{"id":36792,"date":"2025-01-18T18:00:11","date_gmt":"2025-01-18T10:00:11","guid":{"rendered":"https:\/\/17aitech.com\/?p=36792"},"modified":"2025-01-18T18:00:11","modified_gmt":"2025-01-18T10:00:11","slug":"%e8%85%be%e8%ae%af%e6%b7%b7%e5%85%83%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%a0%b8%e5%bf%83%e8%ae%ba%e6%96%87%e6%9b%9d%e5%85%89%ef%bc%9ascaling-law%e3%80%81moe%e3%80%81%e5%90%88%e6%88%90%e6%95%b0%e6%8d%ae","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=36792","title":{"rendered":"\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587\u66dd\u5149\uff1aScaling law\u3001MoE\u3001\u5408\u6210\u6570\u636e\u4ee5\u53ca\u66f4\u591a"},"content":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587\u66dd\u5149\uff1aScaling law\u3001MoE\u3001\u5408\u6210\u6570\u636e\u4ee5\u53ca\u66f4\u591a<\/a><\/p>\n

\u968f\u7740 ChatGPT \u7684\u6a2a\u7a7a\u51fa\u4e16\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u529b\u5f00\u59cb\u5728\u5404\u9879\u9886\u57df\uff08\u4f20\u7edf NLP\u3001\u6570\u5b66\u3001\u4ee3\u7801\u7b49\uff09\u5f97\u5230\u5e7f\u6cdb\u9a8c\u8bc1\uff0c\u76ee\u524d\u5df2\u7ecf\u6df1\u523b\u5f71\u54cd\u5230\u817e\u8baf\u6df7\u5143\u56e2\u961f\u65e5\u5e38\u751f\u6d3b\u7684\u65b9\u65b9\u9762\u9762\u3002\u817e\u8baf\u6df7\u5143\u56e2\u961f\u957f\u671f\u81f4\u529b\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a2\u7d22\u4e4b\u8def\uff0c\u5927\u6a21\u578b\u751f\u4ea7\u7684\u5404\u4e2a\u73af\u8282\u5f00\u5c55\u7814\u7a76\u521b\u65b0\u4ee5\u63d0\u5347\u5176\u57fa\u7840\u80fd\u529b\uff0c\u5e76\u5c06\u6df7\u5143\u5927\u6a21\u578b\u7684\u80fd\u529b\u8ddf\u4e1a\u52a1\u505a\u6df1\u5ea6\u7ed3\u5408\uff0c\u8ba9\u751f\u6210\u5f0f AI \u6210\u4e3a\u4e1a\u52a1\u589e\u957f\u7684\u653e\u5927\u5668\u3002<\/p>\n

\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bbe\u8ba1\u3001\u8bad\u7ec3\u548c\u4f18\u5316\u662f\u4e00\u9879\u590d\u6742\u7684\u7cfb\u7edf\u5de5\u7a0b\uff0c\u6d89\u53ca\u5230\u6a21\u578b\u7ed3\u6784\u521b\u65b0\u3001\u8bad\u7ec3\u8303\u5f0f\u4f18\u5316\u3001\u6570\u636e\u83b7\u53d6\u548c\u8bc4\u6d4b\u8bbe\u8ba1\u3001\u5173\u952e\u80fd\u529b\u63d0\u5347\u548c\u6311\u6218\u6027\u95ee\u9898\u7684\u89e3\u51b3\u7b49\u65b9\u65b9\u9762\u9762\u3002\u817e\u8baf\u6df7\u5143\u56e2\u961f\u5728\u5927\u6a21\u578b\u7814\u7a76\u63a2\u7d22\u4e2d\u79ef\u7d2f\u4e86\u4e30\u5bcc\u7684\u5b9e\u6218\u7ecf\u9a8c\u548c\u521b\u65b0\u6027\u7684\u7814\u7a76\u6210\u679c\uff0c\u76ee\u524d\u7d2f\u8ba1\u53d1\u8868\u8fd1\u767e\u7bc7\u5b66\u672f\u8bba\u6587\uff0c\u4e3a\u63a8\u52a8\u6280\u672f\u7684\u5f00\u653e\u5171\u4eab\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u4e5f\u628a\u5bf9\u5e94\u6210\u679c\u4ee5\u8bba\u6587\u3001\u5f00\u6e90\u6a21\u578b\u548c\u6280\u672f\u62a5\u544a\u7b49\u5f62\u5f0f\u5206\u4eab\u7ed9\u5927\u6a21\u578b\u793e\u533a\u7684\u7814\u7a76\u8005\u3002<\/p>\n

