<\/code> \u65b9\u6848\uff08OpenAI \u4f7f\u7528\uff09\u5bf9\u591a\u8bed\u8a00\u4e0e\u4ee3\u7801\u90fd\u6709\u5f88\u597d\u7684\u517c\u5bb9\u6027\uff0c\u5206\u8bcd\u8d28\u91cf\u5dee\uff0c\u6a21\u578b\u5b66\u5230\u7684\u8bed\u4e49\u4f1a\u788e\u6389\u3002<\/p>\n<h3><strong>\u9884\u8bad\u7ec3\uff08Pre-training\uff09\uff1a\u8ba9\u6a21\u578b\u5b66\u4f1a\u201c\u8bed\u8a00\u4e16\u754c\u7684\u89c4\u5f8b\u201d<\/strong><\/h3>\n<p>\u9884\u8bad\u7ec3\u662f\u6574\u4e2a\u751f\u547d\u5468\u671f\u4e2d<strong>\u6700\u6602\u8d35\u4f46\u6700\u6838\u5fc3\u7684\u4e00\u6b65<\/strong>\uff0c\u76ee\u6807\u662f\u8ba9\u6a21\u578b\u7406\u89e3\u8bed\u8a00\u7684\u7edf\u8ba1\u7279\u5f81\u3001\u903b\u8f91\u5173\u7cfb\u548c\u4e16\u754c\u5e38\u8bc6\u3002<\/p>\n<p><strong>\uff081\uff09\u9884\u8bad\u7ec3\u7684\u76ee\u6807\u51fd\u6570<\/strong><\/p>\n<p>\u4e3b\u8981\u6709\u4e24\u7c7b\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u81ea\u56de\u5f52\u8bed\u8a00\u5efa\u6a21\uff08Causal LM\uff09<\/strong>\uff1a\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u3002<\/p>\n<p>\u4ee3\u8868\u6a21\u578b\uff1aGPT \u7cfb\u5217\u3002<\/p>\n<p>\u635f\u5931\u51fd\u6570\uff1a\u4ea4\u53c9\u71b5\u635f\u5931<\/p>\n<\/li>\n<li>\n<p><strong>\u63a9\u7801\u8bed\u8a00\u5efa\u6a21\uff08Masked LM\uff09<\/strong>\uff1a\u9884\u6d4b\u88ab\u63a9\u76d6\u7684\u8bcd\u3002<\/p>\n<p>\u4ee3\u8868\u6a21\u578b\uff1aBERT\u3002<\/p>\n<p>\u4f18\u70b9\u662f\u53cc\u5411\u4e0a\u4e0b\u6587\u5efa\u6a21\uff0c\u4f46\u4e0d\u9002\u5408\u751f\u6210\u4efb\u52a1\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u9762\u8bd5\u65f6\u5982\u679c\u80fd\u660e\u786e\u533a\u5206\u8fd9\u4e24\u7c7b\u5efa\u6a21\u76ee\u6807\uff0c\u5e76\u80fd\u89e3\u91ca\u5176\u5dee\u5f02\u4e0e\u9002\u7528\u573a\u666f\uff0c\u4f1a\u975e\u5e38\u52a0\u5206\u3002<\/p>\n<p><strong>\uff082\uff09\u8bad\u7ec3\u914d\u7f6e\u4e0e\u5de5\u7a0b\u6311\u6218<\/strong><\/p>\n<ul>\n<li>\u6a21\u578b\u89c4\u6a21\uff1a\u6570\u5341\u4ebf\u5230\u4e0a\u5343\u4ebf\u53c2\u6570\uff1b<\/li>\n<li>\u8bad\u7ec3\u6846\u67b6\uff1aMegatron-LM\u3001DeepSpeed\u3001ColossalAI\u3001vLLM\uff1b<\/li>\n<li>\u5206\u5e03\u5f0f\u8bad\u7ec3\uff1a\u6570\u636e\u5e76\u884c\u3001\u6a21\u578b\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\uff1b<\/li>\n<li>\u4f18\u5316\u5668\uff1aAdamW\u3001LAMB\uff1b<\/li>\n<li>\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\uff1aFP16\/BF16\uff1b<\/li>\n<li>Checkpoint &amp; Resume\uff1a\u4e2d\u65ad\u6062\u590d\u673a\u5236\u3002<\/li>\n<\/ul>\n<p>\u4e00\u53e5\u8bdd\u603b\u7ed3\uff1a \u201c\u9884\u8bad\u7ec3\u662f\u70e7\u94b1\u7684\u827a\u672f\uff0c\u66f4\u662f\u7b97\u529b\u3001\u5de5\u7a0b\u4e0e\u6570\u5b66\u7684\u535a\u5f08\u3002\u201d<\/p>\n<h3><strong>3\ufe0f\u20e3 \u540e\u8bad\u7ec3 \/ \u5bf9\u9f50\uff08Post-training \/ Alignment\uff09<\/strong><\/h3>\n<p>\u9884\u8bad\u7ec3\u8ba9\u6a21\u578b\u201c\u6709\u77e5\u8bc6\u201d\uff0c\u4f46\u5b83\u8fd8\u201c\u6ca1\u6559\u517b\u201d\u3002<\/p>\n<p>\u8981\u8ba9\u6a21\u578b\u80fd\u542c\u61c2\u4eba\u8bdd\u3001\u9075\u5b88\u6307\u4ee4\u3001\u4e0d\u4e71\u8bf4\uff0c\u5c31\u8981\u901a\u8fc7<strong>\u5bf9\u9f50\uff08Alignment\uff09<\/strong>\u9636\u6bb5\u6765\u201c\u77eb\u6b63\u6027\u683c\u201d\u3002<\/p>\n<h4><strong>\uff081\uff09\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09<\/strong><\/h4>\n<p>SFT \u662f\u201c\u6559\u6a21\u578b\u9075\u5b88\u4eba\u7c7b\u6307\u4ee4\u201d\u7684\u7b2c\u4e00\u6b65\uff0c\u7528\u9ad8\u8d28\u91cf\u7684\u300c\u6307\u4ee4 - \u56de\u7b54\u300d\u6570\u636e\u5bf9\uff08Instruction-Response Pair\uff09\u8bad\u7ec3\u6a21\u578b\u3002<\/p>\n<p>\u4f46\u6b64\u65f6\u6570\u636e\u4e0d\u518d\u662f\u968f\u673a\u7f51\u9875\u6587\u672c\uff0c\u800c\u662f\u4eba\u5de5\u6216\u534a\u81ea\u52a8\u751f\u6210\u7684\u201c\u4f18\u8d28\u95ee\u7b54\u201d\uff0cSFT \u7684\u597d\u574f\uff0c\u76f4\u63a5\u51b3\u5b9a\u6a21\u578b\u662f\u5426\u201c\u542c\u8bdd\u201d\u3002<\/p>\n<h4><strong>\uff082\uff09\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\uff08RLHF \/ DPO\uff09<\/strong><\/h4>\n<p>\u63a5\u4e0b\u6765\uff0c\u8ba9\u6a21\u578b\u201c\u4e0d\u4ec5\u542c\u8bdd\uff0c\u8fd8\u61c2\u5206\u5bf8\u201d\uff0c\u8fd9\u4e00\u6b65\u7684\u76ee\u6807\u662f\u8ba9\u6a21\u578b\u8f93\u51fa\u66f4\u7b26\u5408\u4eba\u7c7b\u504f\u597d\u3002<\/p>\n<ul>\n<li><strong>RLHF\uff08Reinforcement Learning from Human Feedback\uff09<\/strong>\u5305\u62ec\u4e09\u6b65\uff1a\n<ol>\n<li>\u751f\u6210\u591a\u6837\u56de\u7b54\uff1b<\/li>\n<li>\u8ba9\u4eba\u5de5\u6807\u6ce8\u54ea\u4e2a\u66f4\u597d\uff1b<\/li>\n<li>\u8bad\u7ec3\u5956\u52b1\u6a21\u578b\uff08Reward Model\uff09+ PPO \u4f18\u5316\u3002<\/li>\n<\/ol>\n<\/li>\n<li><strong>DPO\uff08Direct