{"id":122,"date":"2026-04-25T17:36:59","date_gmt":"2026-04-25T09:36:59","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/?p=122"},"modified":"2026-04-23T17:38:41","modified_gmt":"2026-04-23T09:38:41","slug":"tokenizer-%e8%a7%a3%e6%9e%90","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/25\/tokenizer-%e8%a7%a3%e6%9e%90\/","title":{"rendered":"Tokenizer \u89e3\u6790"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\">\u5199\u5728\u524d\u9762<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u4e0a\u4e00\u7bc7\u6587\u7ae0\u4e2d\uff0c\u5b8c\u6574\u5730\u8d70\u4e86\u4e00\u904d LLM \u7684\u67b6\u6784\u5168\u666f\u56fe\u3002\u90a3\u7bc7\u6587\u7ae0\u63d0\u5230\uff1a<strong>\u539f\u59cb\u6587\u672c\u8fdb\u5165\u6a21\u578b\u7684\u7b2c\u4e00\u6b65\uff0c\u5c31\u662f\u7ecf\u8fc7 Tokenizer \u7684\u5904\u7406<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u8bf4\u5b9e\u8bdd\uff0c\u5f53\u65f6\u53ea\u7528\u51e0\u6bb5\u8bdd\u5e26\u8fc7 Tokenizer\uff0c\u5b9e\u5728\u6709\u70b9\u201c\u4e8f\u5f85\u201d\u8fd9\u4f4d\u9ed8\u9ed8\u65e0\u95fb\u7684\u529f\u81e3\u3002Tokenizer \u867d\u7136\u770b\u8d77\u6765\u53ea\u662f\u505a\u201c\u5207\u8bcd\u201d\u8fd9\u79cd\u7b80\u5355\u7684\u5de5\u4f5c\uff0c\u4f46\u5b83\u5bf9\u6574\u4e2a\u6a21\u578b\u7684\u6027\u80fd\u3001\u6548\u7387\u751a\u81f3\u516c\u5e73\u6027\u90fd\u6709\u7740\u6df1\u8fdc\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4eca\u5929\u8fd9\u7bc7\u6587\u7ae0\uff0c\u6211\u4eec\u5c31\u6765\u4e13\u95e8\u804a\u4e00\u804a Tokenizer\u2014\u2014\u8fd9\u4f4d\u6587\u672c\u7684\u201c\u7ffb\u8bd1\u5b98\u201d\u3001\u5927\u6a21\u578b\u7684\u201c\u5b88\u95e8\u4eba\u201d\u3002\u6211\u4f1a\u4ece\u5b83\u7684\u57fa\u672c\u539f\u7406\u8bb2\u8d77\uff0c\u6df1\u5165\u5230\u4e09\u79cd\u4e3b\u6d41\u5206\u8bcd\u7b97\u6cd5\u7684\u5bf9\u6bd4\uff0c\u518d\u5230\u5b57\u8282\u7ea7 BPE \u7684\u5de5\u7a0b\u5b9e\u8df5\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u4e3a\u4ec0\u4e48\u9700\u8981 Tokenizer\uff1f\u2014\u2014\u4e00\u4e2a\u770b\u4f3c\u7b80\u5355\u5374\u6697\u85cf\u7384\u673a\u7684\u95ee\u9898<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1.1 Tokenizer \u7684\u201c\u672c\u804c\u5de5\u4f5c\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5927\u6a21\u578b\u672c\u8d28\u4e0a\u662f\u4e00\u5806\u77e9\u9635\u8fd0\u7b97\u7684\u96c6\u5408\uff0c\u5b83\u4e0d\u61c2\u6c49\u5b57\u3001\u5b57\u6bcd\u6216\u6807\u70b9\u7b26\u53f7\uff0c\u53ea\u80fd\u5904\u7406\u6570\u5b57\u3002Tokenizer \u7684\u5de5\u4f5c\uff0c\u5c31\u662f\u628a\u4eba\u7c7b\u5199\u7684\u6587\u672c\u8f6c\u6362\u6210\u6a21\u578b\u80fd\u201c\u5403\u201d\u7684\u6570\u5b57\u5e8f\u5217\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u6280\u672f\u4e0a\u8bb2\uff0cTokenization \u662f\u8bad\u7ec3 LLM \u7684\u7b2c\u4e00\u6b65\u3002LLM \u5206\u6790 Token \u4e4b\u95f4\u7684\u8bed\u4e49\u5173\u7cfb\u2014\u2014\u6bd4\u5982\u5b83\u4eec\u4e00\u8d77\u51fa\u73b0\u7684\u9891\u7387\uff0c\u6216\u8005\u662f\u5426\u51fa\u73b0\u5728\u76f8\u4f3c\u7684\u4e0a\u4e0b\u6587\u73af\u5883\u4e2d\u3002\u6709\u4e86\u8fd9\u4e9b\u6a21\u5f0f\u548c\u5173\u7cfb\uff0c\u6a21\u578b\u624d\u80fd\u6839\u636e\u8f93\u5165\u7684 Token \u5e8f\u5217\uff0c\u9884\u6d4b\u5e76\u751f\u6210\u4e0b\u4e00\u4e2a Token\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3e\u4e2a\u4f8b\u5b50\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\u8f93\u5165\uff1a\"\u6211\u7231\u5927\u8bed\u8a00\u6a21\u578b\"\n\n\u2193 Tokenizer \u5904\u7406 \u2193\n\nToken \u5e8f\u5217\uff1a&#91;\"\u6211\", \"\u7231\", \"\u5927\", \"\u8bed\u8a00\", \"\u6a21\u578b\"]\n\n\u2193 \u6620\u5c04\u5230\u8bcd\u8868 \u2193\n\nToken ID\uff1a&#91;231, 452, 789, 1023, 88]<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u95ee\u9898\u6765\u4e86\uff1a<strong>\u600e\u4e48\u201c\u5207\u201d\u624d\u662f\u6700\u597d\u7684\uff1f<\/strong> \u8fd9\u5c31\u662f Tokenizer \u8bbe\u8ba1\u7684\u6838\u5fc3\u95ee\u9898\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.2 \u4e09\u79cd\u7c92\u5ea6\u7684\u53d6\u820d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 NLP \u7684\u53d1\u5c55\u53f2\u4e0a\uff0c\u7814\u7a76\u8005\u4eec\u5c1d\u8bd5\u8fc7\u4e09\u79cd\u4e0d\u540c\u7c92\u5ea6\u7684\u5206\u8bcd\u65b9\u6848\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u5206\u8bcd\u7c92\u5ea6<\/th><th>\u505a\u6cd5<\/th><th>\u4f18\u70b9<\/th><th>\u7f3a\u70b9<\/th><\/tr><\/thead><tbody><tr><td><strong>\u5355\u8bcd\u7ea7<\/strong><\/td><td>\u6309\u7a7a\u683c\u548c\u6807\u70b9\u62c6\u5206<\/td><td>Token \u6570\u91cf\u5c11\uff0c\u8ba1\u7b97\u91cf\u5c0f<\/td><td>\u8bcd\u8868\u5de8\u5927\uff08\u51e0\u5341\u4e07\u751a\u81f3\u4e0a\u767e\u4e07\uff09\uff0cOOV \u95ee\u9898\u4e25\u91cd<\/td><\/tr><tr><td><strong>\u5b57\u7b26\u7ea7<\/strong><\/td><td>\u6309\u5355\u4e2a\u5b57\u7b26\u62c6\u5206<\/td><td>\u8bcd\u8868\u6781\u5c0f\uff08\u51e0\u5341\u5230\u51e0\u767e\u4e2a\uff09\uff0c\u5b8c\u5168\u8986\u76d6\u6240\u6709\u8f93\u5165<\/td><td>Token \u5e8f\u5217\u8d85\u957f\uff0c\u8ba1\u7b97\u91cf\u5927\uff0c\u8bed\u4e49\u4fe1\u606f\u88ab\u7a00\u91ca<\/td><\/tr><tr><td><strong>\u5b50\u8bcd\u7ea7<\/strong><\/td><td>\u4ecb\u4e8e\u5355\u8bcd\u548c\u5b57\u7b26\u4e4b\u95f4<\/td><td>\u5e73\u8861\u8bcd\u8868\u5927\u5c0f\u548c\u5e8f\u5217\u957f\u5ea6\uff0c\u80fd\u5904\u7406 OOV<\/td><td>\u7b97\u6cd5\u76f8\u5bf9\u590d\u6742\uff0c\u8bed\u8a00\u9002\u914d\u6709\u5dee\u5f02<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5b50\u8bcd\u7ea7\u5206\u8bcd<\/strong>\u662f\u76ee\u524d\u5927\u8bed\u8a00\u6a21\u578b\u7684\u4e3b\u6d41\u9009\u62e9\u3002\u5b83\u65e2\u4e0d\u50cf\u5355\u8bcd\u7ea7\u90a3\u6837\u9047\u5230\u751f\u50fb\u8bcd\u5c31\u675f\u624b\u65e0\u7b56\uff0c\u4e5f\u4e0d\u50cf\u5b57\u7b26\u7ea7\u90a3\u6837\u628a\u53e5\u5b50\u62c6\u5f97\u7a00\u788e\u3002\u901a\u8fc7\u5c06\u5e38\u89c1\u8bcd\u4fdd\u7559\u4e3a\u5b8c\u6574 