11 \u6708 5 \u65e5\uff0c\u817e\u8baf\u6df7\u5143\u53d1\u5e03\u4e86\u4e1a\u754c\u6700\u5927\u53c2\u6570\u89c4\u6a21\u7684 MoE \u5f00\u6e90\u6a21\u578b\u817e\u8baf\u6df7\u5143 Large\u3002\u6df7\u5143 Large \u662f\u76ee\u524d\u6700\u5927\u6700\u5f3a\u7684\u5f00\u6e90 Transformer \u7ed3\u6784\u7684 MoE \u5927\u8bed\u8a00\u6a21\u578b\uff0c\u5728\u9ad8\u8d28\u91cf\u5408\u6210\u6570\u636e\u3001\u5148\u8fdb\u7684\u6a21\u578b\u67b6\u6784\u548c\u6df7\u5408\u4e13\u5bb6\u8def\u7531\u7b56\u7565\u3001\u4ee5\u53ca\u4f18\u5316\u7684\u6a21\u578b\u8bad\u7ec3\u7b56\u7565\u5171\u540c\u52a0\u6301\u4e0b\uff0c\u817e\u8baf\u6df7\u5143 Large \u5728\u5e7f\u6cdb\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e0b\u83b7\u5f97\u4e86\u4f18\u5f02\u7684\u6027\u80fd\u3002\u5bf9\u5e94\u7684\u5f00\u6e90\u6a21\u578b\u548c\u6280\u672f\u62a5\u544a\u8be6\u8ff0\u4e86 Hunyuan-Large \u5f3a\u608d\u80fd\u529b\u7684\u6280\u672f\u57fa\u7840\u7ec6\u8282\uff0c\u53ef\u8c13\u817e\u8baf\u6df7\u5143\u56e2\u961f\u591a\u4e2a\u524d\u6cbf\u7814\u7a76\u7684\u96c6\u5927\u6210\u3002\u89c1\u6280\u672f\u62a5\u544a\u300aHunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent\u300b\u3002<\/p>\n

\u5927\u6a21\u578b\u7684\u953b\u9020\u662f\u4e00\u9879\u7cfb\u7edf\u5de5\u7a0b\uff0c\u5728\u8fd9\u4e00\u8fc7\u7a0b\u4e2d\uff0c\u817e\u8baf\u9009\u62e9\u4ece\u96f6\u5f00\u59cb\u81ea\u7814\u7684\u8def\u7ebf\uff0c\u4ed6\u4eec\u5982\u4f55\u5728\u77ed\u65f6\u95f4\u5185\u642d\u5efa\u51fa\u4e07\u4ebf\u53c2\u6570\u89c4\u6a21\u6a21\u578b\uff1f\u53c8\u5982\u4f55\u7a81\u7834\u7b97\u529b\u6781\u9650\uff0c\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u4e0a\u505a\u529f\u592b\uff1f\u9ad8\u6548\u4ea7\u51fa\u591a\u6b3e\u4e1a\u754c\u9886\u5148\u7684\u6a21\u578b\uff1f\u6700\u8fd1\uff0c\u673a\u5668\u4e4b\u5fc3\u62ff\u5230\u4e86\u817e\u8baf\u6df7\u5143\u56e2\u961f\u7814\u53d1\u56e2\u961f\u51e0\u7bc7\u6838\u5fc3\u8bba\u6587\uff0c\u53ef\u4ee5\u5e26\u5927\u5bb6\u4e00\u7aa5\u7a76\u7adf\u3002<\/p>\n

\u5de5\u6b32\u5584\u5176\u4e8b\uff0c\u5fc5\u5148\u5229\u5176\u5668\uff0c\u5927\u6a21\u578b\u65f6\u4ee3\u6781\u5176\u9ad8\u6602\u7684\u8bad\u7ec3\u6210\u672c\u4fc3\u4f7f\u817e\u8baf\u6df7\u5143\u56e2\u961f\u6c89\u4e0b\u5fc3\u6df1\u5165\u63a2\u7d22\u5927\u6a21\u578b\u8bad\u7ec3\u4e2d\u53c2\u6570\u91cf\u3001\u91cd\u8981\u8d85\u53c2\u3001\u8bad\u7ec3 token \u6570\u548c\u6700\u7ec8\u6027\u80fd\u7b49\u5173\u952e\u8981\u7d20\u4e4b\u95f4\u7684\u89c4\u5f8b\u3002\u817e\u8baf\u6df7\u5143\u5728\u300aSurge Phenomenon in Optimal Learning Rate and Batch Size Scaling\u300b\u4e2d\u6df1\u5165\u63a2\u7d22\u4e86\u5927\u6a21\u578b\u6279\u5927\u5c0f\u548c\u6700\u4f73\u5b66\u4e60\u7387\u4e4b\u95f4\u7684 scaling law\uff0c\u6316\u6398\u51fa\u7684\u89c4\u5f8b\u4e5f\u6210\u4e3a Hunyuan \u5927\u6a21\u578b\u7cfb\u5217\u9ad8\u6548\u8bad\u7ec3\u7684\u7406\u8bba\u6307\u5bfc\u3002<\/p>\n

\u4f18\u8d8a\u7684\u6a21\u578b\u7ed3\u6784\u662f\u5927\u6a21\u578b\u6027\u80fd\u7684\u57fa\u77f3\u3002\u817e\u8baf\u6df7\u5143\u56e2\u961f\u4f5c\u4e3a\u56fd\u5185\u6700\u65e9\u90e8\u7f72\u8d85\u5927\u89c4\u6a21 MoE \u67b6\u6784\u5927\u6a21\u578b\u7684\u56e2\u961f\u4e4b\u4e00\uff0c\u4e00\u76f4\u5728\u4e0d\u65ad\u5730\u8fdb\u884c\u6a21\u578b\u67b6\u6784\u4e0a\u7684\u521b\u65b0\uff0c\u5176\u5de5\u4f5c\u300aHMoE: Heterogeneous Mixture of Experts for Language Modeling\u300b\u521b\u65b0\u6027\u5730\u63d0\u51fa\u4e86\u5f02\u6784\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\u5e76\u9a8c\u8bc1\u5176\u6709\u6548\u6027\uff0c\u4e3a\u5927\u6a21\u578b\u793e\u533a\u5e26\u6765\u5f88\u6709\u524d\u666f\u7684\u65b0\u8bbe\u8ba1\u601d\u8def\u3002\u767e\u82b1\u9f50\u653e\u7684\u63a2\u7d22\u624d\u80fd\u6fc0\u53d1\u66f4\u591a\u98a0\u8986\u6027\u7684\u521b\u65b0\u601d\u8003\u3002<\/p>\n

