{"id":7532,"date":"2026-07-30T01:56:50","date_gmt":"2026-07-29T17:56:50","guid":{"rendered":"https:\/\/i007.cc\/wordpress\/?p=7532"},"modified":"2026-07-30T01:56:50","modified_gmt":"2026-07-29T17:56:50","slug":"ai%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%89%8d%e6%b2%bf%e7%9f%a5%e8%af%86","status":"publish","type":"post","link":"https:\/\/i007.cc\/wordpress\/archives\/7532","title":{"rendered":"AI\u5927\u6a21\u578b\u524d\u6cbf\u77e5\u8bc6"},"content":{"rendered":"<div data-rs-index=\"10\" data-index=\"10\">\n<div tabindex=\"0\" role=\"article\" aria-setsize=\"14\" aria-posinset=\"11\" aria-label=\"Message 11 of 14\">\n<div data-test-render-count=\"2\">\n<div class=\"mb-1 mt-[var(--msg-gap,1.5rem)] group group\/message-row\">\n<div class=\"flex flex-col items-end gap-1\">\n<div class=\"group relative inline-flex gap-2 bg-bg-300 dark:[[data-darker-default]_&amp;]:bg-bg-000 rounded-xl pl-2.5 py-[var(--msg-bubble-py,0.625rem)] break-words text-text-100 transition-all max-w-[75ch] flex-col !px-[var(--msg-bubble-px,1rem)] max-w-[85%]\" data-user-message-bubble=\"true\">\n<div class=\"flex flex-row gap-2 relative\">\n<div class=\"flex-1\">\n<div class=\"font-large !font-user-message py-0.5 grid grid-cols-1 gap-2 relative [&amp;_ul]:!space-y-0 [&amp;_ol]:!space-y-0 [&amp;_ul]:pl-8 [&amp;_ol]:pl-8\" data-testid=\"user-message\">\n<h2 class=\"whitespace-pre-wrap break-words\">\u8bb2\u8bb2SFT\u8fd9\u6837\u7684\u7814\u53d1\u8303\u5f0f<\/h2>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<div data-rs-index=\"11\" data-index=\"11\">\n<div tabindex=\"-1\" role=\"article\" aria-setsize=\"14\" aria-posinset=\"12\" aria-label=\"Message 12 of 14\">\n<div data-test-render-count=\"2\">\n<div class=\"group group\/message-row\">\n<div class=\"group relative relative pb-[var(--msg-assistant-pb,0.75rem)]\" data-is-streaming=\"false\">\n<div class=\"font-claude-response relative leading-[1.65rem] [&amp;_pre&gt;div]:bg-bg-000\/50 [&amp;_pre&gt;div]:border-0.5 [&amp;_pre&gt;div]:border-border-400 [&amp;_.ignore-pre-bg&gt;div]:bg-transparent [&amp;_.standard-markdown_:is(p,blockquote,h1,h2,h3,h4,h5,h6)]:pl-2 [&amp;_.standard-markdown_:is(p,blockquote,ul,ol,h1,h2,h3,h4,h5,h6)]:pr-8 [&amp;_.progressive-markdown_:is(p,blockquote,h1,h2,h3,h4,h5,h6)]:pl-2 [&amp;_.progressive-markdown_:is(p,blockquote,ul,ol,h1,h2,h3,h4,h5,h6)]:pr-8\">\n<div>\n<div class=\"standard-markdown grid-cols-1 grid [&amp;_&gt;_*]:min-w-0 gap-3 [&amp;_&gt;_*:last-child]:mb-0 print:block print:[&amp;_&gt;_*_+_*]:mt-3 standard-markdown\">\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">SFT\u5168\u79f0\u662fSupervised