{"id":7978,"date":"2026-09-15T15:00:16","date_gmt":"2026-09-15T07:00:16","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/09\/15\/meta%e6%96%b0%e7%a0%94%e7%a9%b6%ef%bc%9a%e5%ad%97%e8%8a%82%e6%a8%a1%e5%9e%8b%e8%92%b8%e9%a6%8f%e5%90%8e%ef%bc%8c%e5%a4%a9%e8%8a%b1%e6%9d%bf%e7%a0%b4%e4%ba%86\/"},"modified":"2026-09-15T15:00:16","modified_gmt":"2026-09-15T07:00:16","slug":"meta%e6%96%b0%e7%a0%94%e7%a9%b6%ef%bc%9a%e5%ad%97%e8%8a%82%e6%a8%a1%e5%9e%8b%e8%92%b8%e9%a6%8f%e5%90%8e%ef%bc%8c%e5%a4%a9%e8%8a%b1%e6%9d%bf%e7%a0%b4%e4%ba%86","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/09\/15\/meta%e6%96%b0%e7%a0%94%e7%a9%b6%ef%bc%9a%e5%ad%97%e8%8a%82%e6%a8%a1%e5%9e%8b%e8%92%b8%e9%a6%8f%e5%90%8e%ef%bc%8c%e5%a4%a9%e8%8a%b1%e6%9d%bf%e7%a0%b4%e4%ba%86\/","title":{"rendered":"Meta\u65b0\u7814\u7a76\uff1a\u5b57\u8282\u6a21\u578b\u84b8\u998f\u540e\uff0c\u5929\u82b1\u677f\u7834\u4e86"},"content":{"rendered":"
\"Meta\u65b0\u7814\u7a76\uff1a\u5b57\u8282\u6a21\u578b\u84b8\u998f\u540e\uff0c\u5929\u82b1\u677f\u7834\u4e86\"\/<\/div><\/figure>\n
\"Meta\u65b0\u7814\u7a76\uff1a\u5b57\u8282\u6a21\u578b\u84b8\u998f\u540e\uff0c\u5929\u82b1\u677f\u7834\u4e86\"\/<\/div><\/figure>\n
\"Meta\u65b0\u7814\u7a76\uff1a\u5b57\u8282\u6a21\u578b\u84b8\u998f\u540e\uff0c\u5929\u82b1\u677f\u7834\u4e86\"\/<\/div><\/figure>\n
\"Meta\u65b0\u7814\u7a76\uff1a\u5b57\u8282\u6a21\u578b\u84b8\u998f\u540e\uff0c\u5929\u82b1\u677f\u7834\u4e86\"\/<\/div><\/figure>\n

henry \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n

Token\u8bcd\u5143\u66ff\u4ee3\u4e86Byte\u5b57\u8282\uff0c\u4f46\u4e0d\u610f\u5473\u7740\u5927\u6a21\u578b\u90fd\u5f97\u662fToken\u3002<\/p>\n

\u5b57\u8282\u6a21\u578b\uff0c\u7279\u522b\u662f\u84b8\u998f\u540e\uff0c\u6709\u70b9\u513f\u65b0\u8bf4\u6cd5\u3002<\/p>\n

\u6700\u8fd1\uff0c\u6765\u81eaMeta FAIR\u548c\u534e\u76db\u987f\u5927\u5b66\u7684\u8bba\u6587\u300a\u7a81\u7834Token\u5929\u82b1\u677f\uff1a\u84b8\u998f\u51fa\u66f4\u5c0f\u3001\u66f4\u5f3a\u7684\u5b57\u8282\u6a21\u578b\u300b\uff0c\u63d0\u51fa\u4e86\u4e00\u4e2a\u5f88\u6709\u8da3\u7684\u60f3\u6cd5\u2014\u2014<\/p>\n

\u84b8\u998f\u7684\u65f6\u5019\uff0c\u80fd\u4e0d\u80fd\u628a\u5b66\u4e60\u5355\u4f4d\u4ece\u8bcd\u5143\uff08Token\uff09\u6362\u6210\u5b57\u8282\uff0c\u8ba9\u5b66\u751f\u6a21\u578b\u76f4\u63a5\u4ece\u5b57\u8282\uff08byte\uff09\u7ea7\u522b\u7684\u6982\u7387\u5206\u5e03\u5b66\u8d77\uff1f<\/p>\n

\u7ed3\u679c\u8fd8\u771f\u53ef\u4ee5\uff5e<\/p>\n

\u56e2\u961f\u7ed9\u6bcf\u4e2aToken\u52a0\u4e0a\u7ed3\u675f\u6807\u8bb0\uff0c\u628a\u6559\u5e08\u7684Token\u6982\u7387\u5206\u5e03\u5b8c\u6574\u8f6c\u6362\u5230\u5b57\u8282\u7ea7\u522b\u3002\u5728\u4ee5Llama 3-8B\u4e3a\u6559\u5e08\u7684\u84b8\u998f\u5b9e\u9a8c\u4e2d\uff0c\u56e2\u961f\u6839\u636e\u7f29\u653e\u5b9a\u5f8b\u9884\u6d4b\uff1a<\/p>\n