Token\uff0c\u5c06\u7f55\u89c1\u8bcd\u62c6\u5206\u4e3a\u6709\u610f\u4e49\u7684\u5b50\u8bcd\u7247\u6bb5\uff0cTokenizer \u5728\u6548\u7387\u548c\u8bed\u4e49\u5b8c\u6574\u6027\u4e4b\u95f4\u627e\u5230\u4e86\u4e00\u4e2a\u201c\u751c\u871c\u70b9\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.3 \u4e00\u4e2a\u7ecf\u5178\u7684\u4f8b\u5b50<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6765\u770b BERT \u548c GPT \u4e24\u79cd\u5206\u8bcd\u5668\u7684\u5b9e\u9645\u8868\u73b0\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\u8f93\u5165\uff1a\"transformer architecture\"\n\nGPT\uff08BPE\uff09\u5206\u8bcd\uff1a&#91;\"transform\", \"er\", \"architecture\"]\nBERT\uff08WordPiece\uff09\u5206\u8bcd\uff1a&#91;\"transform\", \"##er\", \"arch\", \"##itecture\"]<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ef\u4ee5\u770b\u5230\uff0c\u4e24\u8005\u90fd\u628a\u201ctransformer\u201d\u62c6\u5206\u6210\u4e86\u201ctransform\u201d\u548c\u201cer\u201d\uff0c\u628a\u201carchitecture\u201d\u62c6\u6210\u4e86\u66f4\u5c0f\u7684\u7247\u6bb5\u3002\u8fd9\u79cd\u62c6\u5206\u65b9\u5f0f\u8ba9\u6a21\u578b\u80fd\u591f\u7406\u89e3\uff1a\u5373\u4f7f\u6ca1\u89c1\u8fc7\u201ctransformer\u201d\u8fd9\u4e2a\u8bcd\uff0c\u4f46\u77e5\u9053\u201ctransform\u201d\u548c\u201cer\u201d\u5404\u81ea\u7684\u610f\u601d\uff0c\u4e5f\u80fd\u5927\u81f4\u63a8\u65ad\u51fa\u201ctransformer\u201d\u7684\u8bed\u4e49\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u4e09\u5927\u4e3b\u6d41\u7b97\u6cd5\uff1aBPE\u3001WordPiece\u3001SentencePiece<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8bf4 Tokenizer \u662f\u6574\u4e2a\u6a21\u578b\u7684\u7b2c\u4e00\u9053\u5de5\u5e8f\uff0c\u90a3\u4e48\u5206\u8bcd\u7b97\u6cd5\u5c31\u662f\u8fd9\u9053\u5de5\u5e8f\u7684\u6838\u5fc3\u5f15\u64ce\u3002\u76ee\u524d\u4e3b\u6d41\u7684\u4e09\u79cd\u5b50\u8bcd\u5206\u8bcd\u7b97\u6cd5\uff0c\u5404\u6709\u5404\u7684\u8bbe\u8ba1\u54f2\u5b66\u548c\u5e94\u7528\u573a\u666f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 BPE\uff08Byte Pair Encoding\uff09\uff1a\u4ece\u538b\u7f29\u7b97\u6cd5\u5230\u5206\u8bcd\u795e\u5668<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">BPE \u7684\u8bde\u751f\u5176\u5b9e\u548c\u81ea\u7136\u8bed\u8a00\u5904\u7406\u6ca1\u4ec0\u4e48\u5173\u7cfb\u2014\u2014\u5b83\u6700\u521d\u662f 1994 \u5e74\u63d0\u51fa\u7684\u4e00\u79cd<strong>\u6570\u636e\u538b\u7f29\u7b97\u6cd5<\/strong>\uff0c\u540e\u6765\u88ab\u7814\u7a76\u8005\u53d1\u73b0\u201c\u8bf6\uff0c\u8fd9\u4e1c\u897f\u7528\u6765\u505a\u5206\u8bcd\u597d\u50cf\u4e5f\u4e0d\u9519\u201d\uff0c\u4e8e\u662f\u6447\u8eab\u4e00\u53d8\uff0c\u6210\u4e86\u4eca\u5929\u5927\u8bed\u8a00\u6a21\u578b\u9886\u57df\u4f7f\u7528\u6700\u5e7f\u6cdb\u7684\u5206\u8bcd\u7b97\u6cd5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">BPE \u7684\u6838\u5fc3\u903b\u8f91\u975e\u5e38\u7b80\u5355\uff1a<strong>\u4ece\u5b57\u7b26\u5f00\u59cb\uff0c\u53cd\u590d\u5408\u5e76\u51fa\u73b0\u9891\u7387\u6700\u9ad8\u7684\u76f8\u90bb\u5b57\u7b26\u5bf9\uff0c\u76f4\u5230\u8fbe\u5230\u9884\u8bbe\u7684\u8bcd\u6c47\u8868\u5927\u5c0f<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u521d\u59cb\u5316<\/strong>\uff1a\u5c06\u8bad\u7ec3\u8bed\u6599\u4e2d\u7684\u6bcf\u4e2a\u8bcd\u62c6\u6210\u5b57\u7b26\u5e8f\u5217\uff0c\u672b\u5c3e\u52a0\u4e00\u4e2a\u7279\u6b8a\u7b26\u53f7\uff08\u5982 <code>&lt;\/w><\/code>\uff09\u8868\u793a\u8bcd\u7ed3\u675f\u3002<\/li>\n\n\n\n<li><strong>\u7edf\u8ba1\u9891\u6b21<\/strong>\uff1a\u8ba1\u7b97\u6240\u6709\u76f8\u90bb\u5b57\u7b26\u5bf9\u51fa\u73b0\u7684\u9891\u7387\u3002<\/li>\n\n\n\n<li><strong>\u5408\u5e76<\/strong>\uff1a\u627e\u51fa\u9891\u7387\u6700\u9ad8\u7684\u90a3\u4e00\u5bf9\uff0c\u5c06\u5b83\u4eec\u5408\u5e76\u6210\u4e00\u4e2a\u65b0\u7684\u7b26\u53f7\uff0c\u52a0\u5165\u5230\u8bcd\u8868\u4e2d\u3002<\/li>\n\n\n\n<li><strong>\u5faa\u73af<\/strong>\uff1a\u91cd\u590d\u6b65\u9aa4 2-3\uff0c\u76f4\u5230\u8fbe\u5230\u9884\u8bbe\u7684\u8bcd\u6c47\u8868\u5927\u5c0f\uff08\u6bd4\u5982 32000\uff09\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u6765\u770b\u4e00\u4e2a\u7b80\u5316\u7684\u4f8b\u5b50\u3002\u5047\u8bbe\u8bad\u7ec3\u8bed\u6599\u4e2d\u6709\u4ee5\u4e0b\u8bcd\u6c47\u53ca\u9891\u7387\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>(\"hug\", 10), (\"pug\", 5), (\"pun\", 12), (\"bun\", 4), (\"hugs\", 5)<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">BPE \u7684\u5408\u5e76\u8fc7\u7a0b\u5927\u6982\u662f\u8fd9\u6837\u7684\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u521d\u59cb\u8bcd\u6c47\u8868\uff1a<code>[\"b\", \"g\", \"h\", \"n\", \"p\", \"s\", \"u\"]<\/code><\/li>\n\n\n\n<li>\u7edf\u8ba1\u53d1\u73b0 <code>\"u\"+\"g\"<\/code> \u51fa\u73b0\u6700\u9891\u7e41\uff08\u5728 hug\u3001pug\u3001hugs \u4e2d\uff09\uff0c\u5408\u5e76\u4e3a <code>\"ug\"<\/code><\/li>\n\n\n\n<li>\u63a5\u4e0b\u6765 <code>\"h\"+\"ug\"<\/code> \u51fa\u73b0\u6700\u9891\u7e41\uff0c\u5408\u5e76\u4e3a <code>\"hug\"<\/code><\/li>\n\n\n\n<li>\u7ee7\u7eed\u5408\u5e76\uff0c\u76f4\u5230\u8bcd\u8868\u6ee1\u2026\u2026<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u7ec8\uff0c\u6a21\u578b\u5b66\u5230\u7684\u8bcd\u6c47\u8868\u4e2d\u65e2\u6709\u5b8c\u6574\u7684\u5e38\u89c1\u8bcd\uff08\u5982\u201chug\u201d\uff09\uff0c\u4e5f\u6709\u9ad8\u9891\u5b50\u8bcd\uff08\u5982\u201cug\u201d\uff09\uff0c\u8fd8\u6709\u57fa\u7840\u5b57\u7b26\u3002\u8fd9\u79cd\u8bbe\u8ba1\u8ba9 BPE \u80fd\u591f<strong>\u5929\u7136\u5730\u5904\u7406\u672a\u767b\u5f55\u8bcd\uff08OOV\uff09<\/strong> \u2014\u2014\u5373\u4f7f\u9047\u5230\u8bad\u7ec3\u65f6\u6ca1\u89c1\u8fc7\u7684\u8bcd\uff0c\u4e5f\u53ef\u4ee5\u901a\u8fc7\u7ec4\u5408\u5df2\u77e5\u7684\u5b50\u8bcd\u6765\u7f16\u7801\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>BPE \u7684\u4f18\u52bf<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>OOV \u95ee\u9898\u89e3\u51b3\u5f97\u597d<\/strong>\uff1a\u7406\u8bba\u4e0a\u4efb\u4f55\u8bcd\u90fd\u53ef\u4ee5\u7528\u57fa\u7840\u5b57\u7b26\u7ec4\u5408\u8868\u793a<\/li>\n\n\n\n<li><strong>\u8bcd\u6c47\u91cf\u53ef\u63a7<\/strong>\uff1a\u901a\u8fc7\u8bbe\u7f6e\u5408\u5e76\u6b21\u6570\uff0c\u7cbe\u786e\u63a7\u5236\u8bcd\u8868\u5927\u5c0f<\/li>\n\n\n\n<li><strong>\u591a\u8bed\u8a00\u652f\u6301\u4e0d\u9519<\/strong>\uff1a\u5b57\u8282\u7ea7 BPE \u66f4\u662f\u53ef\u4ee5\u7edf\u4e00\u5904\u7406\u591a\u8bed\u8a00<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>BPE \u7684\u4e0d\u8db3<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5bf9\u4e2d\u6587\u652f\u6301\u6709\u9650<\/strong>\uff1a\u4e2d\u6587\u6ca1\u6709\u5929\u7136\u7684\u5206\u9694\u7b26\uff0c\u5408\u5e76\u6548\u679c\u4e0d\u5982\u82f1\u6587\u81ea\u7136<\/li>\n\n\n\n<li><strong>\u53ef\u80fd\u4ea7\u751f\u4e0d\u5b8c\u6574\u7684\u8bcd<\/strong>\uff1a\u8d2a\u5fc3\u5408\u5e76\u7b56\u7565\u672a\u5fc5\u662f\u8bed\u4e49\u4e0a\u7684\u6700\u4f18\u9009\u62e9<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u76ee\u524d\uff0cOpenAI \u7684 GPT \u7cfb\u5217\u3001Meta \u7684 LLaMA \u7cfb\u5217\u3001\u963f\u91cc\u7684 Qwen \u7cfb\u5217\u7b49\u4e3b\u6d41\u6a21\u578b\uff0c\u90fd\u91c7\u7528\u4e86 BPE \u6216\u5176\u53d8\u4f53\u4f5c\u4e3a\u5206\u8bcd\u65b9\u6848\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 WordPiece\uff1aBERT \u80cc\u540e\u7684\u201c\u5355\u8bcd\u62fc\u56fe\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">WordPiece \u662f Google \u4e3a BERT \u4e13\u95e8\u5f00\u53d1\u7684\u5206\u8bcd\u7b97\u6cd5\u3002\u5b83\u7684\u57fa\u672c\u601d\u8def\u548c BPE \u7c7b\u4f3c\u2014\u2014\u90fd\u662f\u4ece\u5b57\u7b26\u5f00\u59cb\u9010\u6b65\u5408\u5e76\uff0c\u4f46<strong>\u5408\u5e76\u7684\u9009\u62e9\u6807\u51c6\u4e0d\u540c<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">BPE \u7684\u5408\u5e76\u7b56\u7565\u662f\u201c\u9009\u51fa\u73b0\u9891\u7387\u6700\u9ad8\u7684\u90a3\u4e00\u5bf9\u201d\uff1b\u800c WordPiece \u7684\u7b56\u7565\u662f\u201c\u9009\u80fd\u8ba9\u8bad\u7ec3\u8bed\u6599\u6574\u4f53\u6982\u7387\u63d0\u5347\u6700\u5927\u7684\u90a3\u4e00\u5bf9\u201d\u2014\u2014\u4e5f\u5c31\u662f\u8bf4\uff0c\u5b83\u57fa\u4e8e<strong>\u6700\u5927\u4f3c\u7136\u4f30\u8ba1<\/strong>\u6765\u505a\u51b3\u7b56\uff0c\u800c\u4e0d\u662f\u7b80\u5355\u7684\u9891\u7387\u7edf\u8ba1\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u4e00\u4e2a\u663e\u8457\u7684\u533a\u522b\u662f<strong>\u6807\u8bb0\u65b9\u5f0f<\/strong>\u3002WordPiece \u4f7f\u7528 <code>##<\/code> \u524d\u7f00\u6765\u6807\u8bb0\u5b50\u8bcd\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\"playing\" \u2192 &#91;\"play\", \"##ing\"]\n\"unhappiness\" \u2192 &#91;\"un\", \"##happiness\"] \u6216 &#91;\"un\", \"##happy\", \"##ness\"]<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>##<\/code> \u7684\u542b\u4e49\u662f\u201c\u8fd9\u4e0d\u662f\u4e00\u4e2a\u8bcd\u7684\u5f00\u59cb\uff0c\u800c\u662f\u524d\u9762\u67d0\u4e2a\u8bcd\u7684\u5ef6\u7eed\u201d\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u8ba9\u6a21\u578b\u80fd\u660e\u786e\u533a\u5206\u201c\u4e00\u4e2a\u8bcd\u7684\u7b2c\u4e00\u4e2a\u5b50\u8bcd\u201d\u548c\u201c\u540e\u7eed\u5b50\u8bcd\u201d\uff0c\u5bf9\u4e8e\u82f1\u6587\u8fd9\u79cd\u9ad8\u5ea6\u4f9d\u8d56\u8bcd\u7f00\u7684\u8bed\u8a00\u7279\u522b\u6709\u7528\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>WordPiece \u7684\u62c6\u5206\u89c4\u5219\u4e5f\u66f4\u4e25\u683c<\/strong>\u3002\u5982\u679c\u5f53\u524d\u8bcd\u5728\u8bcd\u6c47\u8868\u4e2d\u627e\u4e0d\u5230\uff0cWordPiece \u4f1a\u5c1d\u8bd5\u66f4\u957f\u7684\u5b50\u8bcd\u5339\u914d\uff1b\u800c BPE \u5219\u4f1a\u8d2a\u5a6a\u5730\u5408\u5e76\u5b50\u8bcd\u76f4\u5230\u5b8c\u6210\u3002\u8fd9\u4f7f\u5f97 WordPiece \u7684\u62c6\u5206\u7ed3\u679c\u5f80\u5f80\u6bd4 BPE \u66f4\u201c\u4fdd\u5b88\u201d\uff0c\u503e\u5411\u4e8e\u4fdd\u7559\u66f4\u957f\u7684\u8bed\u4e49\u5355\u5143\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>WordPiece \u7684\u4f18\u52bf<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5bf9\u82f1\u6587\u8bed\u4e49\u7684\u6355\u6349\u66f4\u7cbe\u51c6<\/li>\n\n\n\n<li>\u62c6\u5206\u89c4\u5219\u4e25\u683c\uff0c\u9002\u5408\u7406\u89e3\u7c7b\u4efb\u52a1\uff08\u5982 BERT \u7684 MLM\uff09<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>WordPiece \u7684\u4e0d\u8db3<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8de8\u8bed\u8a00\u6cdb\u5316\u80fd\u529b\u8f83\u5f31\uff0c\u5bf9\u4e2d\u6587\u9700\u8981\u989d\u5916\u7684\u9884\u5206\u8bcd\u5904\u7406<\/li>\n\n\n\n<li>\u8bad\u7ec3\u901f\u5ea6\u6bd4 BPE \u6162<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 SentencePiece\uff1a\u6253\u7834\u7a7a\u683c\u201c\u9738\u6743\u201d\u7684\u591a\u8bed\u8a00\u65b9\u6848<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">BPE \u548c WordPiece \u90fd\u6709\u4e00\u4e2a\u5171\u540c\u7684\u201c\u76f2\u70b9\u201d\uff1a<strong>\u5b83\u4eec\u90fd\u4f9d\u8d56\u7a7a\u683c\u4f5c\u4e3a\u8bcd\u8fb9\u754c\u7684\u9884\u5224\u65ad<\/strong>\u3002\u8fd9\u5bf9\u4e8e\u82f1\u6587\u7b49\u6709\u5929\u7136\u7a7a\u683c\u5206\u9694\u7684\u8bed\u8a00\u6765\u8bf4\u6ca1\u6709\u95ee\u9898\uff0c\u4f46\u5bf9\u4e8e\u4e2d\u6587\u3001\u65e5\u6587\u3001\u6cf0\u6587\u7b49\u65e0\u7a7a\u683c\u8bed\u8a00\uff0c\u5c31\u9700\u8981\u989d\u5916\u9884\u5904\u7406\u2014\u2014\u8981\u4e48\u5148\u5206\u8bcd\u518d\u4e22\u7ed9 Tokenizer\uff0c\u8981\u4e48\u6309\u5b57\u5207\u5206\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f\uff0c<strong>SentencePiece<\/strong> \u5e94\u8fd0\u800c\u751f\u3002\u5b83\u6700\u6838\u5fc3\u7684\u521b\u65b0\u662f\uff1a<strong>\u5c06\u8f93\u5165\u6587\u672c\u89c6\u4e3a