\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u53ca\u5408\u7406\u5168\u9762\u7684\u8bc4\u6d4b\u662f\u5927\u6a21\u578b\u8bad\u7ec3\u7684\u91cd\u4e2d\u4e4b\u91cd\u3002\u817e\u8baf\u6df7\u5143\u56e2\u961f\u5728\u5408\u6210\u6570\u636e\u548c\u6a21\u578b\u8bc4\u6d4b\u4e0a\u8fdb\u884c\u4e86\u5145\u5206\u7684\u79ef\u7d2f\uff0c\u529f\u592b\u4f53\u73b0\u4e8e\u5fae\u672b\u5904\uff0c\u770b\u4f3c\u67af\u71e5\u7684\u5de5\u4f5c\u4e2d\u4e5f\u80fd\u51dd\u7ed3\u51fa\u6280\u672f\u6d1e\u89c1\u4e4b\u82b1\u3002\u817e\u8baf\u6df7\u5143\u5728\u300aDINGO: Towards Diverse and Fine-grained Instruction-Following Evaluation\u300b\u4e2d\u63d0\u4f9b\u4e86\u4e00\u4e2a\u7ec6\u7c92\u5ea6\u4e14\u591a\u6837\u5316\u7684\u6307\u4ee4\u9075\u5faa\u8bc4\u4f30\u6570\u636e\u96c6\uff0c\u57fa\u4e8e\u4ece\u771f\u5b9e\u7528\u6237\u8bf7\u6c42\u4e2d\u603b\u7ed3\u7684\u624b\u52a8\u6ce8\u91ca\u7684\u591a\u7ea7\u5206\u7c7b\u6811\uff0c\u4e3a\u5927\u6a21\u578b\u63d0\u4f9b\u66f4\u5177\u6311\u6218\u7684\u5168\u9762\u7684\u8bc4\u4f30\u3002<\/p>\n

\u6700\u540e\uff0c\u6df7\u5143\u56e2\u961f\u4e5f\u5bf9\u5927\u6a21\u578b\u7ecf\u5178\u80fd\u529b\u548c\u6311\u6218\uff0c\u5982\u5e7b\u89c9\u3001\u957f\u6587\u80fd\u529b\u7b49\u65b9\u9762\u8fdb\u884c\u4e86\u957f\u671f\u7684\u7814\u7a76\u3002\u4ee5\u300aTruth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning\u300b\u4e3a\u4ee3\u8868\uff0c\u5728\u5927\u6a21\u578b\u548c\u591a\u6a21\u6001\u5927\u6a21\u578b\u5e7b\u89c9\u95ee\u9898\u7684\u6570\u636e\u96c6\u6784\u5efa\u3001\u68c0\u6d4b\u548c\u7f13\u89e3\u4e0a\u8fdb\u884c\u4e86\u591a\u9879\u521b\u65b0\u6027\u7814\u7a76\uff0c\u65e8\u5728\u63d0\u5347\u6df7\u5143\u5927\u6a21\u578b\u5728\u771f\u5b9e\u573a\u666f\u4e0b\u7684\u53ef\u9760\u6027\u548c\u53ef\u7528\u6027\u3002<\/p>\n

\u4ee5\u4e0b\u662f\u4e0a\u8ff0\u4e94\u7bc7\u5de5\u4f5c\u7684\u8be6\u7ec6\u89e3\u8bfb\uff1a<\/p>\n

\u00a0Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent<\/p>\n

arXiv\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2411.02265<\/a><\/p>\n

\u5728\u672c\u6587\u4e2d\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u4ecb\u7ecd\u4e86 Hunyuan-Large\uff0c\u8fd9\u662f\u76ee\u524d\u6700\u5927\u7684\u5f00\u6e90\u57fa\u4e8e Transformer \u7684\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff0c\u603b\u53c2\u6570\u91cf\u4e3a 389B\uff0c\u6fc0\u6d3b\u53c2\u6570\u91cf\u4e3a 52B\uff0c\u80fd\u591f\u5904\u7406\u591a\u8fbe 256K \u7684\u4e0a\u4e0b\u6587 tokens\u3002Hunyuan-Large \u7684\u5173\u952e\u6280\u672f\u5305\u62ec\u6bd4\u4ee5\u5f80\u5de5\u4f5c\u5927\u51e0\u4e2a\u6570\u91cf\u7ea7\u7684\u5927\u89c4\u6a21\u5408\u6210\u6570\u636e\u3001\u4f18\u5316\u7684\u6df7\u5408\u4e13\u5bb6\u8def\u7531\u7b56\u7565\u3001\u9ad8\u6548\u7684 KV cache \u538b\u7f29\u6280\u672f\u548c\u4e13\u5bb6\u7279\u5316\u7684\u5b66\u4e60\u7387\u7b56\u7565\u7b49\uff0c\u7ec6\u8282\u6ee1\u6ee1\u3002\u6b64\u5916\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u8fd8\u7814\u7a76\u4e86\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\u7684 scaling law \u548c\u5b66\u4e60\u7387\uff0c\u4e3a\u672a\u6765\u7684 MoE \u5927\u6a21\u578b\u5f00\u53d1\u548c\u4f18\u5316\u63d0\u4f9b\u4e86\u5b9d\u8d35\u7684\u89c1\u89e3\u548c\u6307\u5bfc\u3002\u4e0b\u56fe\u7ed9\u51fa\u4e86 Hunyuan-Large \u4e2d\u7684 Recycle routing \u8def\u7531\u7b56\u7565\u793a\u610f\u56fe\u3002<\/p>\n