Fine-Tuning,\u4e5f\u5c31\u662f&#8221;\u76d1\u7763\u5fae\u8c03&#8221;,\u6307\u7684\u662f\u5728\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u57fa\u7840\u4e0a,\u7528\u4eba\u5de5\u51c6\u5907\u597d\u7684\u9ad8\u8d28\u91cf\u8f93\u5165\u8f93\u51fa\u6837\u672c,\u901a\u8fc7\u76d1\u7763\u5b66\u4e60\u7684\u65b9\u5f0f\u8ba9\u6a21\u578b\u5b66\u4f1a\u7279\u5b9a\u7684\u884c\u4e3a\u6a21\u5f0f\u3002\u653e\u5728\u5927\u6a21\u578b\u7814\u53d1\u7684\u6574\u4f53\u8303\u5f0f\u91cc\u770b,SFT\u662f\u8fde\u63a5&#8221;\u9884\u8bad\u7ec3&#8221;\u548c&#8221;\u5bf9\u9f50&#8221;\u8fd9\u4e24\u4e2a\u9636\u6bb5\u7684\u5173\u952e\u4e00\u73af,\u7406\u89e3\u5b83\u7684\u4f5c\u7528\u548c\u5c40\u9650,\u6709\u52a9\u4e8e\u7406\u89e3\u6574\u4e2a\u8bad\u7ec3\u6d41\u6c34\u7ebf\u4e3a\u4ec0\u4e48\u8981\u8fd9\u6837\u8bbe\u8ba1\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u5148\u8bf4\u5b83\u5728\u6d41\u6c34\u7ebf\u91cc\u7684\u4f4d\u7f6e\u3002\u9884\u8bad\u7ec3\u9636\u6bb5\u6a21\u578b\u5728\u6d77\u91cf\u4e92\u8054\u7f51\u6587\u672c\u4e0a\u5b66\u4e60&#8221;\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd&#8221;,\u8fd9\u4e2a\u8fc7\u7a0b\u8d4b\u4e88\u4e86\u6a21\u578b\u5e7f\u535a\u7684\u77e5\u8bc6\u548c\u8bed\u8a00\u80fd\u529b,\u4f46\u9884\u8bad\u7ec3\u7684\u76ee\u6807\u51fd\u6570\u53ea\u662f\u8ba9\u6a21\u578b\u7684\u8f93\u51fa\u5728\u7edf\u8ba1\u4e0a\u50cf\u8bad\u7ec3\u8bed\u6599,\u5e76\u4e0d\u5929\u7136\u5bf9\u5e94&#8221;\u50cf\u4e00\u4e2a\u4e50\u4e8e\u52a9\u4eba\u7684\u52a9\u624b\u90a3\u6837\u56de\u7b54\u95ee\u9898&#8221;\u3002\u8fd9\u65f6\u5019\u5982\u679c\u76f4\u63a5\u8ddf\u9884\u8bad\u7ec3\u6a21\u578b\u5bf9\u8bdd,\u4f60\u4f1a\u53d1\u73b0\u5b83\u66f4\u50cf\u662f\u5728\u505a\u6587\u672c\u7eed\u5199\u800c\u4e0d\u662f\u597d\u597d\u56de\u7b54\u2014\u2014\u6bd4\u5982\u4f60\u95ee\u4e00\u4e2a\u95ee\u9898,\u5b83\u53ef\u80fd\u4f1a\u7eed\u5199\u51fa\u66f4\u591a\u7c7b\u4f3c\u7684\u95ee\u9898,\u800c\u4e0d\u662f\u7ed9\u51fa\u7b54\u6848\u3002SFT\u8981\u89e3\u51b3\u7684\u6b63\u662f\u8fd9\u4e2a&#8221;\u77e5\u8bc6\u5f88\u591a\u4f46\u4e0d\u77e5\u9053\u600e\u4e48\u6070\u5f53\u5730\u8868\u8fbe\u548c\u5e94\u7528&#8221;\u7684\u95ee\u9898\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u5177\u4f53\u505a\u6cd5\u662f\u51c6\u5907\u4e00\u6279&#8221;\u63d0\u793a-\u7406\u60f3\u56de\u7b54&#8221;\u7684\u6837\u672c\u5bf9,\u8fd9\u4e9b\u6837\u672c\u901a\u5e38\u7531\u4eba\u7c7b\u6807\u6ce8\u8005\u64b0\u5199,\u6216\u8005\u7531\u4eba\u7c7b\u5bf9\u6a21\u578b\u8f93\u51fa\u8fdb\u884c\u4fee\u6539\u6da6\u8272\u800c\u6210,\u6db5\u76d6\u5404\u79cd\u573a\u666f:\u5982\u4f55\u793c\u8c8c\u5730\u56de\u7b54\u95ee\u9898\u3001\u5982\u4f55\u62d2\u7edd\u4e0d\u5408\u9002\u7684\u8bf7\u6c42\u3001\u5982\u4f55\u7ed9\u51fa\u7ed3\u6784\u6e05\u6670\u7684\u89e3\u91ca\u3001\u5982\u4f55\u4f7f\u7528\u5de5\u5177\u7b49\u7b49\u3002\u7136\u540e\u7528\u6807\u51c6\u7684\u76d1\u7763\u5b66\u4e60\u65b9\u5f0f\u2014\u2014\u4e5f\u5c31\u662f\u8ba9\u6a21\u578b\u53bb\u6700\u5927\u5316\u751f\u6210\u8fd9\u4e9b&#8221;\u6807\u51c6\u7b54\u6848&#8221;\u7684\u6982\u7387\u2014\u2014\u53bb\u5fae\u8c03\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u53c2\u6570\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u76f8\u5bf9\u7b80\u5355\u76f4\u63a5,\u8ba1\u7b97\u6210\u672c\u8fdc\u4f4e\u4e8e\u9884\u8bad\u7ec3,\u5374\u80fd\u8ba9\u6a21\u578b\u7684\u884c\u4e3a\u53d1\u751f\u8d28\u7684\u53d8\u5316,\u4ece&#8221;\u7eed\u5199\u673a\u5668&#8221;\u53d8\u6210&#8221;\u80fd\u5bf9\u8bdd\u3001\u542c\u6307\u4ee4\u7684\u52a9\u624b&#8221;\u3002\