{"id":37588,"date":"2025-01-31T14:01:10","date_gmt":"2025-01-31T06:01:10","guid":{"rendered":"https:\/\/17aitech.com\/?p=37588"},"modified":"2025-01-31T14:01:10","modified_gmt":"2025-01-31T06:01:10","slug":"%e4%bb%8e2019%e5%b9%b4%e5%88%b0%e7%8e%b0%e5%9c%a8%ef%bc%8c%e6%98%af%e6%97%b6%e5%80%99%e9%87%8d%e6%96%b0%e5%ae%a1%e8%a7%86tokenization%e4%ba%86","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=37588","title":{"rendered":"\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86"},"content":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86<\/a><\/p>\n

\n
\n
2019 \u5e74\u95ee\u4e16\u7684 GPT-2\uff0c\u5176 tokenizer \u4f7f\u7528\u4e86 BPE \u7b97\u6cd5\uff0c\u8fd9\u79cd\u7b97\u6cd5\u81f3\u4eca\u4ecd\u5f88\u5e38\u89c1\uff0c\u4f46\u8fd9\u79cd\u65b9\u5f0f\u662f\u6700\u4f18\u7684\u5417\uff1f\u6765\u81ea HuggingFace \u7684\u4e00\u7bc7\u6587\u7ae0\u7ed9\u51fa\u4e86\u89e3\u91ca\u3002<\/section>\n<\/section>\n<\/blockquote>\n
<\/section>\n
\u300c9.9 \u548c 9.11 \u5230\u5e95\u54ea\u4e2a\u5927\uff1f\u300d\u8fd9\u4e2a\u95ee\u9898\u4e00\u5ea6\u96be\u574f\u4e86\u5404\u5bb6\u5927\u6a21\u578b\u3002<\/section>\n
<\/section>\n
\u5173\u4e8e\u6a21\u578b\u4e3a\u4ec0\u4e48\u4f1a\u7b54\u9519\uff0c\u7814\u7a76\u4eba\u5458\u7ed9\u51fa\u4e86\u5404\u79cd\u731c\u6d4b\uff0c\u5305\u62ec\u9884\u8bad\u7ec3\u6570\u636e\u7684\u6784\u6210\u548c\u6a21\u578b\u67b6\u6784\u672c\u8eab\u3002<\/section>\n
<\/section>\n
\u5728\u4e00\u7bc7\u65b0\u535a\u5ba2\u4e2d\uff0c\u6765\u81ea HuggingFace \u7684\u7814\u7a76\u8005\u8ba8\u8bba\u4e86\u53ef\u80fd\u9020\u6210\u8fd9\u4e00\u95ee\u9898\u7684\u539f\u56e0\u4e4b\u4e00 \u2014\u2014tokenization\uff0c\u5e76\u91cd\u70b9\u5206\u6790\u4e86\u5b83\u5982\u4f55\u5f71\u54cd\u6a21\u578b\u7684\u6570\u5b66\u80fd\u529b\uff0c\u5c24\u5176\u662f\u7b97\u672f\u80fd\u529b\u3002<\/section>\n
<\/section>\n
<\/section>\n
\u56de\u987e Tokenization<\/strong><\/section>\n
<\/section>\n
\u65e9\u5728 2019 \u5e74\uff0cGPT-2 \u8bba\u6587\u5c31\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u5c06 BPE\uff08byte-pair encoding\uff09\u7528\u4e8e\u8bed\u8a00\u6a21\u578b\u7684 tokenization \u65b9\u6cd5\u3002\u6b64\u65b9\u6cd5\u7684\u5de5\u4f5c\u539f\u7406\u662f\u5c06\u9891\u7e41\u51fa\u73b0\u7684\u5b50\u8bcd\u5408\u5e76\u4e3a\u5355\u4e2a\u5355\u5143\uff0c\u76f4\u5230\u8bcd\u6c47\u91cf\u8fbe\u5230\u76ee\u6807\u5927\u5c0f\u3002<\/section>\n