\u00a0<\/a><\/p>\n

\u817e\u8baf\u6df7\u5143\u56e2\u961f\u5bf9 Hunyuan-Large \u5728\u5404\u79cd\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u5353\u8d8a\u6027\u80fd\u8fdb\u884c\u4e86\u5168\u9762\u8bc4\u4f30\uff0c\u5305\u62ec\u8bed\u8a00\u7406\u89e3\u4e0e\u751f\u6210\u3001\u903b\u8f91\u63a8\u7406\u3001\u6570\u5b66\u95ee\u9898\u3001\u4ee3\u7801\u4efb\u52a1\u3001\u957f\u4e0a\u4e0b\u6587\u548c\u7efc\u5408\u4efb\u52a1\u7b49\u3002\u5728\u8fd9\u4e9b\u65b9\u9762\u4e0a\uff0cHunyuan-Large \u4f18\u4e8e LLama3.1-70B\uff0c\u5e76\u4e14\u5728\u4e0e\u663e\u8457\u66f4\u5927\u7684 LLama3.1-405B \u6a21\u578b\u76f8\u6bd4\u65f6\u8868\u73b0\u51fa\u76f8\u5f53\u7684\u6027\u80fd\u3002<\/p>\n

<\/a><\/p>\n

\u76ee\u524d Hunyuan-Large \u7684\u4ee3\u7801\u548c\u6a21\u578b\u5747\u5df2\u53d1\u5e03\uff0c\u6b22\u8fce\u5404\u4f4d\u7814\u7a76\u8005\u548c\u5f00\u53d1\u8005\u4f7f\u7528\uff0c\u4e5f\u671f\u671b\u6df7\u5143\u56e2\u961f\u7684\u5f00\u6e90\u6210\u679c\u80fd\u591f\u4fc3\u8fdb\u672a\u6765\u7684\u5927\u6a21\u578b\u7814\u7a76\u521b\u65b0\u548c\u843d\u5730\u5e94\u7528\u3002<\/p>\n

\u00a0Surge Phenomenon in Optimal Learning Rate and Batch Size Scaling<\/p>\n

NeurIPS-2024<\/p>\n

\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2405.14578<\/a><\/p>\n

\u5728\u5f53\u524d\u7684\u6df1\u5ea6\u5b66\u4e60\u4efb\u52a1\u4e2d\uff0cAdam \u98ce\u683c\u7684\u4f18\u5316\u5668\uff08\u5982 Adam\u3001Adagrad\u3001RMSProp\u3001Adafactor \u548c Lion \u7b49\uff09\u5df2\u88ab\u5e7f\u6cdb\u7528\u4f5c SGD \u98ce\u683c\u4f18\u5316\u5668\u7684\u66ff\u4ee3\u54c1\u3002\u8fd9\u4e9b\u4f18\u5316\u5668\u901a\u5e38\u4f7f\u7528\u68af\u5ea6\u7684\u7b26\u53f7\u6765\u66f4\u65b0\u6a21\u578b\u53c2\u6570\uff0c\u4ece\u800c\u4ea7\u751f\u66f4\u7a33\u5b9a\u7684\u6536\u655b\u66f2\u7ebf\u3002\u5b66\u4e60\u7387\uff08learning rate\uff09\u548c\u6279\u5927\u5c0f\uff08batch size\uff09\u662f\u4f18\u5316\u5668\u6700\u5173\u952e\u7684\u8d85\u53c2\u6570\uff0c\u9700\u8981\u4ed4\u7ec6\u8c03\u6574\u4ee5\u5b9e\u73b0\u6709\u6548\u7684\u6536\u655b\u3002\u5148\u524d\u7684\u7814\u7a76\u8868\u660e\uff0c\u5bf9\u4e8e SGD \u98ce\u683c\u7684\u4f18\u5316\u5668\uff0c\u6700\u4f73\u5b66\u4e60\u7387\u968f\u7740\u6279\u91cf\u5927\u5c0f\u7ebf\u6027\u589e\u52a0\u6216\u9075\u5faa\u7c7b\u4f3c\u7684\u89c4\u5219\u3002\u7136\u800c\uff0c\u8fd9\u4e00\u7ed3\u8bba\u5e76\u4e0d\u9002\u7528\u4e8e Adam \u98ce\u683c\u7684\u4f18\u5316\u5668\u3002<\/p>\n