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u65e9\u671fInstructGPT\u3001\u4ee5\u53ca\u540e\u6765\u51e0\u4e4e\u6240\u6709\u5bf9\u8bdd\u5f0f\u5927\u6a21\u578b,\u90fd\u4f1a\u628aSFT\u4f5c\u4e3a\u5bf9\u9f50\u6d41\u7a0b\u91cc\u5fc5\u4e0d\u53ef\u5c11\u7684\u7b2c\u4e00\u6b65\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u4f5c\u4e3a\u4e00\u79cd\u7814\u53d1\u8303\u5f0f,SFT\u6709\u51e0\u4e2a\u7279\u70b9\u503c\u5f97\u5c55\u5f00\u8bb2\u3002\u7b2c\u4e00\u662f\u5b83\u9ad8\u5ea6\u4f9d\u8d56\u6570\u636e\u8d28\u91cf\u800c\u975e\u6570\u636e\u6570\u91cf\u3002\u56e0\u4e3aSFT\u6837\u672c\u901a\u5e38\u9700\u8981\u4eba\u5de5\u64b0\u5199\u6216\u7cbe\u5fc3\u7b5b\u9009,\u89c4\u6a21\u8fdc\u5c0f\u4e8e\u9884\u8bad\u7ec3\u8bed\u6599(\u53ef\u80fd\u662f\u51e0\u4e07\u5230\u51e0\u5341\u4e07\u6761,\u800c\u4e0d\u662f\u4e07\u4ebf\u7ea7token),\u6240\u4ee5\u6bcf\u4e00\u6761\u6837\u672c\u7684\u8d28\u91cf\u3001\u591a\u6837\u6027\u3001\u8986\u76d6\u9762\u90fd\u76f4\u63a5\u51b3\u5b9a\u4e86\u5fae\u8c03\u6548\u679c\u2014\u2014\u8fd9\u4e5f\u50ac\u751f\u4e86\u5927\u91cf\u5173\u4e8e&#8221;\u5982\u4f55\u6784\u9020\u66f4\u597d\u7684SFT\u6570\u636e\u96c6&#8221;\u7684\u7814\u7a76,\u5305\u62ec\u8ba9\u66f4\u5f3a\u7684\u6a21\u578b\u751f\u6210\u5019\u9009\u56de\u7b54\u518d\u7531\u4eba\u5de5\u6311\u9009\u4fee\u6539(\u8fd9\u6837\u4eba\u7c7b\u53ea\u9700\u8981\u505a\u5224\u65ad\u548c\u7f16\u8f91,\u800c\u4e0d\u5fc5\u4ece\u96f6\u5199),\u4ee5\u53ca\u540e\u6765\u6d41\u884c\u7684\u7528\u5f3a\u6a21\u578b\u84b8\u998f\u6570\u636e\u7ed9\u8f83\u5f31\u6a21\u578b\u8bad\u7ec3\u7684\u505a\u6cd5\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u7b2c\u4e8c\u662fSFT\u672c\u8d28\u4e0a\u662f\u5728&#8221;\u6a21\u4eff&#8221;,\u800c\u4e0d\u662f\u5728&#8221;\u4f18\u5316\u504f\u597d&#8221;\u3002\u5b83\u6559\u4f1a\u6a21\u578b\u7684\u662f\u590d\u73b0\u8bad\u7ec3\u6837\u672c\u91cc\u90a3\u79cd\u7279\u5b9a\u98ce\u683c\u548c\u5185\u5bb9\u7684\u56de\u7b54,\u4f46\u6ca1\u6709\u4e00\u4e2a\u673a\u5236\u8ba9\u6a21\u578b\u7406\u89e3&#8221;\u8fd9\u4e2a\u56de\u7b54\u6bd4\u90a3\u4e2a\u56de\u7b54\u597d\u5728\u54ea\u91cc&#8221;\u8fd9\u79cd\u76f8\u5bf9\u6bd4\u8f83\u5173\u7cfb\u3002\u8fd9\u6b63\u662f\u4e3a\u4ec0\u4e48SFT\u4e4b\u540e\u5f80\u5f80\u8fd8\u9700\u8981\u63a5\u4e0aRLHF\u6216\u8005DPO\u8fd9\u7c7b\u57fa\u4e8e\u504f\u597d\u5bf9\u6bd4\u7684\u65b9\u6cd5\u2014\u2014SFT\u628a\u6a21\u578b\u4ece&#8221;\u4e0d\u4f1a\u597d\u597d\u8bf4\u8bdd&#8221;\u8c03\u6559\u5230&#8221;\u5927\u81f4\u4f1a\u597d\u597d\u8bf4\u8bdd&#8221;,\u540e\u7eed\u7684\u504f\u597d\u4f18\u5316\u518d\u8fdb\u4e00\u6b65\u6253\u78e8\u5230&#8221;\u8bf4\u5f97\u66f4\u8ba9\u4eba\u6ee1\u610f&#8221;\u3002\u53ef\u4ee5\u628aSFT\u7406\u89e3\u6210\u5148\u5b66\u4f1a\u5199\u89c4\u8303\u4f5c\u6587,\u518d\u901a\u8fc7\u53cd\u590d\u5bf9\u6bd4\u4fee\u6539\u5b66\u4f1a\u5199\u51fa\u66f4\u7cbe\u5f69\u7684\u4f5c\u6587\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u7b2c\u4e09,SFT\u8fd9\u5957\u8303\u5f0f\u7684\u4e00\u4e2a\u9690\u542b\u98ce\u9669\u662f\u6a21\u578b\u53ef\u80fd\u4f1a\u5b66\u5230\u8bad\u7ec3\u6570\u636e\u91cc\u4e00\u4e9b\u4e0d\u5e0c\u671b\u88ab\u5b66\u5230\u7684