\u968f\u7740\u8bad\u7ec3\u8ba1\u7b97\u91cf\u589e\u52a0\uff0c\u5b57\u8282\u7ea7\u84b8\u998f\u5c06\u53cd\u8d85Token\u7ea7\u84b8\u998f\uff0c\u4e0b\u6e38\u5e73\u5747\u51c6\u786e\u7387\u4e0a\u9650\u9ad8\u51fa4\u4e2a\u767e\u5206\u70b9\u3002<\/p>\n

\u4f20\u7edf\u84b8\u998f\u8ba9\u5b66\u751f\u5b66\u4e60\u6559\u5e08\u5728\u5e9e\u5927Token\u8bcd\u8868\u4e0a\u7684\u6982\u7387\u5206\u5e03\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5b57\u8282\u53ea\u6709256\u79cd\u57fa\u7840\u53d6\u503c\uff0c\u5355\u4e2a\u4f4d\u7f6e\u7684\u9884\u6d4b\u7a7a\u95f4\u66f4\u5c0f\uff0c\u57fa\u7840\u53d6\u503c\u4e5f\u4e0d\u968f\u5206\u8bcd\u5668\u6539\u53d8\u3002<\/p>\n

\u6362\u53e5\u8bdd\u8bf4\uff0c\u5b66\u4e60\u7684\u57fa\u672c\u5355\u4f4d\u53cd\u800c\u66f4\u5c0f\u4e86\uff0c\u6a21\u578b\u6700\u7ec8\u80fd\u5b66\u5230\u7684\u80fd\u529b\u4e0a\u9650\u5374\u66f4\u9ad8\u4e86\u3002<\/p>\n

\u8fd9\u662f\u600e\u4e48\u505a\u5230\u7684\uff1f<\/p>\n

\u5176\u5b9e\u5230\u4eca\u5929\uff0c\u84b8\u998f\u5df2\u7ecf\u4e0d\u662f\u4ec0\u4e48\u964c\u751f\u6280\u672f\u4e86\u3002\u6211\u4eec\u90fd\u77e5\u9053\uff0c\u5927\u7684\u6a21\u578b\u80fd\u529b\u66f4\u5f3a\uff0c\u4f46\u628a\u5b83\u90e8\u7f72\u5230\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u663e\u5b58\u3001\u8ba1\u7b97\u548c\u54cd\u5e94\u901f\u5ea6\u90fd\u662f\u6210\u672c\u3002<\/p>\n

\u4e8e\u662f\uff0c\u4e00\u4e2a\u5e38\u89c1\u505a\u6cd5\uff0c\u5c31\u662f\u8ba9\u5927\u6a21\u578b\u5f53\u8001\u5e08\uff0c\u628a\u80fd\u529b\u84b8\u998f\u7ed9\u66f4\u5c0f\u7684\u5b66\u751f\u6a21\u578b\u3002<\/p>\n

\u548c\u666e\u901a\u76d1\u7763\u8bad\u7ec3\u53ea\u544a\u8bc9\u6a21\u578b\u201c\u4e0b\u4e00\u4e2a\u6b63\u786eToken\u662f\u4ec0\u4e48\u201d\u4e0d\u540c\uff0c\u84b8\u998f\u8fd8\u4f1a\u8ba9\u5b66\u751f\u5b66\u4e60\u6559\u5e08\u5bf9\u5404\u4e2a\u5019\u9009Token\u7684\u6982\u7387\u5224\u65ad\u3002<\/p>\n

\u4f46\u95ee\u9898\u4e5f\u51fa\u5728\u8fd9\u91cc\u3002Token\u7684\u5019\u9009\u7a7a\u95f4\u5b9e\u5728\u592a\u5927\u3002\u4ee5Llama 3-8B\u4e3a\u4f8b\uff0c\u5b83\u7684\u8bcd\u8868\u5305\u542b128256\u4e2aToken\uff0c\u8fd9\u610f\u5473\u7740\u6bcf\u4e00\u4e2a\u9884\u6d4b\u4f4d\u7f6e\uff0c\u90fd\u5bf9\u5e94\u5341\u51e0\u4e07\u4e2a\u5019\u9009\u6982\u7387\u3002<\/p>\n

\u5982\u679c\u505a\u79bb\u7ebf\u84b8\u998f\uff0c\u628a\u5b8c\u6574\u6982\u7387\u5206\u5e03\u5168\u90e8\u4fdd\u5b58\u4e0b\u6765\uff0c\u5b58\u50a8\u6210\u672c\u4f1a\u975e\u5e38\u9ad8\u3002\u56e0\u6b64\u5b9e\u9645\u64cd\u4f5c\u4e2d\uff0c\u901a\u5e38\u53ea\u80fd\u4fdd\u7559\u6982\u7387\u6700\u9ad8\u7684\u4e00\u90e8\u5206\uff0c\u4e5f\u5c31\u662ftop-k\u622a\u65ad\u3002<\/p>\n

\u800c\u5b57\u8282\u6a21\u578b\uff0c\u521a\u597d\u628a\u8fd9\u4e2a\u5019\u9009\u7a7a\u95f4\u4e00\u4e0b\u538b\u5c0f\u4e86\u3002<\/p>\n