\u5728\u672c\u6587\u4e2d\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u901a\u8fc7\u7406\u8bba\u5206\u6790\u548c\u5e7f\u6cdb\u7684\u5b9e\u9a8c\uff0c\u9610\u660e\u4e86 Adam \u98ce\u683c\u4f18\u5316\u5668\u7684\u6700\u4f73\u5b66\u4e60\u7387\u4e0e\u6279\u5927\u5c0f\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u9996\u5148\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u63d0\u51fa\u4e86\u68af\u5ea6\u7b26\u53f7\u60c5\u51b5\u4e0b batch size \u4e0e\u6700\u4f73\u5b66\u4e60\u7387\u4e4b\u95f4\u7684\u7f29\u653e\u89c4\u5f8b\uff0c\u5e76\u8bc1\u660e\u4e86\u968f\u7740 batch size \u7684\u589e\u52a0\uff0c\u6700\u4f73\u5b66\u4e60\u7387\u5148\u4e0a\u5347\u540e\u4e0b\u964d\u3002\u6b64\u5916\uff0c\u968f\u7740\u8bad\u7ec3\u7684\u8fdb\u884c\uff0c\u5cf0\u503c\u5c06\u9010\u6e10\u5411\u66f4\u5927\u7684 batch size \u79fb\u52a8\u3002\u5176\u6b21\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u5728\u5404\u79cd\u8ba1\u7b97\u673a\u89c6\u89c9\u548c\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u9a8c\u8bc1\u4e86\u8fd9\u4e00\u7f29\u653e\u89c4\u5f8b\u7684\u6b63\u786e\u6027\u3002<\/p>\n

\u00a0<\/a><\/p>\n

\u817e\u8baf\u6df7\u5143\u56e2\u961f\u5145\u5206\u5229\u7528\u5de5\u4e1a\u754c\u7684\u7b97\u529b\u4f18\u52bf\uff0c\u5728\u8bf8\u591a\u4efb\u52a1\u3001\u6a21\u578b\u548c\u53c2\u6570\u8bbe\u7f6e\u4e0a\u8fdb\u884c\u4e86\u5e7f\u6cdb\u7684\u7f29\u653e\u89c4\u5f8b\u9a8c\u8bc1\u5b9e\u9a8c\uff0c\u7ec6\u8282\u8be6\u89c1\u8bba\u6587\u3002\u8fd9\u4e9b\u7ed3\u8bba\u4e0d\u4ec5\u4e3a\u817e\u8baf\u6df7\u5143\u56e2\u961f\u672a\u6765\u8bad\u7ec3 Hunyuan-Large \u4ee5\u53ca\u5176\u5b83 Hunyuan \u7cfb\u5217\u6a21\u578b\u63d0\u4f9b\u4e86\u575a\u5b9e\u7684\u7406\u8bba\u57fa\u7840\u548c\u7ecf\u9a8c\u7ed3\u8bba\uff0c\u4e5f\u5c06\u5bf9 LLM \u793e\u533a\u672a\u6765\u7684\u5927\u6a21\u578b\u8bad\u7ec3\u63d0\u4f9b\u6307\u5f15\u548c\u542f\u53d1\u3002<\/p>\n

<\/a><\/p>\n

HMoE: Heterogeneous Mixture of Experts for Language Modeling<\/p>\n

arXiv<\/p>\n

\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2408.10681<\/a><\/p>\n

\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08Mixture of Experts, MoE\uff09\u901a\u8fc7\u9009\u62e9\u6027\u6fc0\u6d3b\u5927\u6a21\u578b\u53c2\u6570\u7684\u5b50\u96c6\uff0c\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u6027\u80fd\u548c\u8ba1\u7b97\u6548\u7387\u3002\u4f20\u7edf\u7684 MoE \u6a21\u578b\u4f7f\u7528\u540c\u8d28\u4e13\u5bb6\uff0c\u6bcf\u4e2a\u4e13\u5bb6\u5177\u6709\u76f8\u540c\u7684\u5bb9\u91cf\u3002\u7136\u800c\uff0c\u8003\u8651\u5230\u8f93\u5165\u6570\u636e\u7684\u590d\u6742\u6027\u548c\u5dee\u5f02\u6027\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u9700\u8981\u6709\u4e0d\u540c\u80fd\u529b\u7684\u4e13\u5bb6\uff0c\u800c\u540c\u8d28 MoE \u5f71\u54cd\u4e86\u6709\u6548\u7684\u4e13\u5bb6\u4e13\u95e8\u5316\u548c\u9ad8\u6548\u7684\u53c2\u6570\u5229\u7528\u3002\u5728\u8fd9\u9879\u7814\u7a76\u4e2d\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u5f02\u8d28\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08Heterogeneous Mixture of Experts, HMoE\uff09\uff0c\u4e0d\u540c\u4e13\u5bb6\u5728\u89c4\u6a21\u4e0a\u6709\u6240\u4e0d\u540c\uff0c\u56e0\u6b64\u4e5f\u80fd\u6fc0\u53d1\u51fa\u66f4\u591a\u6837\u7684\u80fd\u529b\u3002\u8fd9\u79cd\u5f02\u8d28\u6027\u5141\u8bb8\u66f4\u4e13\u4e1a\u7684\u4e13\u5bb6\u66f4\u6709\u6548\u5730\u8003\u8651\u5230\u4e0d\u540c\u7684 token \u7684\u590d\u6742\u6027\u3002\u4e3a\u4e86\u89e3\u51b3\u4e13\u5bb6\u6fc0\u6d3b\u4e0d\u5e73\u8861\u7684\u95ee\u9898\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u8bad\u7ec3\u76ee\u6807\uff0c\u9f13\u52b1\u66f4\u9891\u7e41\u5730\u6fc0\u6d3b\u8f83\u5c0f\u7684\u4e13\u5bb6\uff0c\u4ece\u800c\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\u548c\u53c2\u6570\u5229\u7528\u7387\u3002<\/p>\n