Preference Optimization\uff09<\/strong>\u662fRLHF\u7684\u7b80\u5316\u7248\u672c\uff0c\u76f4\u63a5\u901a\u8fc7\u504f\u597d\u5bf9\u4f18\u5316\u76ee\u6807\u8fdb\u884c\u5efa\u6a21\u3002 \u4e0d\u518d\u9700\u8981\u5956\u52b1\u6a21\u578b\uff0c\u8bad\u7ec3\u66f4\u7a33\u5b9a\u3002<\/li>\n<\/ul>\n<p>\u4e00\u53e5\u8bdd\u603b\u7ed3\uff1a\u201cSFT\u8ba9\u6a21\u578b\u542c\u6307\u4ee4\uff0cRLHF\u8ba9\u6a21\u578b\u8bb2\u4eba\u8bdd\u3002\u201d<\/p>\n<h2>\u4e8c\u3001\u8bad\u7ec3\u4e2d\u7684\u5173\u952e\u6982\u5ff5\u4e0e\u7406\u8bba\u6311\u6218<\/h2>\n<p>\u5982\u679c\u4f60\u80fd\u5728\u9762\u8bd5\u4e2d\u8bb2\u5230\u8fd9\u4e00\u5c42\uff0c\u57fa\u672c\u5c31\u662f\u201c\u9ad8\u9636\u9009\u624b\u201d\u4e86\u3002<\/p>\n<h3><strong>1\ufe0f\u20e3 Scaling Laws\uff08\u7f29\u653e\u5b9a\u5f8b\uff09<\/strong><\/h3>\n<p>\u7f29\u653e\u5b9a\u5f8b\u63cf\u8ff0\u4e86<strong>\u6a21\u578b\u6027\u80fd\u4e0e\u6a21\u578b\u89c4\u6a21\u3001\u6570\u636e\u91cf\u3001\u8ba1\u7b97\u91cf\u4e09\u8005\u7684\u5e42\u5f8b\u5173\u7cfb<\/strong>\u3002<\/p>\n<p>\u7b80\u5355\u6765\u8bf4\uff1a\u201c\u53ea\u8981\u94b1\u591f\u591a\uff0c\u6a21\u578b\u4e00\u5b9a\u4f1a\u66f4\u597d\uff0c\u4f46\u8981\u82b1\u5f97\u5728\u5200\u5203\u4e0a\u3002\u201d<\/p>\n<p>\u7ecf\u9a8c\u4e0a\uff1a<\/p>\n<ul>\n<li>\u6a21\u578b\u6027\u80fd \u2248 k \u00d7 (\u53c2\u6570\u91cf)^\u03b1 \u00d7 (\u6570\u636e\u91cf)^\u03b2 \u00d7 (\u8ba1\u7b97\u91cf)^\u03b3<\/li>\n<li>\u6570\u636e\u89c4\u6a21\u4e0d\u8db3\u65f6\uff0c\u589e\u5927\u6a21\u578b\u53cd\u800c\u4f1a\u8fc7\u62df\u5408\uff1b<\/li>\n<li>\u53cd\u4e4b\uff0c\u7b97\u529b\u592a\u5c0f\u3001Batch\u592a\u5c0f\uff0c\u4e5f\u4f1a\u5f71\u54cd\u6536\u655b\u3002<\/li>\n<\/ul>\n<p>\u6240\u4ee5\uff0c\u5927\u5382\u90fd\u4f1a\u6709\u5185\u90e8\u7684\u201cScaling Law Dashboard\u201d\uff0c\u5e2e\u52a9\u51b3\u7b56\u6a21\u578b\u89c4\u6a21\u4e0e\u9884\u7b97\u3002<\/p>\n<h3><strong>\u8fc7\u62df\u5408\u4e0e\u6b63\u5219\u5316\uff08Overfitting &amp; Regularization\uff09<\/strong><\/h3>\n<p><strong>\u8fc7\u62df\u5408\u7684\u75c7\u72b6\uff1a<\/strong><\/p>\n<ul>\n<li>\u8bad\u7ec3\u96c6 loss \u5f88\u4f4e\uff1b<\/li>\n<li>\u9a8c\u8bc1\u96c6\u6027\u80fd\u4e0b\u6ed1\uff1b<\/li>\n<li>\u6a21\u578b\u8f93\u51fa\u201c\u80cc\u4e66\u5f0f\u56de\u7b54\u201d\u3002<\/li>\n<\/ul>\n<p><strong>\u5e38\u89c1\u89e3\u51b3\u65b9\u6848\uff1a<\/strong><\/p>\n<ul>\n<li>\u6570\u636e\u589e\u5f3a\uff08Data