\u4e00\u4e2a\u5b57\u8282\u75318\u4e2a\u6bd4\u7279\u7ec4\u6210\uff0c\u5171\u6709256\u79cd\u53ef\u80fd\u7684\u53d6\u503c\u3002\u5373\u4f7f\u52a0\u4e0a\u5c11\u91cf\u7279\u6b8a\u7b26\u53f7\uff0c\u6bcf\u4e2a\u4f4d\u7f6e\u9700\u8981\u4fdd\u5b58\u7684\u6982\u7387\u4e5f\u4e0d\u8fc7\u4e24\u767e\u591a\u4e2a\uff0c\u5b8c\u6574\u5206\u5e03\u81ea\u7136\u66f4\u5bb9\u6613\u4fdd\u7559\u4e0b\u6765\u3002<\/p>\n

\u4e0d\u8fc7\uff0c\u6559\u5e08\u9884\u6d4b\u7684\u662f\u6574\u4e2aToken\uff0c\u5b66\u751f\u9884\u6d4b\u7684\u662f\u5355\u4e2a\u5b57\u8282\u3002\u8981\u8ba9\u4e24\u8005\u5bf9\u4e0a\uff0c\u4e0d\u80fd\u53ea\u628a\u6587\u672c\u62c6\u5f00\uff0c\u8fd8\u5f97\u628a\u6559\u5e08\u7684\u6982\u7387\u5206\u5e03\u4e00\u8d77\u8f6c\u6362\u8fc7\u53bb\u3002<\/p>\n

\u8fd9\u6b63\u662f\u8bba\u6587\u9996\u5148\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/p>\n

\u5177\u4f53\u7684\uff0c\u56e2\u961f\u63d0\u51fa\u4e86\u4e24\u79cd\u65b9\u6848\uff0c\u5206\u522b\u53eb\u4f5cMarginalize-It\u548cEnd-Of-Token\u3002<\/p>\n

\u4e24\u8005\u7684\u57fa\u672c\u601d\u8def\u57fa\u672c\u76f8\u540c\uff1a\u628a\u6559\u5e08\u5bf9\u4e0d\u540cToken\u7684\u6982\u7387\uff0c\u9010\u6b65\u5206\u89e3\u5230\u5bf9\u5e94\u7684\u5b57\u8282\u4f4d\u7f6e\u4e0a\u3002<\/p>\n

\u4ee5\u8bba\u6587\u91cc\u7684Tiramisu\u4e3a\u4f8b\uff0c\u5b83\u53ef\u80fd\u88abtokenizer\u5207\u6210 T\u3001iram\u3001isu\u51e0\u4e2aToken\u3002<\/p>\n

\u5047\u8bbe\u6559\u5e08\u5728\u9884\u6d4b\u4e0b\u4e00\u4e2a Token \u65f6\uff0c\u5019\u9009\u91cc\u540c\u65f6\u6709 isu\u3001isk \u548c is\u3002<\/p>\n

\u8fd9\u4e09\u4e2a\u5019\u9009\u90fd\u4ee5\u5b57\u8282 i \u5f00\u5934\uff0c\u90a3\u4e48\u9884\u6d4b\u7b2c\u4e00\u4e2a\u5b57\u8282\u65f6\uff0c\u5c31\u53ef\u4ee5\u628a\u6240\u6709\u4ee5 i \u5f00\u5934\u5019\u9009\u7684\u6982\u7387\u52a0\u8d77\u6765\u3002<\/p>\n

\u7b49\u771f\u5b9e\u524d\u7f00\u53d8\u6210 i\uff0c\u518d\u53ea\u4fdd\u7559\u524d\u7f00\u5339\u914d\u7684\u5019\u9009\uff0c\u7ee7\u7eed\u8ba1\u7b97\u4e0b\u4e00\u4e2a\u5b57\u8282\u7684\u6982\u7387\u3002<\/p>\n

\u6cbf\u7740\u771f\u5b9e\u5b57\u8282\u5e8f\u5217\u4e0d\u65ad\u91cd\u590d\uff0c\u5c31\u80fd\u628a\u539f\u672c\u7684 Token \u5206\u5e03\uff0c\u4e00\u6b65\u6b65\u62c6\u6210\u9010\u5b57\u8282\u7684\u8bad\u7ec3\u76ee\u6807\u3002<\/p>\n

\u4e0d\u8fc7\uff0c\u9ebb\u70e6\u51fa\u73b0\u5728\u4e0d\u540cToken\u957f\u5ea6\u4e0d\u4e00\u6837\u7684\u65f6\u5019\u3002<\/p>\n

\u6bd4\u5982\u5df2\u7ecf\u8d70\u5230\u524d\u7f00 is\uff0cisu \u548c isk \u540e\u9762\u90fd\u8fd8\u6709\u5b57\u8282\uff0c\u4f46 is \u8fd9\u4e2aToken\u5df2\u7ecf\u7ed3\u675f\u4e86\u3002<\/p>\n

\u8fd9\u65f6\uff0c\u5982\u679c\u8fd8\u60f3\u77e5\u9053 is \u4e4b\u540e\u63a5\u4ec0\u4e48\uff0c\u5c31\u9700\u8981\u7ee7\u7eed\u67e5\u8be2\u6559\u5e08\u5728\u8fd9\u4e2aToken\u7ed3\u675f\u540e\u7684\u9884\u6d4b\uff1b\u800c\u4e0d\u540c\u5019\u9009\u8def\u5f84\uff0c\u53c8\u53ef\u80fd\u5bf9\u5e94\u4e0d\u540c\u7684\u540e\u7eed\u5206\u5e03\uff0c\u8ba1\u7b97\u91cf\u4f1a\u8fc5\u901f\u589e\u52a0\u3002<\/p>\n