\u8868\u9762\u6a21\u5f0f,\u6bd4\u5982\u6807\u6ce8\u8005\u7684\u4e66\u5199\u4e60\u60ef\u3001\u56de\u7b54\u957f\u5ea6\u504f\u597d\u3001\u751a\u81f3\u65e0\u610f\u8bc6\u7684\u504f\u89c1,\u5982\u679c\u6807\u6ce8\u6570\u636e\u8986\u76d6\u4e0d\u591f\u591a\u6837\u6216\u8005\u5b58\u5728\u7cfb\u7edf\u6027\u504f\u5dee,\u6a21\u578b\u4f1a\u539f\u6837\u7ee7\u627f\u8fd9\u4e9b\u95ee\u9898\u2014\u2014\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u6570\u636e\u6784\u9020\u8fc7\u7a0b\u672c\u8eab\u9700\u8981\u975e\u5e38\u8c28\u614e\u7684\u8d28\u91cf\u63a7\u5236\u548c\u591a\u6837\u6027\u8bbe\u8ba1\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u7b2c\u56db,\u8fd1\u5e74SFT\u8fd9\u4e2a\u8303\u5f0f\u672c\u8eab\u4e5f\u5728\u6f14\u5316,\u5c24\u5176\u662f\u5728\u63a8\u7406\u6a21\u578b\u5174\u8d77\u4e4b\u540e\u3002\u524d\u9762\u804a\u5230\u7684DeepSeek-R1\u91cc,SFT\u88ab\u7528\u4f5c\u5f3a\u5316\u5b66\u4e60\u4e4b\u524d\u7684&#8221;\u51b7\u542f\u52a8&#8221;\u6b65\u9aa4\u2014\u2014\u5148\u7528\u5c11\u91cf\u9ad8\u8d28\u91cf\u7684\u957f\u63a8\u7406\u6837\u672c\u8ba9\u6a21\u578b\u5177\u5907\u57fa\u672c\u7684\u63a8\u7406\u683c\u5f0f\u548c\u81ea\u6211\u68c0\u67e5\u4e60\u60ef,\u518d\u8ba9\u5f3a\u5316\u5b66\u4e60\u53bb\u8fdb\u4e00\u6b65\u4f18\u5316\u6b63\u786e\u7387\u548c\u63a8\u7406\u8d28\u91cf\u3002\u8fd9\u8bf4\u660eSFT\u4e0d\u518d\u53ea\u662f\u7ec8\u70b9\u524d\u7684\u6700\u540e\u4e00\u6b65\u5fae\u8c03,\u4e5f\u53ef\u4ee5\u662f\u66f4\u590d\u6742\u8bad\u7ec3\u6d41\u7a0b\u91cc\u4e00\u4e2a\u627f\u524d\u542f\u540e\u7684\u4e2d\u95f4\u9636\u6bb5,\u5177\u4f53\u89d2\u8272\u548c\u4f4d\u7f6e\u4f1a\u968f\u7740\u4efb\u52a1\u76ee\u6807(\u662f\u505a\u901a\u7528\u5bf9\u8bdd\u52a9\u624b,\u8fd8\u662f\u505a\u4e13\u7cbe\u7684\u63a8\u7406\u6a21\u578b)\u800c\u8c03\u6574\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u603b\u7684\u6765\u8bf4,SFT\u4f5c\u4e3a\u4e00\u79cd\u7814\u53d1\u8303\u5f0f\u7684\u6838\u5fc3\u4ef7\u503c\u5728\u4e8e,\u5b83\u7528\u76f8\u5bf9\u5c11\u91cf\u4f46\u7cbe\u5fc3\u6784\u9020\u7684\u76d1\u7763\u6570\u636e,\u9ad8\u6548\u5730\u628a\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u539f\u59cb\u80fd\u529b&#8221;\u5851\u5f62&#8221;\u6210\u7b26\u5408\u7279\u5b9a\u884c\u4e3a\u89c4\u8303\u7684\u6837\u5b50,\u662f\u76ee\u524d\u51e0\u4e4e\u6240\u6709\u4e3b\u6d41\u5bf9\u8bdd\u5f0f\u5927\u6a21\u578b\u8bad\u7ec3\u6d41\u7a0b\u91cc\u4e0d\u53ef\u6216\u7f3a\u3001\u4f46\u4e5f\u7edd\u975e\u5145\u5206\u7684\u4e00\u73af\u3002<\/p>\n<p dir=\"ltr\">\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<div class=\"flex justify-start ml-0.5 opacity-0 group-hover:opacity-100 group-focus-within:opacity-100 pointer-coarse:opacity-100 transition-opacity duration-150\" data-message-action-bar=\"\">\n<div class=\"text-text-300\">\n<div data-rs-index=\"12\" data-index=\"12\">\n<div tabindex=\"0\" role=\"article\" aria-setsize=\"14\" aria-posinset=\"13\" aria-label=\"Message 13 of 14\">\n<div data-test-render-count=\"2\">\n<div class=\"mb-1 mt-[var(--msg-gap,1.5rem)] group group\/message-row\">\n<div class=\"flex flex-col items-end gap-1\">\n<div class=\"group