<\/a>\u00a0<\/p>\n

\u5e7f\u6cdb\u7684\u5b9e\u9a8c\u8868\u660e\uff0cHMoE \u5728\u6fc0\u6d3b\u66f4\u5c11\u53c2\u6570\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0\u4e86\u66f4\u4f4e\u7684\u635f\u5931\uff0c\u5e76\u5728\u5404\u79cd\u9884\u8bad\u7ec3\u8bc4\u4f30\u57fa\u51c6\u4e0a\u4f18\u4e8e\u4f20\u7edf\u7684\u540c\u8d28 MoE \u6a21\u578b\u3002<\/p>\n

<\/a>\u00a0<\/p>\n

<\/a><\/p>\n

MoE \u7ed3\u6784\u8bbe\u8ba1\u4f18\u5316\u662f\u5176\u80fd\u529b\u63d0\u5347\u7684\u5173\u952e\u3002\u6df7\u5143\u56e2\u961f\u5728\u5927\u6a21\u578b\u7ed3\u6784\u8bbe\u8ba1\u548c\u4f18\u5316\u4e0a\u6295\u5165\u4e86\u5f88\u591a\u7cbe\u529b\uff0c\u4e5f\u5c1d\u8bd5\u4e86\u591a\u79cd\u65b0\u67b6\u6784\u6280\u672f\u65b9\u5411\uff0c\u672a\u6765\u5c06\u9010\u6b65\u5f00\u653e\u5e76\u5206\u4eab\u817e\u8baf\u6df7\u5143\u56e2\u961f\u65b0\u7684\u7814\u7a76\u6210\u679c\uff0c\u5e0c\u671b\u80fd\u4e3a\u5927\u6a21\u578b\u793e\u533a\u5e26\u6765\u66f4\u591a\u767e\u82b1\u9f50\u653e\u7684\u65b0\u9896\u601d\u8def\u3002<\/p>\n

DINGO: Towards Diverse and Fine-Grained Instruction-Following Evaluation<\/p>\n

AAAI-2024<\/p>\n

\u94fe\u63a5\uff1ahttps:\/\/ojs.aaai.org\/index.php\/AAAI\/article\/view\/29768<\/a><\/p>\n

\u7b80\u4ecb\uff1a\u6307\u4ee4\u9075\u5faa\u5bf9\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u652f\u6301\u591a\u6837\u5316\u7684\u7528\u6237\u8bf7\u6c42\u5c24\u4e3a\u91cd\u8981\uff0c\u8fd9\u79cd\u5927\u6a21\u578b\u5728\u771f\u5b9e\u5e94\u7528\u4e0a\u7684\u57fa\u672c\u80fd\u529b\u3002\u5c3d\u7ba1\u73b0\u6709\u7684\u5de5\u4f5c\u5728\u4f7f LLM \u4e0e\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\u65b9\u9762\u53d6\u5f97\u4e86\u4e00\u5b9a\u8fdb\u5c55\uff0c\u4f46\u7531\u4e8e\u73b0\u5b9e\u4e16\u754c\u7528\u6237\u6307\u4ee4\u7684\u590d\u6742\u6027\u548c\u591a\u6837\u6027\uff0c\u8bc4\u4f30\u5b83\u4eec\u7684\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u4ecd\u7136\u662f\u4e00\u4e2a\u6311\u6218\u3002\u73b0\u6709\u7684\u8bc4\u4f30\u65b9\u6cd5\u867d\u7136\u5173\u6ce8\u5927\u6a21\u578b\u7684\u4e00\u822c\u6280\u80fd\uff0c\u4f46\u5b58\u5728\u4e24\u4e2a\u4e3b\u8981\u7f3a\u70b9\uff0c\u5373\u7f3a\u4e4f\u7ec6\u7c92\u5ea6\u7684\u4efb\u52a1\u7ea7\u8bc4\u4f30\u548c\u4f9d\u8d56\u5355\u4e00\u7684\u6307\u4ee4\u8868\u8fbe\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u672c\u6587\u6784\u5efa\u4e86 DINGO\uff0c\u4e00\u4e2a\u7ec6\u7c92\u5ea6\u4e14\u591a\u6837\u5316\u7684\u6307\u4ee4\u9075\u5faa\u8bc4\u4f30\u6570\u636e\u96c6\uff0c\u5177\u6709\u4e24\u4e2a\u4e3b\u8981\u4f18\u52bf\uff1a\uff081\uff09DINGO \u57fa\u4e8e\u624b\u52a8\u6ce8\u91ca\u7684\u3001\u7ec6\u7c92\u5ea6\u7684\u591a\u7ea7\u5206\u7c7b\u6811\uff0c\u8be5\u5206\u7c7b\u6811\u5305\u542b\u4ece\u73b0\u5b9e\u4e16\u754c\u7528\u6237\u8bf7\u6c42\u4e2d\u63d0\u53d6\u7684 130 \u4e2a\u8282\u70b9\uff1b\uff082\uff09DINGO \u5305\u542b\u7531\u5927\u6a21\u578b\u548c\u4eba\u7c7b\u4e13\u5bb6\u751f\u6210\u7684\u591a\u6837\u5316\u6307\u4ee4\u3002<\/p>\n