Marginalize-It\u7684\u5904\u7406\u5f88\u76f4\u63a5\uff1a<\/p>\n

\u5df2\u7ecf\u7ed3\u675f\u7684\u5019\u9009\u5c31\u4e0d\u518d\u5f80\u540e\u8ffd\uff0c\u628a\u5269\u4f59\u5019\u9009\u7684\u6982\u7387\u91cd\u65b0\u5f52\u4e00\u5316\u3002<\/p>\n

\u8fd9\u6837\u53ea\u9700\u8981\u4e00\u6b21\u6559\u5e08\u524d\u5411\u8ba1\u7b97\uff0c\u6548\u7387\u5f88\u9ad8\uff0c\u4f46\u4ee3\u4ef7\u662f\u4f1a\u4e22\u6389\u4e00\u90e8\u5206\u5df2\u7ecf\u7ed3\u675f Token\u7684\u6982\u7387\u4fe1\u606f\uff0c\u56e0\u6b64\u672c\u8d28\u4e0a\u662f\u4e00\u79cd\u8fd1\u4f3c\u3002<\/p>\n

End-Of-Token\u5219\u6362\u4e86\u4e00\u4e2a\u601d\u8def\uff1a\u65e2\u7136Token\u4f1a\u7ed3\u675f\uff0c\u90a3\u5c31\u5e72\u8106\u7ed9\u201c\u7ed3\u675f\u201d\u672c\u8eab\u4e00\u4e2a\u660e\u786e\u7684\u9884\u6d4b\u4f4d\u7f6e\u3002<\/p>\n

\u5177\u4f53\u6765\u8bf4\uff0c\u56e2\u961f\u5728\u6bcf\u4e2aToken\u672b\u5c3e\u52a0\u5165\u4e00\u4e2a\u7279\u6b8a\u7b26\u53f7\u3002\u4e8e\u662f\u539f\u6765\u7684\u4e09\u4e2a\u5019\u9009\u5c31\u53d8\u6210\uff1aisu\u3001isk\u548cis\u3002<\/p>\n

\u8fd9\u6837\uff0c\u5f53\u5df2\u7ecf\u8d70\u5230\u524d\u7f00is\u65f6\uff0c\u5b66\u751f\u63a5\u4e0b\u6765\u65e2\u53ef\u4ee5\u9884\u6d4bu\uff0c\u4e5f\u53ef\u4ee5\u9884\u6d4bk\uff0c\u8fd8\u53ef\u4ee5\u9884\u6d4b\uff0c\u8868\u793a\u5f53\u524dToken\u5230\u8fd9\u91cc\u7ed3\u675f\u3002<\/p>\n

\u539f\u672c\u56e0\u4e3a\u957f\u5ea6\u4e0d\u540c\u800c\u201c\u60ac\u7a7a\u201d\u7684\u90a3\u90e8\u5206\u6982\u7387\uff0c\u5c31\u6709\u4e86\u660e\u786e\u53bb\u5904\u3002<\/p>\n

\u56e0\u6b64\uff0cEnd-Of-Token\u53ef\u4ee5\u5728\u4fdd\u7559Token\u8fb9\u754c\u7684\u540c\u65f6\uff0c\u66f4\u5b8c\u6574\u5730\u628a\u6559\u5e08\u5206\u5e03\u6620\u5c04\u5230\u5b57\u8282\u7a7a\u95f4\u3002<\/p>\n

\u800c\u4e14\uff0c\u4e24\u79cd\u65b9\u6cd5\u90fd\u53ea\u9700\u8981\u6559\u5e08\u505a\u4e00\u6b21\u524d\u5411\u8ba1\u7b97\uff0c\u5c31\u80fd\u5b8c\u6210\u4eceToken\u5206\u5e03\u5230\u5b57\u8282\u5206\u5e03\u7684\u8f6c\u6362\uff0c\u4e0d\u5fc5\u4e3a\u4e86\u4e0d\u540c\u5206\u8bcd\u8def\u5f84\u53cd\u590d\u8c03\u7528\u6559\u5e08\u6a21\u578b\u3002<\/p>\n

\u5230\u8fd9\u91cc\uff0c\u6559\u5e08\u6a21\u578b\u7684\u77e5\u8bc6\uff0c\u603b\u7b97\u53ef\u4ee5\u771f\u6b63\u4ea4\u5230\u5b57\u8282\u5b66\u751f\u624b\u91cc\u4e86\u3002<\/p>\n

\u63a5\u4e0b\u6765\u8981\u770b\u7684\uff0c\u5c31\u662f\u5b83\u5230\u5e95\u80fd\u5b66\u5230\u591a\u5c11\u3002<\/p>\n

\u56e2\u961f\u5206\u522b\u6d4b\u8bd5\u4e86Token\u3001\u666e\u901a\u5b57\u8282\u548c\u5e26\u7684\u5b57\u8282\u4e09\u79cd\u8868\u793a\uff0c\u6bcf\u79cd\u518d\u5206\u4e3a\u76d1\u7763\u8bad\u7ec3\u4e0e\u84b8\u998f\u8bad\u7ec3\uff0c\u4e00\u5171\u516d\u7ec4\u3002<\/p>\n