relative inline-flex gap-2 bg-bg-300 dark:[[data-darker-default]_&amp;]:bg-bg-000 rounded-xl pl-2.5 py-[var(--msg-bubble-py,0.625rem)] break-words text-text-100 transition-all max-w-[75ch] flex-col !px-[var(--msg-bubble-px,1rem)] max-w-[85%]\" data-user-message-bubble=\"true\">\n<div class=\"flex flex-row gap-2 relative\">\n<div class=\"flex-1\">\n<div class=\"font-large !font-user-message py-0.5 grid grid-cols-1 gap-2 relative [&amp;_ul]:!space-y-0 [&amp;_ol]:!space-y-0 [&amp;_ul]:pl-8 [&amp;_ol]:pl-8\" data-testid=\"user-message\">\n<h2 class=\"whitespace-pre-wrap break-words\">\u8bb2\u8bb2Muon<\/h2>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<div data-rs-index=\"13\" data-index=\"13\" data-last-message=\"true\">\n<div tabindex=\"-1\" role=\"article\" aria-setsize=\"14\" aria-posinset=\"14\" aria-label=\"Message 14 of 14\">\n<div data-test-render-count=\"1\">\n<div class=\"group group\/message-row\">\n<div class=\"group relative relative pb-[var(--msg-assistant-pb,0.75rem)]\" data-is-streaming=\"false\">\n<div class=\"font-claude-response relative leading-[1.65rem] [&amp;_pre&gt;div]:bg-bg-000\/50 [&amp;_pre&gt;div]:border-0.5 [&amp;_pre&gt;div]:border-border-400 [&amp;_.ignore-pre-bg&gt;div]:bg-transparent [&amp;_.standard-markdown_:is(p,blockquote,h1,h2,h3,h4,h5,h6)]:pl-2 [&amp;_.standard-markdown_:is(p,blockquote,ul,ol,h1,h2,h3,h4,h5,h6)]:pr-8 [&amp;_.progressive-markdown_:is(p,blockquote,h1,h2,h3,h4,h5,h6)]:pl-2 [&amp;_.progressive-markdown_:is(p,blockquote,ul,ol,h1,h2,h3,h4,h5,h6)]:pr-8\">\n<div>\n<div class=\"grid grid-rows-[auto_auto] min-w-0\">\n<div class=\"row-start-2 col-start-1 relative grid grid-rows-[auto_auto] isolate min-w-0\">\n<div class=\"row-start-1 col-start-1 relative z-[2] min-w-0\">\n<div>\n<div>\n<div class=\"standard-markdown grid-cols-1 grid [&amp;_&gt;_*]:min-w-0 gap-3 [&amp;_&gt;_*:last-child]:mb-0 print:block print:[&amp;_&gt;_*_+_*]:mt-3 standard-markdown\">\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">Muon\u662f\u8fd1\u4e24\u5e74\u5728\u5927\u6a21\u578b\u8bad\u7ec3\u5708\u5b50\u91cc\u5f88\u53d7\u5173\u6ce8\u7684\u4e00\u79cd\u65b0\u578b\u4f18\u5316\u5668,\u5168\u540d\u662f&#8221;Momentum Orthogonalized by Newton-Schulz&#8221;,\u6700\u65e9\u7531Keller Jordan\u5728NanoGPT\u8bad\u7ec3\u901f\u5ea6\u7eaa\u5f55(&#8220;speedrun&#8221;)\u7684\u5b9e\u8df5\u4e2d\u63d0\u51fa,\u540e\u6765\u88abMoonshotAI\u7b49\u673a\u6784\u9a8c\u8bc1\u53ef\u4ee5\u6269\u5c55\u5230\u771f\u6b63\u5927\u89c4\u6a21\u7684LLM\u9884\u8bad\u7ec3\u4e0a,\u76ee\u524d\u88ab\u8ba4\u4e3a\u662fAdam\/AdamW\u4e4b\u5916\u6700\u6709\u5e0c\u671b\u7684\u66ff\u4ee3\u65b9\u6848\u4e4b\u4e00\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u8981\u7406\u89e3Muon,\u5f97\u5148\u8bf4\u6e05\u695a\u5b83\u60f3\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898\u3002Adam\u548cAdamW\u662f\u8fc7\u53bb\u51e0\u5e74\u8bad\u7ec3\u5927\u6a21\u578b\u7684\u4e8b\u5b9e\u6807\u51c6\u4f18\u5316\u5668,\u6838\u5fc3\u601d\u8def\u662f\u7ed9\u6bcf\u4e2a\u53c2\u6570\u7ef4\u62a4\u4e00\u4e2a\u72ec\u7acb\u7684\u81ea\u9002\u5e94\u5b66\u4