Unicode \u5b57\u7b26\u5e8f\u5217\uff0c\u5b8c\u5168\u5ffd\u7565\u7a7a\u683c\u7684\u5b58\u5728<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u610f\u5473\u7740 SentencePiece \u5bf9\u4efb\u4f55\u8bed\u8a00\u90fd\u662f\u201c\u4e00\u89c6\u540c\u4ec1\u201d\u7684\u3002\u65e0\u8bba\u662f\u82f1\u6587\u3001\u4e2d\u6587\u8fd8\u662f\u65e5\u6587\uff0c\u5728\u5b83\u773c\u91cc\u90fd\u53ea\u662f\u4e00\u4e32 Unicode \u5b57\u7b26\u6d41\u3002\u5b83\u65e2\u4e0d\u4f9d\u8d56\u7a7a\u683c\uff0c\u4e5f\u4e0d\u9700\u8981\u9884\u5148\u5206\u8bcd\uff0c\u76f4\u63a5\u4ece\u539f\u59cb\u6587\u672c\u4e2d\u5b66\u4e60\u5b50\u8bcd\u5355\u5143\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SentencePiece \u8fd8\u5185\u7f6e\u4e86\u5bf9 <strong>BPE<\/strong> \u548c <strong>Unigram<\/strong> \u4e24\u79cd\u7b97\u6cd5\u7684\u652f\u6301\uff0c\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u201c\u5206\u8bcd\u6846\u67b6\u201d\u800c\u975e\u5355\u4e00\u7b97\u6cd5\u3002\u5176\u4e2d\uff0cUnigram \u662f\u4e00\u79cd\u57fa\u4e8e\u6982\u7387\u7684\u7b97\u6cd5\uff0c\u4e0e BPE \u7684\u8d2a\u5fc3\u7b56\u7565\u4e0d\u540c\uff0c\u5b83\u901a\u8fc7\u671f\u671b\u6700\u5927\u5316\uff08EM\uff09\u7b97\u6cd5\u8fed\u4ee3\u4f18\u5316\u8bcd\u8868\uff0c\u7406\u8bba\u4e0a\u66f4\u52a0\u4f18\u96c5\u3002\u4e0d\u8fc7\uff0cUnigram \u7684\u5b9e\u73b0\u590d\u6742\u5ea6\u8f83\u9ad8\uff0c\u5728\u5b9e\u8df5\u4e2d BPE \u6a21\u5f0f\u4ecd\u7136\u66f4\u53d7\u6b22\u8fce\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>SentencePiece \u7684\u4f18\u52bf<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8bed\u8a00\u65e0\u5173\u6027<\/strong>\uff1a\u5b8c\u7f8e\u652f\u6301\u4e2d\u6587\u3001\u65e5\u6587\u7b49\u65e0\u7a7a\u683c\u8bed\u8a00<\/li>\n\n\n\n<li><strong>\u53ef\u9006\u8f6c\u6362<\/strong>\uff1a\u53ef\u4ee5\u4ece Token \u5e8f\u5217\u65e0\u635f\u8fd8\u539f\u539f\u59cb\u6587\u672c<\/li>\n\n\n\n<li><strong>\u4e00\u4f53\u5316\u5904\u7406<\/strong>\uff1a\u8bad\u7ec3\u3001\u7f16\u7801\u3001\u89e3\u7801\u90fd\u5728\u540c\u4e00\u4e2a\u6846\u67b6\u5185\u5b8c\u6210<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>SentencePiece \u7684\u4e0d\u8db3<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bad\u7ec3\u901f\u5ea6\u76f8\u5bf9\u8f83\u6162<\/li>\n\n\n\n<li>\u6a21\u578b\u6587\u4ef6\u8f83\u5927<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">LLaMA\u3001Mistral\u3001Gemma\u3001T5\u3001ALBERT \u7b49\u4f17\u591a\u6a21\u578b\u90fd\u4f7f\u7528\u4e86 SentencePiece \u4f5c\u4e3a\u5206\u8bcd\u65b9\u6848\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.4 \u4e09\u79cd\u7b97\u6cd5\u7684\u6a2a\u5411\u5bf9\u6bd4<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7279\u6027<\/th><th>BPE<\/th><th>WordPiece<\/th><th>SentencePiece<\/th><\/tr><\/thead><tbody><tr><td><strong>\u6838\u5fc3\u7b56\u7565<\/strong><\/td><td>\u9ad8\u9891\u5b57\u7b26\u5408\u5e76<\/td><td>\u6700\u5927\u4f3c\u7136\u5408\u5e76<\/td><td>\u8bed\u8a00\u65e0\u5173\u7684 Unicode \u5904\u7406<\/td><\/tr><tr><td><strong>\u8bed\u8a00\u652f\u6301<\/strong><\/td><td>\u82f1\u6587\u4e3a\u4e3b\uff0c\u591a\u8bed\u8a00\u5c1a\u53ef<\/td><td>\u82f1\u6587\u4e3a\u4e3b<\/td><td>\u591a\u8bed\u8a00\u539f\u751f\u652f\u6301<\/td><\/tr><tr><td><strong>\u7a7a\u683c\u5904\u7406<\/strong><\/td><td>\u4f9d\u8d56\u7a7a\u683c<\/td><td>\u4f9d\u8d56\u7a7a\u683c<\/td><td>\u4e0d\u4f9d\u8d56\u7a7a\u683c<\/td><\/tr><tr><td><strong>\u8bad\u7ec3\u901f\u5ea6<\/strong><\/td><td>\u5feb<\/td><td>\u4e2d\u7b49<\/td><td>\u6162<\/td><\/tr><tr><td><strong>\u4ee3\u8868\u6a21\u578b<\/strong><\/td><td>GPT, LLaMA, Qwen<\/td><td>BERT<\/td><td>LLaMA, T5, ALBERT<\/td><\/tr><tr><td><strong>\u7279\u6b8a\u6807\u8bb0<\/strong><\/td><td>\u65e0\u7279\u6b8a\u6807\u8bb0<\/td><td><code>##<\/code> \u524d\u7f00\u6807\u8bb0\u5b50\u8bcd<\/td><td>\u652f\u6301\u81ea\u5b9a\u4e49\u7279\u6b8a\u6807\u8bb0<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u8868\u683c\u53ef\u4ee5\u770b\u51fa\uff0c\u4e0d\u540c\u7b97\u6cd5\u7684\u9009\u62e9\u5f80\u5f80\u53d6\u51b3\u4e8e\u6a21\u578b\u7684\u76ee\u6807\u573a\u666f\uff1a\u7406\u89e3\u7c7b\u4efb\u52a1\u591a\u7528 WordPiece\uff0c\u751f\u6210\u7c7b\u4efb\u52a1\u591a\u7528 BPE\uff0c\u591a\u8bed\u8a00\u573a\u666f\u5219\u503e\u5411\u4e8e SentencePiece\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u5b57\u8282\u7ea7 BPE\uff1a\u73b0\u4ee3 LLM \u7684\u201c\u771f\u00b7\u6807\u914d\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 \u4ece\u5b57\u7b26\u7ea7\u5230\u5b57\u8282\u7ea7\u7684\u8fdb\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f20\u7edf\u7684 BPE \u662f\u5728<strong>\u5b57\u7b26<\/strong>\u5c42\u9762\u4e0a\u8fdb\u884c\u5408\u5e76\u7684\u3002\u4f46\u5b57\u7b26\u7684\u8fb9\u754c\u662f\u4eba\u4e3a\u5b9a\u4e49\u7684\uff08\u6bd4\u5982 Unicode \u5b57\u7b26\uff09\uff0c\u4e0d\u540c\u8bed\u8a00\u7684\u5b57\u7b26\u96c6\u5927\u5c0f\u5dee\u5f02\u5f88\u5927\uff0c\u800c\u4e14\u603b\u6709\u4e00\u4e9b\u751f\u50fb\u5b57\u7b26\u53ef\u80fd\u8d85\u51fa\u8bcd\u8868\u7684\u8986\u76d6\u8303\u56f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u7814\u7a76\u8005\u4eec\u63d0\u51fa\u4e86<strong>\u5b57\u8282\u7ea7 BPE\uff08Byte-level BPE\uff09<\/strong> \uff1a\u76f4\u63a5\u5728 UTF-8 \u7f16\u7801\u7684<strong>\u5b57\u8282\u5e8f\u5217<\/strong>\u4e0a\u5e94\u7528 BPE \u7b97\u6cd5\uff0c\u800c\u975e\u5728\u5b57\u7b26\u7ea7\u522b\u4e0a\u64cd\u4f5c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">UTF-8 \u7f16\u7801\u6709\u4e00\u4e2a\u7f8e\u5999\u7684\u6027\u8d28\uff1a<strong>\u4efb\u4f55\u5b57\u7b26\u90fd\u53ef\u4ee5\u8868\u793a\u4e3a 1 \u5230 4 \u4e2a\u5b57\u8282<\/strong>\u3002\u6574\u4e2a UTF-8 \u7684\u5b57\u8282\u7a7a\u95f4\u53ea\u6709 0 \u5230 255 \u5171 256 \u79cd\u53d6\u503c\u3002\u8fd9\u610f\u5473\u7740\uff0c\u5b57\u8282\u7ea7 BPE \u7684\u8bcd\u8868\u57fa\u7840\u5c31\u662f\u8fd9 256 \u4e2a\u5b57\u8282\u2014\u2014\u7406\u8bba\u4e0a\uff0c\u5b83\u53ef\u4ee5\u8868\u793a\u4efb\u610f Unicode \u5b57\u7b26\uff0c\u65e0\u8bba\u8fd9\u4e2a\u5b57\u7b26\u6709\u591a\u751f\u50fb\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 \u5b57\u8282\u7ea7 BPE \u7684\u8fd0\u4f5c\u673a\u5236<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5 Qwen3 \u4e3a\u4f8b\uff0c\u5b83\u7684\u5206\u8bcd\u5668\u91c7\u7528\u4e86\u5b57\u8282\u7ea7 BPE \u5b9e\u73b0\uff0c\u6838\u5fc3\u67b6\u6784\u5305\u542b\u4e09\u4e2a\u7ec4\u4ef6\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5b57\u8282\u7f16\u7801\u5668<\/strong>\uff1a\u5c06\u8f93\u5165\u6587\u672c\u8f6c\u6362\u4e3a UTF-8 \u5b57\u8282\u5e8f\u5217<\/li>\n\n\n\n<li><strong>BPE \u5408\u5e76\u8868<\/strong>\uff1a\u5b58\u50a8\u8bad\u7ec3\u5f97\u5230\u7684\u5b57\u8282\u5408\u5e76\u89c4\u5219<\/li>\n\n\n\n<li><strong>\u8bcd\u6c47\u6620\u5c04\u8868<\/strong>\uff1a\u7ef4\u62a4 token_id \u4e0e\u5b57\u8282\u5e8f\u5217\u7684\u5bf9\u5e94\u5173\u7cfb<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u67b6\u6784\u4f7f\u5f97\u6a21\u578b\u53ef\u4ee5\u9ad8\u6548\u5904\u7406\u6df7\u5408\u8bed\u8a00\u6587\u672c\uff0c\u5c24\u5176\u662f\u5728\u5305\u542b\u4e13\u4e1a\u672f\u8bed\u548c\u7279\u6b8a\u7b26\u53f7\u7684\u573a\u666f\u4e0b\u8868\u73b0\u4f18\u5f02\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5904\u7406\u4e2d\u6587\u5b57\u7b26\u201c\u4f60\u597d\u201d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>SentencePiece\uff08\u5b57\u7b26\u7ea7 BPE\uff09\uff1a\u9700\u8981 1 \u6b21\u5408\u5e76\uff08\u5b57\u7b26\u7ea7\uff09<\/li>\n\n\n\n<li>Qwen3\uff08\u5b57\u8282\u7ea7 BPE\uff09\uff1a\u9700\u8981\u5904\u7406 6 \u4e2a\u539f\u59cb\u5b57\u8282\uff0c\u7ea6 5 \u6b21\u5b57\u8282\u7ea7\u5408\u5e76<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u867d\u7136\u6b65\u9aa4\u66f4\u591a\uff0c\u4f46\u5b57\u8282\u7ea7 BPE \u6362\u6765\u7684\u597d\u5904\u662f\u5de8\u5927\u7684\uff1a<strong>\u5b8c\u5168\u8986\u76d6\u6240\u6709\u5b57\u7b26\uff0c\u6ca1\u6709\u4efb\u4f55\u5b57\u7b26\u4f1a\u88ab\u6807\u8bb0\u4e3a\u201c\u672a\u77e5\u201d<\/strong> \u3002\u5728\u5b57\u8282\u7ea7 BPE \u4e2d\uff0cOOV\uff08Out of Vocabulary\uff09\u8fd9\u4e2a\u6982\u5ff5\u51e0\u4e4e\u6d88\u5931\u4e86\u2014\u2014\u56e0\u4e3a\u4efb\u4f55\u6587\u672c\u90fd\u53ef\u4ee5\u7528 256 \u4e2a\u57fa\u7840\u5b57\u8282\u7ec4\u5408\u8868\u793a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 \u5b57\u8282\u7ea7 BPE \u6210\u4e3a\u201c\u6807\u914d\u201d\u7684\u539f\u56e0<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u4eca\u51e0\u4e4e\u6240\u6709\u4e3b\u6d41\u7684\u89e3\u7801\u5668\u6a21\u578b\u90fd\u91c7\u7528 BPE \u4f5c\u4e3a\u5206\u8bcd\u7b97\u6cd5\uff0c\u800c\u5b57\u8282\u7ea7 BPE \u66f4\u662f\u6210\u4e3a\u4e8b\u5b9e\u4e0a\u7684\u201c\u6807\u914d\u201d\u3002\u539f\u56e0\u6709\u4e09\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5168\u8986\u76d6\u6027<\/strong>\uff1a256 \u4e2a\u57fa\u7840\u5b57\u8282\u53ef\u4ee5\u8868\u793a\u4efb\u610f Unicode \u5b57\u7b26\uff0c\u5f7b\u5e95\u89e3\u51b3\u4e86 OOV \u95ee\u9898<\/li>\n\n\n\n<li><strong>\u8bad\u7ec3\u9ad8\u6548<\/strong>\uff1a\u76f8\u6bd4\u4e8e\u8bad\u7ec3\u4e00\u4e2a\u8bed\u8a00\u6a21\u578b\u9700\u8981\u6570\u4e07\u4ebf Token\uff0c\u8bad\u7ec3\u4e00\u4e2a\u597d\u7684 BPE \u5206\u8bcd\u5668\u53ea\u9700\u8981\u51e0\u767e\u4e07 Token<\/li>\n\n\n\n<li><strong>\u5de5\u7a0b\u6210\u719f<\/strong>\uff1aHugging Face \u7684 <code>tokenizers<\/code>\u3001OpenAI \u7684 <code>tiktoken<\/code>\u3001Google \u7684 <code>sentencepiece<\/code> \u7b49\u5e93\u63d0\u4f9b\u4e86\u6210\u719f\u6613\u7528\u7684\u5b9e\u73b0<\/li>\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u4ece\u201c\u9759\u6001\u8bcd\u5178\u201d\u5230\u201c\u667a\u80fd\u9002\u914d\u201d\uff1a\u6700\u65b0\u7814\u7a76\u524d\u6cbf<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8bf4\u524d\u4e09\u7ae0\u8bb2\u7684\u662f\u201c\u5df2\u7ecf\u6210\u719f\u7684\u7ecf\u5178\u6280\u672f\u201d\uff0c\u90a3\u4e48\u8fd9\u4e00\u7ae0\u8981\u804a\u7684\u5c31\u662f\u201c\u6b63\u5728\u53d1\u751f\u7684\u672a\u6765\u201d\u3002Tokenizer \u9886\u57df\u8fd1\u4e24\u5e74\u7684\u7814\u7a76\u5f02\u5e38\u6d3b\u8dc3\uff0c\u4e3b\u8981\u6cbf\u7740\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u5411\u63a8\u8fdb\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u591a\u8bed\u8a00\u516c\u5e73\u6027\uff1aTokenizer \u4e0d\u53ea\u662f\u6280\u672f\u95ee\u9898<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">2025 \u5e74\u4e00\u9879\u8986\u76d6 200 \u591a\u79cd\u8bed\u8a00\u7684\u5927\u89c4\u6a21\u7814\u7a76\u53d1\u73b0\uff1a<strong>\u62c9\u4e01\u5b57\u6bcd\u8bed\u8a00\u7684\u5206\u8bcd\u6548\u7387\u8fdc\u9ad8\u4e8e\u975e\u62c9\u4e01\u5b57\u6bcd\u548c\u5f62\u6001\u590d\u6742\u7684\u8bed\u8a00<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u6765\u8bf4\uff0c\u4ee5\u82f1\u8bed\u4e3a\u57fa\u51c6\uff08RTC = 1.0\uff09\uff0c\u5f88\u591a\u975e\u62c9\u4e01\u8bed\u8a00\u7684\u201c\u76f8\u5bf9\u5206\u8bcd\u6210\u672c\u201d\uff08Relative Tokenization Cost\uff09\u9ad8\u8fbe 3 \u5230 5 \u500d\u3002\u8fd9\u610f\u5473\u7740\uff0c\u8bf4\u8fd9\u4e9b\u8bed\u8a00\u7684\u7528\u6237\uff0c\u5728\u4f7f\u7528\u540c\u6837\u7684 LLM \u65f6\uff0c\u9700\u8981\u6d88\u8017\u66f4\u591a Token\u3001\u66f4\u591a\u7b97\u529b\u3001\u66f4\u591a\u8d39\u7528\uff0c\u800c\u4e14\u6709\u6548\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e5f\u88ab\u5927\u5e45\u538b\u7f29\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e0d\u4ec5\u4ec5\u662f\u6280\u672f\u95ee\u9898\uff0c\u66f4\u662f\u4e00\u4e2a<strong>\u57fa\u7840\u8bbe\u65bd\u5c42\u9762\u7684\u516c\u5e73\u6027\u95ee\u9898<\/strong>\u3002\u7814\u7a76\u8005\u547c\u5401\uff0c\u672a\u6765\u7684 