\u8fd9\u6837\u65e2\u80fd\u6bd4\u8f83\u84b8\u998f\u6709\u6ca1\u6709\u5e2e\u52a9\uff0c\u4e5f\u80fd\u89c2\u5bdf\u6539\u53d8\u5b66\u4e60\u5355\u4f4d\u540e\uff0c\u6a21\u578b\u7684\u8bad\u7ec3\u8868\u73b0\u5982\u4f55\u53d8\u5316\u3002<\/p>\n

\u6559\u5e08\u7edf\u4e00\u4f7f\u7528Llama 3-8B\u3002\u4e09\u7c7b\u5b66\u751f\u7684Transformer\u5c42\u53c2\u6570\u91cf\u76f8\u540c\uff0c\u5747\u7ea6\u4e3a12.8\u4ebf\u3002<\/p>\n

\u8ba1\u5165\u8bcd\u8868\u76f8\u5173\u53c2\u6570\u540e\uff0c\u603b\u53c2\u6570\u91cf\u5219\u6709\u6240\u4e0d\u540c\uff1aToken\u5b66\u751f\u7ea618.1\u4ebf\uff0c\u5b57\u8282\u5b66\u751f\u7ea612.8\u4ebf\u3002<\/p>\n

\u7814\u7a76\u9010\u6b65\u589e\u52a0\u8bad\u7ec3\u6570\u636e\u548c\u8ba1\u7b97\u91cf\uff0c\u5b57\u8282\u6a21\u578b\u7684\u8bad\u7ec3\u8986\u76d6\u7ea6\u4e07\u4ebf\u5b57\u8282\u89c4\u6a21\uff0c\u518d\u5728\u9009\u62e9\u9898\u95ee\u7b54\u3001\u8bed\u8a00\u751f\u6210\u548c\u673a\u5668\u7ffb\u8bd1\u4e09\u7c7b\u4efb\u52a1\u4e2d\u5b8c\u6210\u516b\u9879\u8bc4\u6d4b\u3002<\/p>\n

\u5148\u770b\u8bad\u7ec3\u8fc7\u7a0b\u3002<\/p>\n

Token\u6a21\u578b\u5728\u8ba1\u7b97\u91cf\u8f83\u5c0f\u65f6\u5b66\u5f97\u66f4\u5feb\uff0c\u4f46\u5f88\u5feb\u5f00\u59cb\u63a5\u8fd1\u5e73\u53f0\u671f\uff1b\u5b57\u8282\u6a21\u578b\u8d77\u6b65\u66f4\u6162\uff0c\u5374\u968f\u7740\u8ba1\u7b97\u91cf\u589e\u52a0\u6301\u7eed\u6539\u5584\u3002<\/p>\n

\u4e5f\u5c31\u662f\u8bf4\uff0cToken\u6a21\u578b\u8d62\u5728\u524d\u671f\u6548\u7387\uff0c\u5b57\u8282\u6a21\u578b\u5219\u8868\u73b0\u51fa\u4e86\u66f4\u597d\u7684Scaling\u6f5c\u529b\u3002<\/p>\n

\u4e0d\u8fc7\uff0c\u5982\u679c\u53ea\u770bBPB\uff0c\u4e5f\u5c31\u662f\u7edf\u4e00\u5230\u6bcf\u5b57\u8282\u4e0a\u7684\u9884\u6d4b\u635f\u5931\uff0c\u51e0\u79cd\u5b57\u8282\u6a21\u578b\u5f88\u65e9\u5c31\u5df2\u7ecf\u8d85\u8fc7Token\u6a21\u578b\uff0c\u4f46\u771f\u6b63\u62ff\u53bb\u505a\u9898\u65f6\uff0c\u6210\u7ee9\u5374\u4ecd\u7136\u66f4\u4f4e\u3002<\/p>\n

\u4e4b\u6240\u4ee5\u8fd9\u6837\uff0c\u662f\u56e0\u4e3aBPB\u8861\u91cf\u7684\u662f\u6a21\u578b\u5bf9\u6587\u672c\u9884\u6d4b\u5f97\u51c6\u4e0d\u51c6\uff0c\u800c\u4e0b\u6e38\u4efb\u52a1\u770b\u7684\u662f\u6700\u540e\u7b54\u6848\u80fd\u4e0d\u80fd\u9009\u5bf9\uff0c\u4e24\u8005\u5e76\u4e0d\u662f\u4e00\u56de\u4e8b\u3002<\/p>\n

\u6240\u4ee5\uff0c\u56e2\u961f\u6ca1\u6709\u76f4\u63a5\u62ffBPB\u5224\u65ad\u8c01\u66f4\u5f3a\uff0c\u800c\u662f\u8fdb\u4e00\u6b65\u62df\u5408\u4e86\uff1a\u8bad\u7ec3\u8ba1\u7b97\u91cf \u2192 BPB \u2192 \u4e0b\u6e38\u4efb\u52a1\u6210\u7ee9\uff0c\u518d\u7528\u8fd9\u6761\u5173\u7cfb\u9884\u6d4b\u4e0d\u540c\u65b9\u6848\u7ee7\u7eedScaling\u540e\u7684\u80fd\u529b\u4e0a\u9650\u3002<\/p>\n