Augmentation\uff09\uff1b<\/li>\n<li>Dropout\uff1b<\/li>\n<li>\u6743\u91cd\u8870\u51cf\uff08L1\/L2 Regularization\uff09\uff1b<\/li>\n<li>Early Stopping\uff1b<\/li>\n<li>Mixout \/ LayerNorm \u8c03\u6574\u3002<\/li>\n<\/ul>\n<p>\u5728\u9762\u8bd5\u4e2d\u4f60\u53ef\u4ee5\u5f3a\u8c03\u4e00\u70b9\uff1a<\/p>\n<blockquote>\n<p>\u201c\u6211\u4eec\u4e00\u822c\u901a\u8fc7\u52a8\u6001\u76d1\u63a7\u9a8c\u8bc1\u96c6\u635f\u5931\u66f2\u7ebf\uff0c\u81ea\u52a8\u65e9\u505c\u6765\u9632\u6b62\u8fc7\u62df\u5408\u3002\u201d<\/p>\n<\/blockquote>\n<p>\u8fd9\u8bf4\u660e\u4f60\u6709\u5b9e\u6218\u7ecf\u9a8c\u3002<\/p>\n<h3><strong>\u68af\u5ea6\u95ee\u9898\uff08Vanishing \/ Exploding Gradients\uff09<\/strong><\/h3>\n<p>\u8bad\u7ec3\u5927\u6a21\u578b\u6700\u5e38\u89c1\u7684\u201c\u9690\u5f62\u6740\u624b\u201d\uff0c\u5c31\u662f\u68af\u5ea6\u4e0d\u7a33\u5b9a\u3002<\/p>\n<p><strong>\u6210\u56e0\uff1a<\/strong><\/p>\n<ul>\n<li>\u94fe\u5f0f\u6cd5\u5219\u5bfc\u81f4\u68af\u5ea6\u9010\u5c42\u8870\u51cf\u6216\u653e\u5927\uff1b<\/li>\n<li>\u6fc0\u6d3b\u51fd\u6570\uff08\u5982sigmoid\u3001tanh\uff09\u9971\u548c\uff1b<\/li>\n<li>\u7f51\u7edc\u5c42\u6570\u592a\u6df1\u3001\u53c2\u6570\u521d\u59cb\u5316\u4e0d\u5408\u7406\u3002<\/li>\n<\/ul>\n<p><strong>\u89e3\u51b3\u624b\u6bb5\uff1a<\/strong><\/p>\n<ul>\n<li>\u6b8b\u5dee\u8fde\u63a5\uff08ResNet-style\uff09\uff1b<\/li>\n<li>\u68af\u5ea6\u88c1\u526a\uff08Gradient Clipping\uff09\uff1b<\/li>\n<li>\u5f52\u4e00\u5316\u5c42\uff08LayerNorm, RMSNorm\uff09\uff1b<\/li>\n<li>\u6743\u91cd\u521d\u59cb\u5316\uff08Xavier, Kaiming\uff09\uff1b<\/li>\n<li>\u6539\u7528ReLU\/GELU\u7b49\u975e\u9971\u548c\u6fc0\u6d3b\u51fd\u6570\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e00\u5757\u5982\u679c\u80fd\u8bf4\u51fa\u4f60\u8c03\u8fc7\u68af\u5ea6\u7206\u70b8\u7684\u771f\u5b9e\u6848\u4f8b\uff0c\u9762\u8bd5\u5b98\u4f1a\u7acb\u523b\u89c9\u5f97\u4f60\u201c\u771f\u5e72\u8fc7\u201d\u3002<\/p>\n<h2>\u4e09\u3001\u603b\u7ed3\uff1a\u5982\u4f55\u5728\u9762\u8bd5\u4e2d\u7cfb\u7edf\u56de\u7b54\u8fd9\u9053\u9898\uff1f<\/h2>\n<p>\u4e00\u5206\u949f\u9ad8\u8d28\u91cf\u56de\u7b54\u6a21\u677f\uff1a<\/p>\n<blockquote>\n<p>\u201c\u4ece\u96f6\u8bad\u7ec3\u4e00\u4e2a\u5927\u6a21\u578b\uff0c\u53ef\u4ee5\u5206\u4e3a\u4e09\u6b65\uff1a<\/p>\n<p>\u7b2c\u4e00\u9636\u