\u00a0\u00a0<\/a><\/p>\n

\u901a\u8fc7\u5e7f\u6cdb\u7684\u5b9e\u9a8c\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u8bc1\u660e\u4e86 DINGO \u4e0d\u4ec5\u53ef\u4ee5\u4e3a\u5927\u6a21\u578b\u63d0\u4f9b\u66f4\u5177\u6311\u6218\u6027\u548c\u5168\u9762\u6027\u7684\u8bc4\u4f30\uff0c\u8fd8\u53ef\u4ee5\u63d0\u4f9b\u4efb\u52a1\u7ea7\u7684\u7ec6\u7c92\u5ea6\u65b9\u5411\u4ee5\u8fdb\u4e00\u6b65\u6539\u8fdb\u5927\u6a21\u578b\u3002\u6570\u636e\u96c6\u5df2\u7ecf\u5f00\u6e90\u3002<\/p>\n

<\/a><\/p>\n

<\/p>\n

Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning<\/p>\n

AAAI-2024<\/p>\n

\u94fe\u63a5\uff1ahttps:\/\/ojs.aaai.org\/index.php\/AAAI\/article\/view\/30087<\/a><\/p>\n

\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5728\u5404\u79cd\u4efb\u52a1\u4e2d\u53d6\u5f97\u4e86\u5de8\u5927\u6210\u529f\uff0c\u4f46\u5b83\u4eec\u5728\u751f\u6210\u5e7b\u89c9\u65b9\u9762\u4ecd\u5b58\u5728\u95ee\u9898\u3002\u817e\u8baf\u6df7\u5143\u56e2\u961f\u4ecb\u7ecd\u4e86\u4e00\u79cd\u540d\u4e3a Truth Forest \u7684\u65b9\u6cd5\uff0c\u901a\u8fc7\u4f7f\u7528\u591a\u7ef4\u6b63\u4ea4\u63a2\u9488\u63ed\u793a\u9690\u85cf\u7684\u771f\u5b9e\u8868\u793a\u6765\u589e\u5f3a LLMs \u7684\u771f\u5b9e\u6027\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5b83\u901a\u8fc7\u5728\u63a2\u9488\u4e2d\u52a0\u5165\u6b63\u4ea4\u7ea6\u675f\uff0c\u521b\u5efa\u591a\u4e2a\u6b63\u4ea4\u57fa\u6765\u5efa\u6a21\u771f\u5b9e\u4fe1\u606f\u3002\u6b64\u5916\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u5f15\u5165\u4e86 Random Peek\uff0c\u5b83\u8003\u8651\u4e86\u5e8f\u5217\u4e2d\u66f4\u5e7f\u6cdb\u7684\u4f4d\u7f6e\u4fe1\u606f\uff0c\u51cf\u5c11\u4e86\u5728 LLMs \u4e2d\u8fa8\u522b\u548c\u751f\u6210\u771f\u5b9e\u7279\u5f81\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002<\/p>\n

<\/a>\u00a0\u901a\u8fc7\u91c7\u7528\u8fd9\u79cd\u65b9\u6cd5\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u5c06 Llama-2-7B \u5728 TruthfulQA \u4e0a\u7684\u771f\u5b9e\u6027\u4ece 40.8% \u63d0\u9ad8\u5230\u4e86 74.5%\u3002\u540c\u6837\uff0c\u5728\u5fae\u8c03\u6a21\u578b\u4e2d\uff0c\u817e\u8baf\u6df7\u5143\u56e2\u961f\u4e5f\u89c2\u5bdf\u5230\u4e86\u5e7b\u89c9\u6307\u6807\u4e0a\u7684\u663e\u8457\u6539\u8fdb\u3002<\/p>\n

<\/a>\u00a0\u817e\u8baf\u6df7\u5143\u56e2\u961f\u8fd8\u4f7f\u7528\u63a2\u9488\u5bf9\u771f\u5b9e\u7279\u5f81\u8fdb\u884c\u4e86\u6df1\u5165\u5206\u6790\u3002\u817e\u8baf\u6df7\u5143\u56e2\u961f\u7684\u53ef\u89c6\u5316\u7ed3\u679c\u663e\u793a\uff0c\u6b63\u4ea4\u63a2\u9488\u6355\u6349\u5230\u4e86\u4e92\u8865\u7684\u4e0e\u771f\u5b9e\u4fe1\u606f\u76f8\u5173\u7684\u7279\u5f81\uff0c\u5f62\u6210\u4e86\u5b9a\u4e49\u660e\u786e\u7684\u7c07\uff0c\u63ed\u793a\u4e86\u6570\u636e\u96c6\u7684\u5185\u5728\u7ed3\u6784\u3002<\/p>\n

<\/a><\/p>\n

\u5927\u6a21\u578b\u53ca\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u5e7b\u89c9\u95ee\u9898\u662f\u5927\u6a21\u578b\u771f\u5b9e\u5e94\u7528\u573a\u666f\u4e0b\u7684\u4e00\u9879\u6838\u5fc3\u6311\u6218\u3002\u6df7\u5143\u56e2\u961f\u5c06\u6301\u7eed\u5728\u5927\u6a21\u578b\u5e7b\u89c9\u7b49\u5173\u952e\u4efb\u52a1\u4e0a\u8fdb\u884c\u6df1\u5165\u63a2\u7d22\uff0c\u65e8\u5728\u4e0d\u65ad\u63d0\u5347\u5927\u6a21\u578b\u5728\u771f\u5b9e\u573a\u666f\u4e0b\u7684\u53ef\u9760\u6027\u548c\u53ef\u7528\u6027\u3002<\/p>\n