\u7ed3\u679c\u8868\u660e\uff0c\u4e09\u79cd\u84b8\u998f\u65b9\u6848\u7684\u9884\u6d4b\u5e73\u5747\u51c6\u786e\u7387\u4e0a\u9650\u5206\u522b\u662f\uff1a<\/p>\n

Token\u84b8\u998f48.4%\uff0cMarginalize-It\u84b8\u998f50.5%\uff0cEnd-Of-Token\u84b8\u998f52.4%\u3002<\/p>\n

\u4e5f\u5c31\u662f\u8bf4\uff0cEnd-Of-Token\u6700\u7ec8\u6bd4\u4f20\u7edfToken\u84b8\u998f\u9ad8\u51fa4\u4e2a\u767e\u5206\u70b9\uff0c\u4e5f\u662f\u516d\u79cd\u65b9\u6848\u91cc\u9884\u6d4b\u4e0a\u9650\u6700\u9ad8\u7684\u3002<\/p>\n

\u4e0d\u8fc7\uff0c\u84b8\u998f\u7684\u4f18\u52bf\u5e76\u975e\u5728\u6240\u6709\u9884\u6d4b\u4e2d\u90fd\u6210\u7acb\u3002<\/p>\n

\u666e\u901a\u5b57\u8282\u76d1\u7763\u6a21\u578b\u7684\u6e10\u8fd1\u51c6\u786e\u7387\u9884\u6d4b\u4e3a51.2%\uff0c\u9ad8\u4e8eMarginalize-It\u84b8\u998f\u768450.5%\u3002<\/p>\n

\u8bba\u6587\u8ba4\u4e3a\uff0c\u8fd9\u91cc\u7684\u539f\u56e0\u5f88\u53ef\u80fd\u5c31\u662f\u524d\u9762\u63d0\u5230\u7684\uff1aMarginalize-It\u4e3a\u4e86\u7701\u8ba1\u7b97\uff0c\u4e22\u6389\u4e86\u4e00\u90e8\u5206Token\u7ed3\u675f\u540e\u7684\u6982\u7387\u4fe1\u606f\u3002<\/p>\n

\u800c\u80fd\u591f\u66f4\u5b8c\u6574\u4fdd\u7559\u6559\u5e08\u5206\u5e03\u7684End-Of-Token\uff0c\u5219\u540c\u65f6\u8d85\u8fc7\u4e86\u81ea\u5df1\u7684\u76d1\u7763\u7248\u672c\u548c\u5176\u4ed6\u6240\u6709\u65b9\u6848\u3002<\/p>\n

\u8fdb\u4e00\u6b65\uff0c\u6309\u7167Scaling\u66f2\u7ebf\u7ee7\u7eed\u5f80\u5916\u63a8\uff0cEnd-Of-Token\u5927\u7ea6\u57286.33\u00d710\u00b2\u00b2 FLOPs\u65f6\uff0c\u5c31\u80fd\u8ffd\u5e73Token\u84b8\u998f48.4%\u7684\u9884\u6d4b\u4e0a\u9650\u3002<\/p>\n

\u5728\u6700\u540e\uff0c\u5b9e\u9a8c\u8fd8\u8868\u660e\u5b57\u8282\u7ea7\u84b8\u998f\u80fd\u663e\u8457\u964d\u4f4e\u6570\u636e\u548c\u6559\u5e08\u5206\u5e03\u7684\u5b58\u50a8\u538b\u529b\u3002<\/p>\n

\u56e2\u961f\u8fd8\u628a\u8fd9\u4e9b\u6e10\u8fd1\u9884\u6d4b\uff0c\u4e0e\u73b0\u6709\u5f00\u653e\u6743\u91cd\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002<\/p>\n

\u6309\u7167\u8bba\u6587\u7684\u9884\u6d4b\u8ba1\u7b97\u9884\u7b97\uff0cEnd-Of-Token\u5904\u7406\u7684\u5b9e\u9645\u6587\u672c\u91cf\u7ea6\u4e3aToken\u65b9\u6848\u7684\u516d\u5206\u4e4b\u4e00\uff1b<\/p>\n

\u5728Token\u53ea\u4fdd\u5b58top-600\u3001\u5b57\u8282\u4fdd\u5b58\u5b8c\u6574\u5206\u5e03\u7684\u8bbe\u7f6e\u4e0b\uff0c\u5b58\u50a8\u91cf\u4e5f\u53ea\u6709\u7ea6\u4e94\u5206\u4e4b\u4e00\u3002<\/p>\n

\u4e0d\u8fc7\uff0c\u66f4\u5c11\u7684\u6570\u636e\uff0c\u4e5f\u4e0d\u4ee3\u8868\u66f4\u5c11\u7684\u8ba1\u7b97\u3002<\/p>\n

\u5c24\u5176End-Of-Token\u8fd8\u5728\u6bcf\u4e2aToken\u540e\u52a0\u5165\u4e86\uff0c\u5904\u7406\u76f8\u540c\u6587\u672c\u91cf\u65f6\uff0c\u8bad\u7ec3\u8ba1\u7b97\u91cf\u6bd4\u666e\u901a\u5b57\u8282\u6a21\u578b\u8fd8\u8981\u518d\u9ad8\u7ea630.94%\u3002<\/p>\n