6bb5\u662f\u6570\u636e\u51c6\u5907\uff0c\u5305\u542b\u6570\u636e\u6536\u96c6\u3001\u6e05\u6d17\u3001\u5206\u8bcd\u548c\u914d\u6bd4\uff0c\u662f\u6a21\u578b\u80fd\u529b\u7684\u4e0a\u9650\uff1b<\/p>\n<p>\u7b2c\u4e8c\u9636\u6bb5\u662f\u9884\u8bad\u7ec3\uff0c\u76ee\u6807\u662f\u5b66\u4e60\u8bed\u8a00\u4e0e\u4e16\u754c\u77e5\u8bc6\uff0c\u5173\u952e\u5728\u4efb\u52a1\u8bbe\u8ba1\u4e0e\u5206\u5e03\u5f0f\u8bad\u7ec3\uff1b<\/p>\n<p>\u7b2c\u4e09\u9636\u6bb5\u662f\u540e\u8bad\u7ec3\uff0c\u4e5f\u5c31\u662f\u5bf9\u9f50\uff0c\u5305\u62ecSFT\u548cRLHF\uff0c\u8ba9\u6a21\u578b\u4ece\u2018\u4f1a\u8bf4\u8bdd\u2019\u5230\u2018\u61c2\u4eba\u8bdd\u2019\uff1b<\/p>\n<p>\u8fc7\u7a0b\u4e2d\u8fd8\u8981\u8003\u8651\u7f29\u653e\u5b9a\u5f8b\u3001\u8fc7\u62df\u5408\u4e0e\u68af\u5ea6\u7a33\u5b9a\u6027\u95ee\u9898\uff0c\u8fd9\u4e9b\u90fd\u4f1a\u5f71\u54cd\u6a21\u578b\u6700\u7ec8\u6027\u80fd\u3002\u201d<\/p>\n<\/blockquote>\n<p>\u8fd9\u6837\u7684\u56de\u7b54\uff0c\u65e2\u7cfb\u7edf\u53c8\u6709\u5b9e\u64cd\u5473\uff0c\u9762\u8bd5\u5b98\u542c\u5b8c\u57fa\u672c\u4f1a\u70b9\u5934\u8ba4\u53ef\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4eca\u5929\u804a\u4e00\u7c7b\u9762\u8bd5\u5b98\u975e\u5e38\u559c\u6b22\u3001\u4f46\u7edd\u5927\u591a\u6570\u5019\u9009\u4eba\u7b54\u4e0d\u597d\u7684\u95ee\u9898\uff1a \u9762\u8bd5\u5b98\u95ee\uff1a\u201c\u8bf7\u7cfb\u7edf\u6027\u5730\u9610\u8ff0\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\u4e00\u4e2a\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5b8c\u6574\u6d41\u7a0b\u3002\u201d \u8fd9\u7c7b\u95ee\u9898\u7684   \u2026 &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ngg_post_thumbnail":0},"categories":[225],"tags":[],"_links":{"self":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9456"}],"collection":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=9456"}],"version-history":[{"count":1,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9456\/revisions"}],"predecessor-version":[{"id":9457,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9456\/revisions\/9457"}],"wp:attachment":[{"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=9456"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=9456"},{"taxonomy":"post_tag","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=9456"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}