e60\u7387,\u6839\u636e\u68af\u5ea6\u7684\u4e00\u9636\u548c\u4e8c\u9636\u77e9\u4f30\u8ba1\u6765\u8c03\u6574\u66f4\u65b0\u6b65\u957f\u3002\u4f46Adam\u672c\u8d28\u4e0a\u662f\u628a\u6bcf\u4e2a\u53c2\u6570\u5f53\u4f5c\u72ec\u7acb\u7684\u6807\u91cf\u6765\u770b\u5f85\u548c\u8c03\u6574\u7684,\u6ca1\u6709\u5229\u7528\u5230\u795e\u7ecf\u7f51\u7edc\u91cc\u6743\u91cd\u77e9\u9635\u672c\u8eab\u7684\u7ed3\u6784\u4fe1\u606f\u2014\u2014\u800c\u73b0\u4ee3\u795e\u7ecf\u7f51\u7edc\u7684\u9690\u85cf\u5c42\u6743\u91cd\u6070\u6070\u5927\u591a\u662f\u4e8c\u7ef4\u77e9\u9635\u3002Muon\u7684\u51fa\u53d1\u70b9\u5c31\u662f:\u5982\u679c\u628a\u6743\u91cd\u77e9\u9635\u5f53\u4f5c\u4e00\u4e2a\u6574\u4f53\u7684\u51e0\u4f55\u5bf9\u8c61\u6765\u5904\u7406,\u800c\u4e0d\u662f\u62c6\u6563\u6210\u4e00\u5806\u4e92\u76f8\u72ec\u7acb\u7684\u6807\u91cf,\u662f\u4e0d\u662f\u80fd\u627e\u5230\u66f4\u597d\u7684\u66f4\u65b0\u65b9\u5f0f\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">Muon\u7684\u6838\u5fc3\u673a\u5236\u53ef\u4ee5\u6982\u62ec\u6210\u4e09\u6b65\u3002\u7b2c\u4e00\u6b65\u548c\u666e\u901a\u7684\u5e26\u52a8\u91cf\u7684SGD\u4e00\u6837,\u5148\u79ef\u7d2f\u4e00\u4e2a\u52a8\u91cf\u5316\u7684\u68af\u5ea6\u66f4\u65b0\u91cf\u3002\u7b2c\u4e8c\u6b65\u662f\u5173\u952e\u521b\u65b0:\u5bf9\u8fd9\u4e2a\u52a8\u91cf\u66f4\u65b0\u77e9\u9635\u505a&#8221;\u6b63\u4ea4\u5316&#8221;\u5904\u7406,\u4e5f\u5c31\u662f\u628a\u5b83\u66ff\u6362\u6210\u5947\u5f02\u503c\u5206\u89e3(SVD)\u91cc\u7684UV\u8f6c\u7f6e\u90e8\u5206,\u76f4\u89c2\u7406\u89e3\u5c31\u662f\u4fdd\u7559\u8fd9\u4e2a\u66f4\u65b0\u77e9\u9635\u7684&#8221;\u65b9\u5411\u4fe1\u606f&#8221;,\u4f46\u628a\u5404\u4e2a\u5947\u5f02\u503c\u90fd\u62c9\u5e73\u6210\u76f8\u8fd1\u7684\u5927\u5c0f,\u907f\u514d\u66f4\u65b0\u5728\u67d0\u4e9b\u65b9\u5411\u4e0a\u8fc7\u5ea6\u81a8\u80c0\u3001\u5728\u53e6\u4e00\u4e9b\u65b9\u5411\u4e0a\u53c8\u51e0\u4e4e\u4e0d\u52a8\u3002\u7b2c\u4e09\u6b65\u662f\u5de5\u7a0b\u4e0a\u7684\u5de7\u601d:\u76f4\u63a5\u7b97SVD\u4ee3\u4ef7\u5f88\u9ad8,\u6240\u4ee5Muon\u7528Newton-Schulz\u8fed\u4ee3(\u901a\u5e38\u4e94\u6b65\u5de6\u53f3)\u5728bfloat16\u7cbe\u5ea6\u4e0b\u53bb\u9ad8\u6548\u8fd1\u4f3c\u8fd9\u4e2a\u6b63\u4ea4\u5316\u8fc7\u7a0b,\u800c\u4e0d\u9700\u8981\u771f\u7684\u505a\u4e00\u6b21\u6602\u8d35\u7684\u77e9\u9635\u5206\u89e3\u3002\u8fd9\u4e09\u6b65\u5408\u8d77\u6765,\u5c31\u662f&#8221;\u52a8\u91cf&#8221;\u52a0&#8221;\u7ecfNewton-Schulz\u8fd1\u4f3c\u6b63\u4ea4\u5316&#8221;\u7684\u5b8c\u6574\u542b\u4e49\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u503c\u5f97\u5f3a\u8c03\u7684\u662f,Muon\u5e76\u4e0d\u662f\u8981\u53d6\u4ee3Adam\u6765\u4f18\u5316\u6240\u6709\u53c2\u6570,\u800c\u662f\u6709\u610f\u505a\u4e86\u5206\u5de5:\u5b83\u53ea\u7528\u6765\u4f18\u5316\u795e\u7ecf\u7f51\u7edc\u9690\u85cf\u5c42\u91cc\u7684\u4e8c\u7ef4\u6743\u91cd\u77e9\u9635,\u81f3\u4e8e\u50cfLayerNorm\u7684\u7f29\u653e\u53c2\u6570\u3001\u504f\u7f6e\u8fd9\u7c7b\u6807\u91cf\u6216\u5411\u91cf\u53c2\u6570,\u4ee5\u53ca\u6a21\u578b\u7684\u8f93\u5165\u5d4c\u5165\u5c42\u548c\u8f93\u51fa\u5c42,Jordan\u7684\u5efa\u8bae\u662f\u4ecd\u7136\u4ea4\