Tokenizer \u8bbe\u8ba1\u5e94\u8be5\u66f4\u591a\u5730\u8003\u8651\u8bed\u8a00\u5f62\u6001\u5b66\u7684\u591a\u6837\u6027\uff0c\u4ece\u6e90\u5934\u51cf\u5c11\u8fd9\u79cd\u201c\u8ba1\u7b97\u4e0a\u7684\u4e0d\u516c\u5e73\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u901a\u7528\u591a\u8bed\u8a00 Tokenizer\uff1a\u4e00\u4e2a Tokenizer \u7edf\u6cbb\u6240\u6709\u8bed\u8a00\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Cohere \u56e2\u961f\u5728 2025 \u5e74\u63d0\u51fa\u4e86\u4e00\u4e2a\u6709\u8da3\u7684\u60f3\u6cd5\uff1a<strong>\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u5c31\u4f7f\u7528\u4e00\u4e2a\u8986\u76d6\u66f4\u591a\u8bed\u8a00\u7684\u901a\u7528 Tokenizer<\/strong>\uff0c\u5373\u4f7f\u8fd9\u4e9b\u8bed\u8a00\u5e76\u4e0d\u662f\u4e3b\u8981\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ed6\u4eec\u7684\u5b9e\u9a8c\u8868\u660e\uff1a\u4f7f\u7528\u901a\u7528 Tokenizer \u53ef\u4ee5\u8ba9\u6a21\u578b\u5728\u540e\u7eed\u9002\u5e94\u65b0\u8bed\u8a00\u65f6\uff0c\u80dc\u7387\u63d0\u5347\u9ad8\u8fbe 20.2%\u2014\u2014\u5373\u4f7f\u5728 Tokenizer \u548c\u9884\u8bad\u7ec3\u4e2d\u5b8c\u5168\u6ca1\u89c1\u8fc7\u7684\u8bed\u8a00\u4e0a\uff0c\u4e5f\u80fd\u83b7\u5f97\u7ea6 5% \u7684\u80dc\u7387\u63d0\u5347\u3002\u8fd9\u79cd\u201c\u8bed\u8a00\u53ef\u5851\u6027\u201d\uff08language plasticity\uff09\u7684\u53d1\u73b0\uff0c\u4e3a\u4f4e\u6210\u672c\u6269\u5c55 LLM \u7684\u8bed\u8a00\u8986\u76d6\u8303\u56f4\u63d0\u4f9b\u4e86\u65b0\u7684\u601d\u8def\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.3 \u8bcd\u8868\u9002\u5e94\uff1a\u8ba9\u8001 Tokenizer \u5b66\u4f1a\u65b0\u8bcd<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531\u4e8e Tokenizer \u7684\u8bbe\u8ba1\u5728\u6a21\u578b\u8bad\u7ec3\u4e4b\u524d\u5c31\u56fa\u5b9a\u4e86\uff0c\u4ece\u5934\u8bad\u7ec3\u4e00\u4e2a\u65b0 Tokenizer \u7684\u4ee3\u4ef7\u6781\u5176\u9ad8\u6602\u3002\u90a3\u4e48\uff0c\u6709\u6ca1\u6709\u529e\u6cd5\u5728\u4e0d\u91cd\u65b0\u8bad\u7ec3\u6574\u4e2a\u6a21\u578b\u7684\u60c5\u51b5\u4e0b\uff0c\u8ba9 Tokenizer\u201c\u5b66\u4f1a\u201d\u65b0\u8bcd\u5462\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>AdaptBPE<\/strong> \u63d0\u51fa\u4e86\u4e00\u79cd\u540e\u8bad\u7ec3\u9002\u5e94\u7b56\u7565\uff1a\u9009\u62e9\u6027\u66ff\u6362\u8bcd\u8868\u4e2d\u5229\u7528\u7387\u4f4e\u7684 Token\uff0c\u7528\u76ee\u6807\u9886\u57df\u6216\u8bed\u8a00\u4e2d\u66f4\u76f8\u5173\u7684 Token \u6765\u66ff\u4ee3\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u8fd9\u79cd\u65b9\u6cd5\u80fd\u5728\u4fdd\u6301\u8bcd\u6c47\u8868\u5927\u5c0f\u4e0d\u53d8\u7684\u60c5\u51b5\u4e0b\uff0c\u663e\u8457\u63d0\u5347\u7279\u5b9a\u9886\u57df\u7684\u7f16\u7801\u6548\u7387\uff0c\u76f8\u5f53\u4e8e\u4e00\u79cd\u201c\u8bcd\u8868\u5fae\u8c03\u201d\u8fc7\u7a0b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7c7b\u4f3c\u5730\uff0c<strong>VRCP\uff08Vocabulary Replacement Continued Pretraining\uff09<\/strong> \u4e5f\u63d0\u51fa\u901a\u8fc7\u8bcd\u8868\u66ff\u6362\u52a0\u7ee7\u7eed\u9884\u8bad\u7ec3\u7684\u65b9\u5f0f\uff0c\u9ad8\u6548\u5730\u5c06\u82f1\u8bed\u4e2d\u5fc3\u7684\u6a21\u578b\u9002\u914d\u5230\u5176\u4ed6\u8bed\u8a00\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.4 \u629b\u5f03 Tokenizer\uff1a\u5b57\u8282\u7ea7\u6a21\u578b\u7684\u65b0\u6d6a\u6f6e<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f\u5f53\u524d Tokenizer \u9886\u57df\u6700\u6fc0\u8fdb\u4e5f\u6700\u4ee4\u4eba\u5174\u594b\u7684\u65b9\u5411\uff1a<strong>\u5b8c\u5168\u629b\u5f03 Tokenizer\uff0c\u8ba9\u6a21\u578b\u76f4\u63a5\u5904\u7406\u539f\u59cb\u5b57\u8282\u6d41<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">2026 \u5e74 ICLR \u6536\u5f55\u7684 <strong>ByteFlow<\/strong> \u662f\u8fd9\u4e00\u65b9\u5411\u7684\u4ee3\u8868\u3002\u5b83\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u7684\u5c42\u6b21\u5316\u67b6\u6784\uff0c\u8ba9\u6a21\u578b\u5b66\u4e60\u5bf9\u539f\u59cb\u5b57\u8282\u6d41\u8fdb\u884c\u81ea\u9002\u5e94\u7684\u538b\u7f29\u5f0f\u5206\u6bb5\u2014\u2014\u57fa\u4e8e\u6f5c\u5728\u8868\u793a\u7684\u7f16\u7801\u7387\u52a8\u6001\u51b3\u5b9a\u5728\u54ea\u91cc\u5207\u5206\uff0c\u800c\u4e0d\u662f\u4f9d\u8d56\u9884\u5148\u5b9a\u4e49\u597d\u7684\u9759\u6001\u8bcd\u8868\u3002\u5b9e\u9a8c\u663e\u793a\uff0c\u8fd9\u79cd\u538b\u7f29\u5f0f\u5206\u6bb5\u7b56\u7565\u5728\u6027\u80fd\u4e0a\u4e0d\u4ec5\u8d85\u8d8a\u4e86 BPE \u6a21\u578b\uff0c\u8fd8\u8d85\u8d8a\u4e86\u4e4b\u524d\u6240\u6709\u7684\u5b57\u8282\u7ea7\u67b6\u6784\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u4e00\u4e2a\u6709\u8da3\u7684\u5de5\u4f5c\u662f <strong>Kathleen<\/strong>\uff0c\u4e00\u4e2a\u4ec5 73 \u4e07\u53c2\u6570\u3001\u4e0d\u4f7f\u7528 Attention\u3001\u4e5f\u4e0d\u4f7f\u7528 Tokenizer \u7684\u6587\u672c\u5206\u7c7b\u6a21\u578b\u3002\u5b83\u76f4\u63a5\u5728 UTF-8 \u5b57\u8282\u4e0a\u901a\u8fc7\u9891\u57df\u5904\u7406\u6765\u5b66\u4e60\uff0c\u5728 IMDB \u60c5\u611f\u5206\u7c7b\u4efb\u52a1\u4e0a\u751a\u81f3\u8d85\u8fc7\u4e86 16 \u500d\u53c2\u6570\u91cf\u7684\u4f20\u7edf\u6a21\u578b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b64\u5916\uff0c<strong>MambaByte<\/strong> \u5c06 Mamba \u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u9002\u914d\u5230\u5b57\u8282\u5e8f\u5217\u4e0a\uff0c<strong>SpaceByte<\/strong> \u5219\u5728\u5b57\u8282\u7ea7\u89e3\u7801\u5668\u4e0a\u53d6\u5f97\u4e86\u63a5\u8fd1 Token \u7ea7 Transformer \u7684\u6027\u80fd\u3002\u8fd9\u4e00\u7cfb\u5217\u5de5\u4f5c\u4f20\u9012\u4e86\u4e00\u4e2a\u6e05\u6670\u7684\u4fe1\u53f7\uff1a<strong>\u7aef\u5230\u7aef\u7684\u3001\u65e0 Tokenizer \u7684\u6a21\u578b\u4e0d\u4ec5\u662f\u53ef\u884c\u7684\uff0c\u800c\u4e14\u6b63\u5728\u53d8\u5f97\u8d8a\u6765\u8d8a\u6709\u7ade\u4e89\u529b<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.5 \u89c6\u89c9 Tokenizer\uff1a\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u201c\u65b0\u6218\u573a\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u7740\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u5d1b\u8d77\uff0c\u201c\u5982\u4f55\u628a\u56fe\u50cf\u3001\u89c6\u9891\u3001\u97f3\u9891\u53d8\u6210 Token\u201d\u6210\u4e86\u65b0\u7684\u7814\u7a76\u70ed\u70b9\u30022025 \u5e74\uff0c\u79bb\u6563 Tokenizer \u9886\u57df\u7684\u8fdb\u5c55\u5c24\u4e3a\u663e\u8457\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u89c6\u89c9 Tokenizer \u7684\u6838\u5fc3\u6311\u6218<\/strong>\uff1a\u7528\u4e8e\u56fe\u50cf\u751f\u6210\u7684 Tokenizer\uff08\u5982 VQ-VAE \u98ce\u683c\uff09\u64c5\u957f\u6355\u6349\u4f4e\u7ea7\u611f\u77e5\u7ec6\u8282\uff0c\u4f46\u7f3a\u4e4f\u9ad8\u7ea7\u8bed\u4e49\u7406\u89e3\u80fd\u529b\uff1b\u7528\u4e8e\u7406\u89e3\u7684 Tokenizer \u5219\u76f8\u53cd\u3002\u5982\u4f55\u5728\u4e00\u4e2a Tokenizer \u4e2d\u7edf\u4e00\u4e24\u8005\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>DualToken<\/strong> \u5c1d\u8bd5\u7528\u53cc\u91cd\u89c6\u89c9\u8bcd\u6c47\u8868\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002Apple \u56e2\u961f\u63d0\u51fa\u7684 <strong>ATOKEN<\/strong> \u5219\u53f7\u79f0\u662f\u9996\u4e2a\u80fd\u591f\u5728\u6240\u6709\u4e3b\u8981\u89c6\u89c9\u6a21\u6001\uff08\u56fe\u50cf\u3001\u89c6\u9891\u30013D\uff09\u4e0a\u8fdb\u884c\u7edf\u4e00\u5904\u7406\u7684\u89c6\u89c9\u5206\u8bcd\u5668\u3002<strong>TA-Tok\uff08Text-Aligned Tokenizer\uff09<\/strong> \u5219\u5c06\u89c6\u89c9 Token \u4e0e LLM \u7684\u8bcd\u8868\u5bf9\u9f50\uff0c\u4f7f\u56fe\u50cf\u53ef\u4ee5\u76f4\u63a5\u4f5c\u4e3a LLM \u7684\u8f93\u5165\u6765\u5904\u7406\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728<strong>\u97f3\u9891 Tokenizer<\/strong> \u9886\u57df\uff0c<strong>StableToken<\/strong> \u5f15\u5165\u4e86\u57fa\u4e8e\u5171\u8bc6\u9a71\u52a8\u673a\u5236\u7684\u591a\u5206\u652f\u67b6\u6784\uff0c\u5728\u591a\u79cd\u566a\u58f0\u6761\u4ef6\u4e0b\u663e\u8457\u63d0\u5347\u4e86 Token \u7684\u7a33\u5b9a\u6027\uff0c\u8ba9\u8bed\u97f3 LLM \u9762\u5bf9\u5608\u6742\u73af\u5883\u65f6\u66f4\u52a0\u7a33\u5065\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u7bc7 2025 \u5e74\u7684\u7efc\u8ff0\u5bf9\u79bb\u6563 Tokenizer \u7684\u8bbe\u8ba1\u539f\u5219\u3001\u5e94\u7528\u548c\u6311\u6218\u8fdb\u884c\u4e86\u7cfb\u7edf\u6027\u68b3\u7406\uff0c\u6db5\u76d6\u4e86\u751f\u6210\u3001\u7406\u89e3\u3001\u63a8\u8350\u548c\u4fe1\u606f\u68c0\u7d22\u7b49\u591a\u4e2a\u9886\u57df\uff0c\u6807\u5fd7\u7740 Tokenizer \u6b63\u5728\u4ece\u7eaf NLP \u6280\u672f\u6f14\u53d8\u4e3a\u6a2a\u8de8\u591a\u6a21\u6001 AI \u7684\u57fa\u7840\u7ec4\u4ef6\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u5b9e\u6218\u6307\u5357\uff1a\u5982\u4f55\u9009\u62e9\u5408\u9002\u7684 Tokenizer\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bb2\u4e86\u8fd9\u4e48\u591a\u7406\u8bba\uff0c\u6700\u540e\u6765\u70b9\u201c\u63a5\u5730\u6c14\u201d\u7684\u5185\u5bb9\u2014\u2014\u5728\u5b9e\u9645\u9879\u76ee\u4e2d\uff0c\u8be5\u600e\u4e48\u9009\u62e9\u548c\u914d\u7f6e Tokenizer\uff1f<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 \u573a\u666f\u9a71\u52a8\u7684\u9009\u62e9\u7b56\u7565<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u573a\u666f<\/th><th>\u63a8\u8350\u65b9\u6848<\/th><th>\u539f\u56e0<\/th><\/tr><\/thead><tbody><tr><td><strong>\u4e2d\u6587\u4e3a\u4e3b\u7684\u5e94\u7528<\/strong><\/td><td>SentencePiece (BPE)<\/td><td>\u4e0d\u4f9d\u8d56\u7a7a\u683c\uff0c\u4e2d\u6587\u9002\u914d\u597d<\/td><\/tr><tr><td><strong>\u82f1\u6587\u4e3a\u4e3b\u7684\u5e94\u7528<\/strong><\/td><td>WordPiece \u6216 BPE<\/td><td>WordPiece \u8bed\u4e49\u66f4\u7cbe\u51c6\uff0cBPE \u6548\u7387\u66f4\u9ad8<\/td><\/tr><tr><td><strong>\u591a\u8bed\u8a00\u6df7\u5408\u573a\u666f<\/strong><\/td><td>SentencePiece \u6216\u5b57\u8282\u7ea7 BPE<\/td><td>\u8bed\u8a00\u65e0\u5173\uff0c\u8de8\u8bed\u8a00\u8868\u73b0\u7a33\u5b9a<\/td><\/tr><tr><td><strong>\u4e13\u4e1a\u9886\u57df\uff08\u6cd5\u5f8b\/\u533b\u7597\uff09<\/strong><\/td><td>\u9886\u57df\u8bed\u6599\u8bad\u7ec3\u7684 BPE + \u8bcd\u6c47\u6269\u5c55<\/td><td>\u4e13\u4e1a\u672f\u8bed\u9700\u8981\u4e13\u95e8\u8986\u76d6<\/td><\/tr><tr><td><strong>\u4f4e\u8d44\u6e90\u8bed\u8a00\u9002\u914d<\/strong><\/td><td>\u8bcd\u8868\u66ff\u6362 + \u7ee7\u7eed\u9884\u8bad\u7ec3<\/td><td>\u6210\u672c\u6700\u4f4e\u7684\u9002\u914d\u65b9\u6848<\/td><\/tr><tr><td><strong>\u6781\u81f4\u6027\u80fd\u9700\u6c42<\/strong><\/td><td>tiktoken\uff08Rust \u5b9e\u73b0\uff09<\/td><td>\u63a8\u7406\u901f\u5ea6\u6700\u5feb<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 \u8bcd\u8868\u5927\u5c0f\u7684\u9009\u62e9<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bcd\u8868\u5927\u5c0f\u662f Tokenizer \u8bad\u7ec3\u4e2d\u6700\u91cd\u8981\u7684\u8d85\u53c2\u6570\u4e4b\u4e00\u3002\u4e00\u822c\u6765\u8bf4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u592a\u5c0f<\/strong>\uff08\u6bd4\u5982 &lt; 8000\uff09\uff1a\u5bfc\u81f4\u6bcf\u4e2a\u8bcd\u88ab\u62c6\u5f97\u592a\u788e\uff0c\u5e8f\u5217\u8fc7\u957f\uff0c\u8ba1\u7b97\u91cf\u589e\u52a0<\/li>\n\n\n\n<li><strong>\u592a\u5927<\/strong>\uff08\u6bd4\u5982 > 100000\uff09\uff1a\u8bcd\u8868\u672c\u8eab\u5360\u7528\u5927\u91cf\u663e\u5b58\uff0c\u4e14\u5bb9\u6613\u8fc7\u62df\u5408<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3b\u6d41 LLM \u7684\u8bcd\u8868\u5927\u5c0f\u901a\u5e38\u5728 <strong>32K \u5230 256K<\/strong> \u4e4b\u95f4\u3002LLaMA \u7cfb\u5217\u4f7f\u7528 32K\uff0cGPT-4 \u636e\u8bf4\u8fbe\u5230\u4e86 100K \u4ee5\u4e0a\u3002\u9009\u62e9\u65f6\u9700\u8981\u6743\u8861\uff1a\u5bf9\u4e8e\u591a\u8bed\u8a00\u6a21\u578b\uff0c\u8bcd\u8868\u5e94\u8be5\u5927\u4e00\u4e9b\u4ee5\u8986\u76d6\u66f4\u591a\u5b57\u7b26\uff1b\u5bf9\u4e8e\u4e13\u4e1a\u9886\u57df\u6a21\u578b\uff0c\u53ef\u4ee5\u9002\u5f53\u7f29\u5c0f\u8bcd\u8868\u4f46\u63d0\u9ad8\u9886\u57df\u8bcd\u6c47\u7684\u8986\u76d6\u7387\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.3 \u8bcd\u6c47\u6269\u5c55\u7684\u5de5\u7a0b\u5b9e\u8df5<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u9700\u8981\u5411\u5df2\u6709\u6a21\u578b\u6dfb\u52a0\u65b0\u7684\u4e13\u4e1a\u672f\u8bed\uff08\u6bd4\u5982\u201c\u91cf\u5b50\u8ba1\u7b97\u201d\u201c\u795e\u7ecf\u7f51\u7edc\u201d\uff09\uff0c\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u65b9\u6cd5\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u4f7f\u7528 Hugging Face tokenizers \u6dfb\u52a0\u65b0\u8bcd\nnew_tokens = &#91;\"\u91cf\u5b50\u8ba1\u7b97\", \"\u795e\u7ecf\u7f51\u7edc\", \"transformer\"]\ntokenizer.add_tokens(new_tokens)\nmodel.resize_token_embeddings(len(tokenizer))<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6dfb\u52a0\u7684 Token \u4f1a\u83b7\u5f97\u6bd4 BPE \u5206\u8bcd\u66f4\u9ad8\u7684\u4f18\u5148\u7ea7\uff0c\u4ece\u800c\u4fdd\u8bc1\u4e13\u4e1a\u672f\u8bed\u4e0d\u88ab\u9519\u8bef\u62c6\u5206\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e\u9700\u8981\u5927\u89c4\u6a21\u66f4\u65b0\u8bcd\u8868\u7684\u573a\u666f\uff0c\u53ef\u4ee5\u8003\u8651\u57fa\u4e8e\u73b0\u6709\u8bcd\u8868\u8fdb\u884c<strong>\u589e\u91cf\u8bad\u7ec3<\/strong>\uff0c\u751f\u6210\u65b0\u7684\u5408\u5e76\u89c4\u5219\u6587\u4ef6\uff0c\u7136\u540e\u66ff\u6362\u6a21\u578b\u539f\u6709\u7684\u5206\u8bcd\u914d\u7f6e\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u603b\u7ed3\u4e0e\u5c55\u671b<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">6.1 Tokenizer \u7684\u201c\u4e09\u91cd\u8eab\u4efd\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u987e\u5168\u6587\uff0cTokenizer \u5728 LLM \u4f53\u7cfb\u4e2d\u626e\u6f14\u7740\u4e09\u91cd\u89d2\u8272\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7ffb\u8bd1\u5b98<\/strong>\uff1a\u5c06\u4eba\u7c7b\u8bed\u8a00\u7ffb\u8bd1\u6210\u6a21\u578b\u80fd\u61c2\u7684\u6570\u5b57\u5e8f\u5217<\/li>\n\n\n\n<li><strong>\u5b88\u95e8\u4eba<\/strong>\uff1a\u51b3\u5b9a\u54ea\u4e9b\u4fe1\u606f\u80fd\u8fdb\u5165\u6a21\u578b\u3001\u4ee5\u4ec0\u4e48\u7c92\u5ea6\u8fdb\u5165<\/li>\n\n\n\n<li><strong>\u516c\u5e73\u6027\u7684\u5b88\u62a4\u8005<\/strong>\uff1a\u4e0d\u540c\u8bed\u8a00\u7684\u5206\u8bcd\u6548\u7387\uff0c\u76f4\u63a5\u5f71\u54cd\u7740 AI \u7684\u5168\u7403\u53ef\u53ca\u6027<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u4e2a\u8bbe\u8ba1\u5f97\u5f53\u7684 Tokenizer\uff0c\u80fd\u591f\u8ba9\u6a21\u578b\u8bad\u7ec3\u66f4\u9ad8\u6548\u3001\u63a8\u7406\u66f4\u5feb\u901f\u3001\u8de8\u8bed\u8a00\u80fd\u529b\u66f4\u5f3a\uff1b\u800c\u4e00\u4e2a\u7cdf\u7cd5\u7684 Tokenizer\uff0c\u5219\u53ef\u80fd\u8ba9\u6a21\u578b\u5728\u9762\u5bf9\u751f\u50fb\u8bcd\u3001\u591a\u8bed\u8a00\u6df7\u5408\u6587\u672c\u3001\u4e13\u4e1a\u672f\u8bed\u65f6\u6349\u895f\u89c1\u8098\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">6.2 \u672a\u6765\u8d8b\u52bf\u4e00\u89c8<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u65b9\u5411<\/th><th>\u4ee3\u8868\u5de5\u4f5c<\/th><th>\u6838\u5fc3\u7a81\u7834<\/th><\/tr><\/thead><tbody><tr><td><strong>\u591a\u8bed\u8a00\u516c\u5e73\u6027<\/strong><\/td><td>Tokenization Disparities \u7814\u7a76<\/td><td>\u91cf\u5316\u4e86 Tokenizer \u5bf9\u4e0d\u540c\u8bed\u8a00\u7684\u8ba1\u7b97\u4e0d\u516c\u5e73<\/td><\/tr><tr><td><strong>\u901a\u7528\u591a\u8bed\u8a00 Tokenizer<\/strong><\/td><td>One Tokenizer To Rule Them All<\/td><td>\u4e00\u4e2a\u8bcd\u8868\u8986\u76d6\u591a\u8bed\u8a00\uff0c\u63d0\u5347\u9002\u5e94\u80fd\u529b<\/td><\/tr><tr><td><strong>\u8bcd\u8868\u9002\u5e94<\/strong><\/td><td>AdaptBPE, VRCP<\/td><td>\u4e0d\u91cd\u65b0\u8bad\u7ec3\uff0c\u8ba9\u8001\u8bcd\u8868\u5b66\u4f1a\u65b0\u8bcd<\/td><\/tr><tr><td><strong>\u629b\u5f03 Tokenizer<\/strong><\/td><td>ByteFlow, Kathleen, MambaByte<\/td><td>\u7aef\u5230\u7aef\u5b57\u8282\u7ea7\u5efa\u6a21\uff0c\u4e0d\u518d\u9700\u8981\u9759\u6001\u8bcd\u8868<\/td><\/tr><tr><td><strong>\u89c6\u89c9\/\u97f3\u9891 Tokenizer<\/strong><\/td><td>DualToken, TA-Tok, StableToken<\/td><td>\u5c06\u79bb\u6563 Token \u6269\u5c55\u5230\u591a\u6a21\u6001\u9886\u57df<\/td><\/tr><\/tbody><\/table><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>\u5199\u5728\u524d\u9762 \u5728\u4e0a\u4e00\u7bc7\u6587\u7ae0\u4e2d\uff0c\u5b8c\u6574\u5730\u8d70\u4e86\u4e00\u904d LLM \u7684\u67b6\u6784\u5168\u666f\u56fe\u3002\u90a3\u7bc7\u6587\u7ae0\u63d0\u5230\uff1a\u539f\u59cb\u6587\u672c\u8fdb\u5165\u6a21\u578b\u7684\u7b2c\u4e00\u6b65\uff0c\u5c31\u662f\u7ecf [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":124,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[37,38],"class_list":["post-122","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb","tag-llm","tag-tokenizer"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/122","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/comments?post=122"}],"version-history":[{"count":1,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/122\/revisions"}],"predecessor-version":[{"id":123,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/122\/revisions\/123"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/124"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=122"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=122"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=122"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}