<\/section>\n
\u7136\u800c\uff0c\u8fd9\u79cd\u505a\u6cd5\u751f\u6210\u7684\u8bcd\u6c47\u8868\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e\u8f93\u5165\u5230 tokenizer \u4e2d\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u4ece\u800c\u5bfc\u81f4\u4e86\u5728\u6570\u5b57\u7f16\u7801\u65b9\u5f0f\u4e0a\u7684\u4e0d\u4e00\u81f4\u6027\u3002\u4f8b\u5982\uff0c\u5728\u8bad\u7ec3\u6570\u636e\u4e2d\u5e38\u89c1\u7684\u6570\u5b57\uff08\u4f8b\u5982 1-100\u30011943 \u5e74\u8fd9\u6837\u7684\u8868\u793a\uff09\u5f88\u53ef\u80fd\u88ab\u8868\u793a\u4e3a\u5355\u4e2a token\uff0c\u800c\u8f83\u5c11\u89c1\u5230\u7684\u6570\u5b57\u5219\u88ab\u62c6\u5206\u6210\u591a\u4e2a token\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/section>\n
<\/a><\/section>\n
\u56db\u5e74\u540e\uff0cLlama \u7cfb\u5217\u6765\u4e86\uff01Llama \u548c Llama 2 \u4f7f\u7528 SentencePiece \uff08\u4e00\u4e2a\u7528\u4e8e\u57fa\u4e8e\u6587\u672c\u751f\u6210\u7684\u65e0\u76d1\u7763\u6587\u672c tokenizer \uff09\u7684 BPE \u5b9e\u73b0\uff0c\u5e76\u5bf9\u6570\u5b57\u8fdb\u884c\u4e86\u663e\u8457\u7684\u8c03\u6574\uff1a\u5b83\u4eec\u5c06\u6240\u6709\u6570\u5b57\u62c6\u5206\u4e3a\u5355\u4e2a\u6570\u5b57\u3002\u8fd9\u610f\u5473\u7740\u53ea\u6709 10 \u4e2a\u552f\u4e00 token\uff080-9\uff09\u6765\u8868\u793a\u4efb\u4f55\u6570\u5b57\uff0c\u4ece\u800c\u7b80\u5316\u4e86 LLM \u7684\u6570\u5b57\u8868\u793a\u3002Deepseek \u540e\u6765\u53d1\u5e03\u4e86\u4e00\u4e2a\u6a21\u578b (DeepSeek-V2)\uff0c\u5b83\u6709\u4e00\u4e2a\u7c7b\u4f3c\u7684\u5355\u4f4d\u6570\uff08single-digit\uff09\u7684 tokenizer \u3002<\/section>\n
<\/section>\n
\u540e\u6765\uff0cLlama 3 \u91c7\u7528\u4e86\u4e0d\u540c\u7684\u65b9\u6cd5\u6765\u5904\u7406\u6570\u5b57\uff0c\u5c06\u5b83\u4eec tokenizing \u4e3a\u4e09\u4f4d\u6570\u3002\u56e0\u6b64\uff0c\u4ece 1 \u5230 999 \u7684\u6570\u5b57\u6bcf\u4e2a\u6570\u90fd\u6709\u552f\u4e00\u7684 token\uff0c\u800c\u4ece 1000 \u5f00\u59cb\u7684\u6570\u5b57\u7531\u8fd9\u4e9b token \u7ec4\u6210\u3002<\/section>\n
<\/section>\n
\u4e00\u4e2a\u65b0\u7684\u8303\u5f0f\uff1a\u4ece\u53f3\u5230\u5de6\u7684 Tokenization<\/strong><\/section>\n
<\/section>\n
\u5230\u76ee\u524d\u4e3a\u6b62\uff0c\u6211\u4eec\u6240\u770b\u5230\u7684 tokenization \u65b9\u6cd5\u90fd\u662f\u4ece\u5de6\u5230\u53f3\u5904\u7406\u6587\u672c\u7684\u3002\u4f8b\u5982\uff0c\u5982\u679c\u4e09\u4f4d\u6570\u5b57\u7684\u5206\u8bcd\u6cd5\u9047\u5230\u5e8f\u5217 12345\uff0c\u5b83\u5c06\u4ece\u5f00\u5934\u626b\u63cf\uff0c\u5c06\u5176\u5206\u89e3\u4e3a 123 \u548c 45 \u8fd9\u6837\u7684\u7247\u6bb5\u3002<\/section>\n