\u4e0e\u6b64\u540c\u65f6\uff0c\u63a8\u7406\u6210\u672c\u4e5f\u662f\u540c\u6837\u7684\u95ee\u9898\uff0c\u8bba\u6587\u76ee\u524d\u8fd8\u6ca1\u6709\u5b8c\u6210\u7b49\u63a8\u7406\u6210\u672c\u4e0b\u7684\u516c\u5e73\u6bd4\u8f83\u3002<\/p>\n

\u6240\u4ee5\u603b\u7684\u6765\u8bf4\uff0c\u8fd9\u9879\u5de5\u4f5c\u5c55\u793a\u4e86\u4e00\u6761\u5c0f\u6a21\u578b\u8bad\u7ec3\u7684\u65b0\u8def\u5f84\uff1a\u628a\u6559\u5e08\u7684\u77e5\u8bc6\u8f6c\u6362\u5230\u5b57\u8282\u7ea7\u522b\uff0c\u8ba9\u5b66\u751f\u7528\u66f4\u5c0f\u7684\u9884\u6d4b\u7a7a\u95f4\u3001\u66f4\u5c11\u7684\u8bad\u7ec3\u6587\u672c\uff0c\u4e89\u53d6\u66f4\u9ad8\u7684\u80fd\u529b\u4e0a\u9650\u3002<\/p>\n

End-Of-Token\u5728\u8fd9\u7ec4\u5b9e\u9a8c\u7684\u5916\u63a8\u4e2d\u8868\u73b0\u6700\u597d\uff0c\u4f46\u8981\u5151\u73b0\u8fd9\u4e00\u6f5c\u529b\uff0c\u8fd8\u9700\u8981\u7ee7\u7eed\u6295\u5165\u8ba1\u7b97\u3002<\/p>\n

\u6700\u540e\uff0c\u8ba9\u6211\u4eec\u4ecb\u7ecd\u4e00\u4e0b\u6587\u7ae0\u7684\u4f5c\u8005\u4eec\u3002<\/p>\n

\u8bba\u6587\u4e00\u4f5cKalyani Marathe\u662f\u534e\u76db\u987f\u5927\u5b66\u535a\u58eb\u751f\uff0c\u5e08\u4eceLuke Zettlemoyer\uff0c\u7814\u7a76\u5174\u8da3\u4e3b\u8981\u96c6\u4e2d\u5728\u5927\u6a21\u578b\u8bad\u7ec3\u3001\u6a21\u578b\u67b6\u6784\u4e0eScaling\u3002<\/p>\n

\u5979\u6b64\u524d\u66fe\u5728Meta AI Research\u5b9e\u4e60\uff0c\u4e5f\u5728Amazon Robotics \u7684 Foundation Models\u56e2\u961f\u5de5\u4f5c\u8fc7\u3002<\/p>\n

Artidoro Pagnoni\u662f\u534e\u76db\u987f\u5927\u5b66\u8ba1\u7b97\u673a\u7cfb\u535a\u58eb\u751f\u3001Meta\u8bbf\u95ee\u7814\u7a76\u5458\uff0c\u7814\u7a76\u91cd\u70b9\u662f\u5927\u6a21\u578b\u7684\u8d44\u6e90\u6548\u7387\u4e0eScaling\uff0c\u6b64\u524d\u66fe\u5c31\u8bfb\u4e8eCMU\u548c\u54c8\u4f5b\u5927\u5b66\u3002<\/p>\n

\u4ed6\u4e5f\u662fByte Latent Transformer\uff08BLT\uff09\u548cQLoRA\u7684\u6838\u5fc3\u4f5c\u8005\u4e4b\u4e00\uff0c\u540e\u8005\u5c06\u5927\u6a21\u578b\u5fae\u8c03\u7684\u663e\u5b58\u9700\u6c42\u5927\u5e45\u964d\u4f4e\u3002<\/p>\n

Tomasz Limisiewicz\u662f\u534e\u76db\u987f\u5927\u5b66\u548cMeta\u7684\u535a\u58eb\u540e\u7814\u7a76\u5458\uff0c\u6b64\u524d\u5728\u67e5\u7406\u5927\u5b66\u83b7\u5f97\u535a\u58eb\u5b66\u4f4d\u3002<\/p>\n

\u4ed6\u4e3b\u8981\u7814\u7a76\u5927\u8bed\u8a00\u6a21\u578b\u5982\u4f55\u5b66\u4e60\u548c\u8fc1\u79fb\u80fd\u529b\uff0c\u5c24\u5176\u5173\u6ce8\u6570\u636e\u3001\u6a21\u578b\u5185\u90e8\u673a\u5236\u3001Tokenization\uff0c\u4ee5\u53ca\u4f4e\u8d44\u6e90\u8bed\u8a00\u548c\u6a21\u578b\u9c81\u68d2\u6027\u3002<\/p>\n

Margaret Li\u662f\u534e\u76db\u987f\u5927\u5b66\u535a\u58eb\u751f\uff0c\u5e08\u4eceLuke Zettlemoyer\u548cTim Althoff\uff0c\u540c\u65f6\u4e5f\u662fMeta FAIR\u7684\u8bbf\u95ee\u7814\u7a76\u5458\u3002<\/p>\n