u7ed9AdamW\u8fd9\u7c7b\u6807\u51c6\u65b9\u6cd5\u5904\u7406\u3002\u8fd9\u79cd&#8221;\u770b\u53c2\u6570\u7684\u5f62\u72b6\u51b3\u5b9a\u7528\u4ec0\u4e48\u4f18\u5316\u5668&#8221;\u7684\u6df7\u5408\u7b56\u7565,\u662fMuon\u5728\u5b9e\u8df5\u4e2d\u771f\u6b63\u88ab\u4f7f\u7528\u7684\u65b9\u5f0f,\u800c\u4e0d\u662f\u4e00\u4e2a\u5355\u4e00\u4f18\u5316\u5668\u5305\u6253\u5929\u4e0b\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u4ece\u6548\u679c\u4e0a\u770b,Muon\u6700\u65e9\u662f\u5728CIFAR-10\u8bad\u7ec3\u901f\u5ea6\u7eaa\u5f55\u548cNanoGPT\u901f\u8dd1\u8fd9\u7c7b\u5c0f\u89c4\u6a21\u3001\u8ffd\u6c42\u6781\u81f4\u8bad\u7ec3\u6548\u7387\u7684\u573a\u666f\u91cc\u8bc1\u660e\u81ea\u5df1\u7684\u2014\u2014\u5b83\u8fde\u7eed\u5237\u65b0\u4e86\u591a\u9879\u901f\u5ea6\u7eaa\u5f55,\u62a5\u544a\u7684\u6837\u672c\u6548\u7387\u63d0\u5347\u5927\u7ea6\u57281.35\u500d\u5de6\u53f3,\u800c\u4e14\u5b83\u5e26\u6765\u7684\u989d\u5916\u8ba1\u7b97\u5f00\u9500\u975e\u5e38\u5c0f,\u5728\u5178\u578b\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u573a\u666f\u4e0b\u4f4e\u4e8e1%,\u8fd9\u610f\u5473\u7740\u6362\u6389Adam\u57fa\u672c\u4e0d\u9700\u8981\u4ed8\u51fa\u989d\u5916\u7684\u663e\u8457\u4ee3\u4ef7\u3002\u771f\u6b63\u8ba9\u5b83\u53d7\u5230\u5e7f\u6cdb\u5173\u6ce8\u7684,\u662fMoonshotAI(\u6708\u4e4b\u6697\u9762)\u540e\u7eed\u53d1\u8868\u7684\u6280\u672f\u62a5\u544a,\u9a8c\u8bc1\u4e86Muon\u53ef\u4ee5\u6269\u5c55\u5230\u771f\u6b63\u7684\u5927\u6a21\u578b\u9884\u8bad\u7ec3\u89c4\u6a21\u3002\u4ed6\u4eec\u53d1\u73b0\u8981\u8ba9Muon\u5728\u5927\u89c4\u6a21\u4e0b\u7a33\u5b9a\u6709\u6548,\u9700\u8981\u505a\u4e24\u4e2a\u5173\u952e\u6539\u52a8:\u4e00\u662f\u7ed9Muon\u52a0\u4e0a\u6743\u91cd\u8870\u51cf(weight decay),\u5426\u5219\u6743\u91cd\u4f1a\u5728\u957f\u65f6\u95f4\u8bad\u7ec3\u4e2d\u65e0\u8282\u5236\u589e\u957f;\u4e8c\u662f\u9488\u5bf9\u4e0d\u540c\u5f62\u72b6\u7684\u6743\u91cd\u77e9\u9635,\u6309\u221amax(\u884c\u6570,\u5217\u6570)\u5bf9\u66f4\u65b0\u5e45\u5ea6\u505a\u5f52\u4e00\u5316\u7f29\u653e,\u8fd9\u6837\u624d\u80fd\u76f4\u63a5\u590d\u7528\u5df2\u7ecf\u8c03\u597d\u7684AdamW\u8d85\u53c2\u6570,\u4e0d\u7528\u91cd\u65b0\u8c03\u53c2\u3002\u4ed6\u4eec\u7528\u6539\u826f\u540e\u7684Muon\u8bad\u7ec3\u4e86\u4e00\u4e2a3B\/16B\u53c2\u6570\u7684\u6df7\u5408\u4e13\u5bb6\u6a21\u578bMoonlight,\u7528\u4e865.7\u4e07\u4ebftoken,\u62a5\u544a\u663e\u793a\u76f8\u6bd4AdamW,Muon\u8fbe\u5230\u540c\u7b49\u6548\u679c\u5927\u7ea6\u53ea\u9700\u898152%\u7684\u8bad\u7ec3\u7b97\u529b,\u5373\u63a5\u8fd1\u4e24\u500d\u7684\u8ba1\u7b97\u6548\u7387\u63d0\u5347,\u540c\u65f6\u5728MMLU\u3001GSM8K\u7b49\u8bc4\u6d4b\u4e0a\u4e5f\u8d85\u8fc7\u4e86\u540c\u89c4\u6a21\u7684\u5bf9\u6bd4\u6a21\u578b\u3002<\/p>\n<p class=\"font-claude-response-body break-words