<\/section>\n
\u4e0e\u4ece\u5de6\u5230\u53f3\uff08L2R\uff09\u7684\u5206\u8bcd\u65b9\u6cd5\u4e0d\u540c\uff0c\u4ece\u53f3\u5230\u5de6\uff08R2L\uff09\u7684\u5206\u8bcd\u65b9\u6cd5\u4ee5\u4e09\u4e2a\u5b57\u7b26\u4e3a\u4e00\u7ec4\uff0c\u4ece\u6587\u672c\u7684\u672b\u5c3e\u5f00\u59cb\u5411\u5f00\u5934\u5904\u7406\u3002\u4f7f\u7528 R2L \u5206\u8bcd\uff0c\u5e8f\u5217 12345 \u5c06\u901a\u8fc7\u4ece\u53f3\u4fa7\u626b\u63cf\u8fdb\u884c\u5206\u8bcd\uff0c\u9996\u5148\u5206\u5272\u51fa 345\uff0c\u7136\u540e\u518d\u5904\u7406 12\u3002\u6700\u8fd1\uff0c\u4e00\u4e9b\u524d\u6cbf\u7684\u95ed\u6e90\u6a21\u578b\u4e5f\u5728\u63a2\u7d22\u4f7f\u7528\u8fd9\u79cd R2L \u5206\u8bcd\u65b9\u6cd5\uff0c\u8fd9\u5df2\u7ecf\u88ab\u8bc1\u660e\u5bf9\u67d0\u4e9b\u7b97\u672f\u8fd0\u7b97\u6709\u76ca\uff0c\u56e0\u4e3a R2L \u8868\u793a\u53ef\u4ee5\u9632\u6b62\u64cd\u4f5c\u6570\u7684\u9519\u4f4d\u3002\u8fd8\u6709\u4f20\u8a00\u79f0 Claude \u4f7f\u7528\u4e86\u8fd9\u79cd R2L \u5206\u8bcd\u65b9\u6cd5\u3002<\/section>\n
<\/section>\n
\u4e3a\u4e86\u66f4\u597d\u5730\u7406\u89e3\u9519\u4f4d\u662f\u4ec0\u4e48\u6837\u5b50\u7684\uff0c\u8ba9\u6211\u4eec\u4ee5 3789 + 8791 \u4e3a\u4f8b\uff1a<\/section>\n
<\/a><\/section>\n
\u5982\u4e0a\u6240\u793a\uff0c\u5728\u4e09\u4f4d\u6570\u4ece\u5de6\u5230\u53f3\uff08L2R\uff09\u7684\u4f8b\u5b50\u4e2d\uff0c9 + 1 \u5e94\u8be5\u6620\u5c04\u5230\u6570\u5b57 0\uff0c\u4f46\u5b9e\u9645\u4e0a\u5374\u4e0e 8 \u7ec4\u5408\u5728\u4e00\u8d77\u5f62\u6210\u4e86 80\uff0c\u56e0\u4e3a\u524d\u9762\u7684\u4e09\u4e2a token\uff08125\uff09\u5df2\u7ecf\u88ab\u5206\u5728\u4e00\u8d77\u4e86\u3002tokenization \u8fb9\u754c\u7684\u504f\u79fb\u5728\u5b66\u4e60\u8fc7\u7a0b\u4e2d\u5f15\u5165\u4e86\u989d\u5916\u7684\u590d\u6742\u6027\uff0c\u5df2\u7ecf\u8bc1\u660e\u51c6\u786e\u6027\u662f\u6709\u5bb3\u7684\u3002<\/section>\n
<\/section>\n
\u800c\u5728\u4ece\u53f3\u5230\u5de6\uff08R2L\uff09\u7684\u4f8b\u5b50\u4e2d\uff0c\u6570\u5b57 580 \u548c\u5bf9\u5e94\u7684\u5b50\u64cd\u4f5c\u6570 789 \u548c 791 \u5f88\u597d\u5730\u5bf9\u9f50\u4e86\u3002<\/section>\n
<\/section>\n
\u4ee5\u4e0b\u662f\u7528\u4e8e\u5904\u7406\u6570\u5b57 tokenization \u7684\u6280\u672f\u6982\u8ff0\uff1a<\/section>\n