\u6781\u81f4\u6027\u4ef7\u6bd4\uff0c\u6df7\u5143 Turbo \u6781\u5927\u964d\u4f4e\u5546\u7528\u95e8\u69db<\/h4>\n

\u53e6\u5916\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u8fd1\u671f\u817e\u8baf\u6df7\u5143\u4e5f\u540c\u6b65\u63a8\u51fa\u4e86\u4e0e\u5f00\u6e90\u6df7\u5143 Large \u540c\u5b97\u540c\u6e90\u7684\u5546\u7528\u65d7\u8230\u7248\u672c \u2014\u2014 \u6df7\u5143 Turbo\uff0c\u4f5c\u4e3a\u56fd\u5185\u7387\u5148\u91c7\u7528 MoE \u7ed3\u6784\u5927\u6a21\u578b\u7684\u516c\u53f8\uff0c\u817e\u8baf\u7ee7\u7eed\u5728\u8fd9\u4e00\u6280\u672f\u8def\u7ebf\u4e0a\u8fdb\u884c\u6280\u672f\u521b\u65b0\u3002<\/p>\n

<\/a><\/p>\n

\u516c\u5f00\u4fe1\u606f\u663e\u793a\uff0c\u5f53\u524d\u6df7\u5143 Turbo \u6a21\u578b\u5728\u4e1a\u754c\u516c\u8ba4\u7684 benchmark \u6307\u6807\u4e0a\u5904\u4e8e\u56fd\u5185\u884c\u4e1a\u9886\u5148\u5730\u4f4d\uff0c\u4e0e\u56fd\u5916\u5934\u90e8\u6a21\u578b\u5982 GPT-4o \u7b49\u76f8\u6bd4\u4e5f\u5904\u4e8e\u7b2c\u4e00\u68af\u961f\u3002\u53e6\u5916\uff0c\u5728\u521a\u521a\u53d1\u5e03\u7684\u56fd\u5185\u7b2c\u4e09\u65b9\u6743\u5a01\u8bc4\u6d4b\u673a\u6784\u8bc4\u6d4b\u4e2d\uff0c\u6df7\u5143 Turbo \u6a21\u578b\u4f4d\u5217\u56fd\u5185\u7b2c\u4e00\u3002<\/p>\n

\u66f4\u591a\u5927\u6a21\u578b\u7814\u7a76\u5e72\u8d27\u8bf7\u6301\u7eed\u5173\u6ce8\u6df7\u5143\u56e2\u961f\u7684\u6301\u7eed\u53d1\u5e03\u3002<\/p>\n

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587\u66dd\u5149\uff1aScaling law\u3001MoE\u3001\u5408\u6210\u6570\u636e\u4ee5\u53ca\u66f4\u591a<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587 […]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[73,68,65,67,66],"class_list":["post-36792","post","type-post","status-publish","format-standard","hentry","category-news","tag-73","tag-68","tag-65","tag-67","tag-66"],"yoast_head":"\n\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587\u66dd\u5149\uff1aScaling law\u3001MoE\u3001\u5408\u6210\u6570\u636e\u4ee5\u53ca\u66f4\u591a - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=36792\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=36792\",\"url\":\"https:\/\/17aitech.com\/?p=36792\",\"name\":\"\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587\u66dd\u5149\uff1aScaling law\u3001MoE\u3001\u5408\u6210\u6570\u636e\u4ee5\u53ca\u66f4\u591a - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=36792#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=36792#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6b158414ad099d9ea7e3e85fed69bc93.png\",\"datePublished\":\"2025-01-18T10:00:11+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=36792#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=36792\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=36792#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6b158414ad099d9ea7e3e85fed69bc93.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6b158414ad099d9ea7e3e85fed69bc93.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=36792#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587\u66dd\u5149\uff1aScaling law\u3001MoE\u3001\u5408\u6210\u6570\u636e\u4ee5\u53ca\u66f4\u591a\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587\u66dd\u5149\uff1aScaling law\u3001MoE\u3001\u5408\u6210\u6570\u636e\u4ee5\u53ca\u66f4\u591a - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=36792","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=36792","url":"https:\/\/17aitech.com\/?p=36792","name":"\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587\u66dd\u5149\uff1aScaling law\u3001MoE\u3001\u5408\u6210\u6570\u636e\u4ee5\u53ca\u66f4\u591a - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=36792#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=36792#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6b158414ad099d9ea7e3e85fed69bc93.png","datePublished":"2025-01-18T10:00:11+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=36792#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=36792"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=36792#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6b158414ad099d9ea7e3e85fed69bc93.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6b158414ad099d9ea7e3e85fed69bc93.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=36792#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u817e\u8baf\u6df7\u5143\u5927\u6a21\u578b\u6838\u5fc3\u8bba\u6587\u66dd\u5149\uff1aScaling law\u3001MoE\u3001\u5408\u6210\u6570\u636e\u4ee5\u53ca\u66f4\u591a"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/36792","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=36792"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/36792\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=36792"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=36792"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=36792"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}