\u5979\u7684\u7814\u7a76\u6d89\u53ca\u5927\u6a21\u578b\u9ad8\u6548\u8bad\u7ec3\u4e0e\u751f\u6210\u3001\u6761\u4ef6\u8ba1\u7b97\u3001\u53ef\u63a7\u6027\u548c\u5b89\u5168\uff0c\u6b64\u524d\u8fd8\u66fe\u62c5\u4efbFAIR\u7814\u7a76\u5de5\u7a0b\u5e08\uff0c\u53c2\u4e0e\u5f00\u653e\u57df\u5bf9\u8bdd\u7b49\u65b9\u5411\u7684\u7814\u7a76\u3002<\/p>\n

Mike Lewis\u662fMeta FAIR\u7814\u7a76\u79d1\u5b66\u5bb6\uff0c\u4e5f\u662fLlama 3\u9884\u8bad\u7ec3\u8d1f\u8d23\u4eba\u4e4b\u4e00\u3002<\/p>\n

\u4ed6\u6b64\u524d\u53c2\u4e0e\u4e86RoBERTa\u3001BART\u3001CICERO\u3001kNN-LM\u3001Attention Sinks\u548cTop-k Sampling\u7b49\u591a\u9879\u4ee3\u8868\u6027\u5de5\u4f5c\u3002<\/p>\n

Luke Zettlemoyer\u662f\u534e\u76db\u987f\u5927\u5b66Allen School\u6559\u6388\uff0c\u540c\u65f6\u62c5\u4efbMeta FAIR Seattle\u8d1f\u8d23\u4eba\uff0c\u957f\u671f\u7814\u7a76\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u3001\u8bed\u4e49\u89e3\u6790\u3001\u95ee\u7b54\u548c\u591a\u8bed\u8a00NLP\u3002<\/p>\n

\u4ed6\u66fe\u83b7\u5f97\u7f8e\u56fd\u603b\u7edf\u9752\u5e74\u79d1\u5b66\u5bb6\u4e0e\u5de5\u7a0b\u5e08\u5956\uff08PECASE\uff09\uff0c\u5e76\u5728\u9876\u7ea7NLP\u4f1a\u8bae\u4e0a\u83b7\u5f97\u5341\u4f59\u9879\u8bba\u6587\u5956\u3002<\/p>\n

Srinivasan Iyer\u662fMeta FAIR\u7814\u7a76\u79d1\u5b66\u5bb6\uff0c\u7814\u7a76\u65b9\u5411\u6db5\u76d6\u8bed\u8a00\u5efa\u6a21\u3001\u6587\u672c\u751f\u6210\u3001\u6458\u8981\u3001\u53ef\u89e3\u91ca\u6027\u548c\u95ee\u7b54\u3002<\/p>\n

\u4ed6\u5728\u534e\u76db\u987f\u5927\u5b66\u83b7\u5f97\u8ba1\u7b97\u673a\u535a\u58eb\u5b66\u4f4d\uff0c\u5e08\u4eceLuke Zettlemoyer\u548cAlvin Cheung\uff0c\u535a\u58eb\u9636\u6bb5\u91cd\u70b9\u7814\u7a76\u81ea\u7136\u8bed\u8a00\u4e0e\u901a\u7528\u6e90\u4ee3\u7801\u7684\u8054\u5408\u5efa\u6a21\u3002<\/p>\n

\u53c2\u8003\u94fe\u63a5<\/p>\n

[1]https:\/\/arxiv.org\/pdf\/2609.12303<\/p>\n

\n

\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n

\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n

\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n

\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"

henry \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n

Token\u8bcd\u5143\u66ff\u4ee3\u4e86Byte\u5b57\u8282\uff0c\u4f46\u4e0d\u610f\u5473\u7740\u5927\u6a21\u578b\u90fd\u5f97\u662fToken\u3002<\/p>\n

\u5b57\u8282\u6a21\u578b\uff0c\u7279\u522b\u662f\u84b8\u998f\u540e\uff0c\u6709\u70b9\u513f\u65b0\u8bf4\u6cd5\u3002<\/p>\n

\u6700\u8fd1\uff0c\u6765\u81eaMeta FAIR\u548c\u534e\u76db\u987f\u5927\u5b66\u7684\u8bba\u6587\u300a\u7a81\u7834Token\u5929\u82b1\u677f\uff1a\u84b8\u998f\u51fa\u66f4\u5c0f\u3001\u66f4\u5f3a\u7684\u5b57\u8282\u6a21\u578b\u300b\uff0c\u63d0\u51fa\u4e86\u4e00\u4e2a\u5f88\u6709\u8da3\u7684\u60f3\u6cd5\u2014\u2014<\/p>\n

\u84b8\u998f\u7684\u65f6\u5019\uff0c\u80fd\u4e0d\u80fd\u628a\u5b66\u4e60\u5355\u4f4d\u4ece\u8bcd\u5143\uff08Token\uff09\u6362\u6210\u5b57\u8282\uff0c\u8ba9\u5b66\u751f\u6a21\u578b\u76f4\u63a5\u4ece\u5b57\u8282\uff08byte\uff09\u7ea7\u522b\u7684\u6982\u7387\u5206\u5e03\u5b66\u8d77\uff1f<\/p>\n

\u7ed3\u679c\u8fd8\u771f\u53ef\u4ee5\uff5e<\/p>\n","protected":false},"author":1,"featured_media":7974,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[121,519,1211],"class_list":["post-7978","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx","tag-meta","tag-token","tag-1211"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/7978","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=7978"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/7978\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/7974"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=7978"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=7978"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=7978"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}