<\/a><\/section>\n
\u4e0d\u540c\u65b9\u6cd5\u7684\u6bd4\u8f83<\/strong><\/section>\n
<\/section>\n
\u8be5\u7814\u7a76\u65e8\u5728\u6bd4\u8f83\u591a\u4e2a tokenizer \u4ee5\u53ca\u5b83\u4eec\u5904\u7406\u6570\u5b57\u7684\u4e0d\u540c\u65b9\u5f0f\uff0c\u4ee5\u5c3d\u91cf\u51cf\u5c11\u6a21\u578b\u67b6\u6784\u3001\u8bad\u7ec3\u914d\u7f6e\u548c\u9884\u8bad\u7ec3\u6570\u636e\u7b49\u5916\u90e8\u56e0\u7d20\u5728\u8bc4\u4f30\u7ed3\u679c\u4e2d\u7684\u5f71\u54cd\u3002\u56e0\u6b64\uff0c\u6bcf\u4e2a\u6a21\u578b\u4e4b\u95f4\u552f\u4e00\u7684\u533a\u522b\u5e94\u8be5\u662f tokenizer\u3002<\/section>\n
<\/section>\n
\u5b9e\u9a8c\u9009\u62e9\u4e86 3 \u79cd tokenizer\uff0c\u5206\u522b\u662f GPT-2 \u7684 BPE tokenizer\u3001Llama 3 \u7684\u4e09\u4f4d\u6570 tokenizer\uff08three-digit tokenizer\uff09\u548c Deepseek \u7684\u5355\u4f4d\u6570 tokenizer\uff08single-digit tokenizer\uff09\u3002<\/section>\n
\n
from transformers import AutoTokenizer<\/code><\/section>
from tokenizers import pre_tokenizers, Regex<\/code>\r\n<\/code>\r\n<\/code># Initialize all tokenizers<\/code><\/section>
tokenizer = AutoTokenizer.from_pretrained (\"meta-llama\/Meta-Llama-3-8B\")<\/code>\r\n<\/code>\r\n<\/code># Add an extra step to the existing pre-tokenizer steps<\/code><\/section>
tokenizer._tokenizer.pre_tokenizer = pre_tokenizers.Sequence (<\/code><\/section>
            [<\/code><\/section>
                # Added step: split by R2L digits<\/code><\/section>
                pre_tokenizers.Split (pattern = Regex (r\"d {1,3}(?=(d {3})*b)\"),<\/code><\/section>
\u00a0<\/code>                  behavior=\"isolated\", invert = False),<\/code><\/section>
                # Below: Existing steps from Llama 3's tokenizer<\/code><\/section>
                pre_tokenizers.Split (pattern=Regex (r\"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^rnp {L}p {N}]?p {L}+|p {N}{1,3}| ?[^sp {L}p {N}]+[rn]*|s*[rn]+|s+(?!S)|s+\"),<\/code><\/section>
\u00a0<\/code>                  behavior=\"isolated\", invert=False),<\/code><\/section>
\u00a0<\/code>                pre_tokenizers.ByteLevel (add_prefix_space=False, trim_offsets=True, use_regex=False)<\/code><\/section>
            ]<\/code><\/section>
        )<\/code>\r\n<\/code>\r\n<\/code>print (tokenizer.tokenize (\"42069\")) # [42, 069]<\/code><\/section><\/pre>\n<\/section>\n
<\/section>\n