whitespace-normal\" dir=\"ltr\">\u653e\u5728\u66f4\u5927\u7684\u7814\u53d1\u8303\u5f0f\u89c6\u89d2\u770b,Muon\u4ee3\u8868\u7684\u662f\u4e00\u7c7b&#8221;\u5229\u7528\u53c2\u6570\u77e9\u9635\u7ed3\u6784\u4fe1\u606f&#8221;\u7684\u4f18\u5316\u5668\u7814\u7a76\u65b9\u5411,\u8ddfAdam\u8fd9\u7c7b\u9010\u53c2\u6570\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7684\u601d\u8def\u662f\u4e0d\u540c\u7684\u8bbe\u8ba1\u54f2\u5b66\u3002\u76ee\u524d\u8fd9\u4e2a\u65b9\u5411\u4ecd\u7136\u975e\u5e38\u6d3b\u8dc3,\u5305\u62ec\u5bf9Newton-Schulz\u8fed\u4ee3\u7cfb\u6570\u7684\u8fdb\u4e00\u6b65\u8c03\u4f18\u3001\u81ea\u9002\u5e94\u7248\u672c(\u6bd4\u5982AdamMuon\u8fd9\u7c7b\u7ed3\u5408\u4e24\u8005\u4f18\u70b9\u7684\u5c1d\u8bd5)\u7b49\u540e\u7eed\u5de5\u4f5c\u90fd\u5728\u9646\u7eed\u51fa\u73b0,\u4e5f\u5df2\u7ecf\u6709\u4e00\u4e9b\u516c\u5f00\u7684\u5927\u6a21\u578b\u8bad\u7ec3(\u5305\u62ec\u6708\u4e4b\u6697\u9762\u7684Kimi\u7cfb\u5217)\u91c7\u7528\u4e86Muon\u6216\u5176\u53d8\u4f53\u4f5c\u4e3a\u751f\u4ea7\u73af\u5883\u7684\u4f18\u5316\u5668\u3002\u4e0d\u8fc7\u9700\u8981\u8bf4\u660e\u7684\u662f,\u8fd9\u4e2a\u9886\u57df\u8fdb\u5c55\u5f88\u5feb,\u76ee\u524d\u516c\u5f00\u7684\u8bc1\u636e\u5927\u591a\u8fd8\u662f\u6765\u81ea\u4e2a\u522b\u673a\u6784\u548c\u56e2\u961f\u7684\u62a5\u544a,\u4e1a\u754c\u662f\u5426\u4f1a\u50cf\u5f53\u5e74Adam\u53d6\u4ee3SGD\u90a3\u6837\u666e\u904d\u8f6c\u5411Muon,\u8fd8\u9700\u8981\u66f4\u591a\u72ec\u7acb\u590d\u73b0\u548c\u66f4\u957f\u65f6\u95f4\u7684\u68c0\u9a8c\u3002<\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<div data-rs-index=\"12\" data-index=\"12\">\n<div tabindex=\"-1\" role=\"article\" aria-setsize=\"14\" aria-posinset=\"13\" aria-label=\"Message 13 of 14\">\n<div data-test-render-count=\"2\">\n<div class=\"mb-1 mt-[var(--msg-gap,1.5rem)] group group\/message-row\">\n<div class=\"flex flex-col items-end gap-1\">\n<div class=\"group relative inline-flex gap-2 bg-bg-300 dark:[[data-darker-default]_&amp;]:bg-bg-000 rounded-xl pl-2.5 py-[var(--msg-bubble-py,0.625rem)] break-words text-text-100 transition-all max-w-[75ch] flex-col !px-[var(--msg-bubble-px,1rem)] max-w-[85%]\" data-user-message-bubble=\"true\">\n<div class=\"flex flex-row gap-2 relative\">\n<div class=\"flex-1\">\n<div class=\"font-large !font-user-message py-0.5 grid grid-cols-1 gap-2 relative [&amp;_ul]:!space-y-0 [&amp;_ol]:!space-y-0 [&amp;_ul]:pl-8 [&amp;_ol]:pl-8\" data-testid=\"user-message\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u8bb2\u8bb2SFT\u8fd9\u6837\u7684\u7814\u53d1\u8303\u5f0f SFT\u5168\u79f0\u662fS<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"colormag_page_container_layout":"default_layout","colormag_page_sidebar_layout":"default_layout","footnotes":""},"categories":[24],"tags":[],"class_list":["post-7532","post","type-post","status-publish","format-standard","hentry","category-value_docs"],"_links":{"self":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/7532","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/comments?post=7532"}],"version-history":[{"count":0,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/7532\/revisions"}],"wp:attachment":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/media?parent=7532"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/categories?post=7532"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/tags?post=7532"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}