\u8bad\u7ec3\u6a21\u578b\u4f7f\u7528\u4e86\u539f\u59cb\u7684 Llama \u67b6\u6784\uff0c\u6b64\u5916\uff0c\u8be5\u7814\u7a76\u8fd8\u8c03\u6574\u4e86\u9690\u85cf\u5c42\u7684\u6570\u91cf\uff0c\u4ee5\u786e\u4fdd\u6bcf\u4e2a\u6a21\u578b\u5927\u81f4\u5177\u6709\u76f8\u540c\u6570\u91cf\u7684\u53c2\u6570\uff08\u7ea6 14.5 \u4ebf\uff09\u3002<\/section>\n
<\/a><\/section>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u4e3a\u4e86\u4fdd\u6301\u6052\u5b9a\u7684\u8ba1\u7b97\u9884\u7b97\uff0c\u672c\u6587\u51cf\u5c11\u4e86\u5177\u6709\u66f4\u5927\u8bcd\u6c47\u8868\u6a21\u578b\u4e2d\u7684\u9690\u85cf\u5c42\u6570\u91cf\u3002<\/sup><\/em><\/section>\n
<\/section>\n
\u7ed3\u679c<\/strong><\/section>\n
<\/section>\n
\u7b97\u672f\u95ee\u9898<\/strong><\/section>\n
<\/section>\n
\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5355\u4f4d\u6570 tokenization \u4f18\u4e8e\u5176\u4ed6 tokenizer \u65b9\u6cd5\u3002<\/section>\n
<\/a><\/section>\n
<\/a><\/section>\n
<\/a><\/section>\n
\u7ed3\u679c\u663e\u793a\uff0c\u867d\u7136\u5728\u8f83\u7b80\u5355\u7684\u95ee\u9898\u4e0a\u5dee\u5f02\u4e0d\u592a\u660e\u663e\uff0c\u4f46\u968f\u7740\u95ee\u9898\u590d\u6742\u6027\u7684\u589e\u52a0\uff0c\u8868\u73b0\u6700\u4f73\u7684 tokenizer\uff08\u5355\u4f4d\u6570\u5206\u8bcd\uff09\u4e0e\u5176\u4ed6 tokenizer \u4e4b\u95f4\u7684\u5dee\u8ddd\u8d8a\u6765\u8d8a\u5927\u3002\u8fd9\u8868\u660e\u5355\u4f4d\u6570\u5206\u8bcd\u5bf9\u4e8e\u8f93\u5165\u6570\u636e\u957f\u5ea6\u7684\u53d8\u5316\u66f4\u4e3a\u9c81\u68d2\uff0c\u5e76\u4e14\u80fd\u591f\u66f4\u597d\u5730\u6355\u6349\u590d\u6742\u7684\u6a21\u5f0f\uff0c\u4ece\u800c\u5728\u5176\u4ed6\u5206\u8bcd\u65b9\u6cd5\u96be\u4ee5\u5e94\u5bf9\u7684\u573a\u666f\u4e2d\u63d0\u5347\u6027\u80fd\u3002<\/section>\n
<\/section>\n
\u6b64\u5916\uff0c\u672c\u6587\u8fd8\u53d1\u73b0\u6d6e\u70b9\u6570\u548c\u6574\u6570\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u8ddd\u5728\u6240\u6709 tokenizer \u4e2d\u90fd\u662f\u76f8\u4f3c\u7684\u3002\u8fd9\u8868\u660e\u5728\u8fd9\u4e24\u4e2a\u7c7b\u522b\u4e2d\u9009\u62e9 tokenizer \u65f6\uff0c\u5e76\u4e0d\u5b58\u5728\u56fa\u6709\u7684\u6743\u8861\uff0c\u5373\u5bf9\u4e8e\u6574\u6570\u6700\u4f18\u7684 tokenizer \u5bf9\u4e8e\u6d6e\u70b9\u6570\u4e5f\u662f\u6700\u4f18\u7684\u3002<\/section>\n
<\/section>\n
\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u4e09\u4f4d\u6570 R2L tokenization \u6bd4\u6807\u51c6\u4e09\u4f4d\u6570 L2R tokenization \u5177\u6709\u66f4\u597d\u7684\u6027\u80fd\u3002<\/section>\n
<\/a><\/section>\n
\u672c\u6587\u53d1\u73b0\uff0c\u4e0e\u4f7f\u7528\u9ed8\u8ba4 L2R token \u6570\u636e\u8fdb\u884c\u8bad\u7ec3\u76f8\u6bd4\uff0c\u4f7f\u7528 R2L token \u6570\u636e\u8fdb\u884c\u8bad\u7ec3\u7684\u6a21\u578b\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6539\u8fdb\uff08\u4e58\u6cd5\u9664\u5916\uff09\u3002\u8fd9\u8868\u660e\uff0c\u4e0e\u5178\u578b\u7684\u4ece\u5de6\u5230\u53f3\u7f16\u7801\u76f8\u6bd4\uff0c\u5b83\u662f\u7b97\u672f\u8fd0\u7b97\u7684\u6700\u4f73\u8bbe\u7f6e\u3002<\/section>\n
<\/section>\n
\u5f53\u6570\u5b57\u88ab\u4ece\u53f3\u5411\u5de6\u6bcf 3 \u4f4d\u4e00\u7ec4\u8fdb\u884c\u5206\u5757\u65f6\uff0cPure-BPE\uff08Byte Pair Encoding\uff09tokenizer \u663e\u793a\u51fa\u4e0d\u4e00\u81f4\u7684\u6027\u80fd\u3002<\/section>\n
<\/a><\/section>\n
\u663e\u7136\uff0c\u6ca1\u6709\u4efb\u4f55\u989d\u5916\u6570\u5b57\u9884\u5904\u7406\u7684\u7eaf\u57fa\u4e8e BPE \u7684 tokenizer \u4e0d\u4f1a\u4ece\u4f7f\u7528 R2L token \u5316\u4e2d\u53d7\u76ca\u3002\u4e00\u4e2a\u53ef\u80fd\u7684\u89e3\u91ca\u662f\uff0c\u8fd9\u4e9b tokenizer \u4e2d\u6570\u5b57\u5206\u7ec4\u7684\u65b9\u5f0f\u7f3a\u4e4f\u7ed3\u6784\u3002<\/section>\n
<\/section>\n
\u57fa\u4e8e\u5355\u8bcd\u7684\u95ee\u9898<\/strong><\/section>\n
<\/a><\/strong><\/section>\n
\u867d\u7136\u5728\u57fa\u4e8e\u5355\u8bcd\u7684\u95ee\u9898\u4e0a\uff0c\u4e0d\u540c tokenizer \u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u8ddd\u4e0d\u592a\u660e\u663e\uff0c\u4f46\u672c\u6587\u89c2\u5bdf\u5230\u5355\u4f4d\u6570 tokenizer \u548c\u4e09\u4f4d\u6570 tokenizer \u901a\u5e38\u4f18\u4e8e\u57fa\u4e8e BPE \u7684 tokenizer\u3002\u8fd9\u8868\u660e\uff0c\u65e0\u8bba\u662f\u5355\u8bcd\u95ee\u9898\u8fd8\u662f\u6570\u5b57\u95ee\u9898\uff0c\u8fd9\u79cd\u8d8b\u52bf\u90fd\u662f\u4e00\u81f4\u7684\u3002<\/section>\n
<\/section>\n
Llama 3 R2L \u63a8\u7406<\/strong><\/section>\n
<\/section>\n
\u63a5\u4e0b\u6765\u672c\u6587\u8fdb\u884c\u4e86\u53e6\u4e00\u9879\u6d4b\u8bd5\uff0c\u5373\u73b0\u6709\u7684\u9884\u8bad\u7ec3 \/ \u6307\u4ee4\u6a21\u578b\u5728\u63a5\u53d7\u4e0e\u6700\u521d\u8bad\u7ec3\u65b9\u6848\u4e0d\u540c\u7684 token \u5316\u65b9\u6848\u65f6\u8868\u73b0\u5982\u4f55\uff0c\u800c\u65e0\u9700\u91cd\u65b0\u8bad\u7ec3\u6216\u5fae\u8c03\u3002\u56e0\u6b64\uff0c\u672c\u6587\u57fa\u4e8e Llama3 8B Instruct \u6a21\u578b\uff0c\u5e76\u4f7f\u7528\u4e0a\u8ff0\u76f8\u540c\u7684\u4ee3\u7801\u4fee\u6539\u5176 tokenizer\uff0c\u4ee5\u5728\u63a8\u7406\u671f\u95f4\u6267\u884c R2L tokenization\uff0c\u800c\u65e0\u9700\u91cd\u65b0\u8bad\u7ec3\u65b0\u6a21\u578b\u3002<\/section>\n
<\/section>\n
\u5728\u4e09\u4f4d\u6570 tokenization \u65b9\u6848\u4e2d\u8fdb\u884c\u4e24\u4e2a\u6570\u76f8\u52a0\u9700\u8981\u6ce8\u610f\u7684\u662f\uff1a\u7ed3\u679c\u6709\u65f6\u4f1a\u4ea7\u751f\u6bd4\u8f93\u5165\u6570\u5b57\u66f4\u591a\u7684 token\u3002\u4f8b\u5982\u5c06 999 \u548c 111 \u76f8\u52a0\u65f6\uff0c\u5b83\u4eec\u5355\u72ec\u53ea\u9700\u8981\u4e00\u4e2a token\uff0c\u4f46\u662f\u5f53\u5b83\u4eec\u76f8\u52a0\u4ea7\u751f 1110 \u65f6\uff0c\u9700\u8981\u4e24\u4e2a token\uff081 \u548c 110\uff09\u3002\u57fa\u4e8e\u8fd9\u4e2a\u89c2\u5bdf\uff0c\u672c\u6587\u60f3\u63a2\u7d22\u5728\u4f7f\u7528 L2R \u548c R2L tokenization \u5bf9\u4e0d\u540c\u7684 token \u957f\u5ea6\u6267\u884c\u52a0\u6cd5\u65f6\uff0c\u4f1a\u4ea7\u751f\u591a\u5927\u7684\u5dee\u5f02\u3002<\/section>\n
<\/section>\n
\u63a5\u4e0b\u6765\uff0c\u672c\u6587\u5c06\u628a\u5bfc\u81f4\u989d\u5916 token \u7684\u52a0\u6cd5\u79f0\u4e3a\u8fdb\u4f4d\uff08carry\uff09\u52a0\u6cd5\uff0c\u800c\u90a3\u4e9b\u6ca1\u6709\u8fdb\u4f4d\u7684\u52a0\u6cd5\u79f0\u4e3a\u65e0\u8fdb\u4f4d\uff08without carry\uff09\u52a0\u6cd5\u3002<\/section>\n
<\/section>\n
\u672c\u6587\u7528 Llama3 8B Instruct \u6267\u884c\u4e86\u4e0d\u540c\u6570\u5b57\u957f\u5ea6\u548c\u8fdb\u4f4d\u8bbe\u7f6e\u7684\u7b97\u672f\u4efb\u52a1\u3002\u7ed3\u679c\u53d1\u73b0\uff0c\u51cf\u6cd5\u3001\u4e58\u6cd5\u6216\u9664\u6cd5\u6ca1\u6709\u4efb\u4f55\u663e\u8457\u7684\u6027\u80fd\u5dee\u5f02\uff0c\u56e0\u6b64\u7ed3\u679c\u53ea\u5c55\u793a\u4e86\u52a0\u6cd5\u3002<\/section>\n
<\/a><\/section>\n
\u5bf9\u4e8e\u975e\u8fdb\u4f4d\u52a0\u6cd5\uff0c\u6570\u5b57\u4e2a\u6570\u4e3a 3 \u7684\u500d\u6570\u4f1a\u4ea7\u751f\u5b8c\u5168\u76f8\u540c\u7684\u7ed3\u679c\uff0c\u56e0\u4e3a\u50cf 528\u3001491 \u8fd9\u6837\u7684\u6570\u5b57\u65e0\u8bba token \u5316\u65b9\u5411\u5982\u4f55\u90fd\u5177\u6709\u76f8\u540c\u7684 token\u3002<\/section>\n
<\/a><\/section>\n
\u54ea\u79cd tokenization \u65b9\u6cd5\u9002\u5408\u6570\u5b66<\/strong><\/section>\n
<\/section>\n
\u867d\u7136 BPE \u4ecd\u7136\u662f\u4e00\u79cd\u6d41\u884c\u7684 tokenization \u65b9\u6cd5\uff0c\u4f46\u5982\u679c\u4f60\u5fc5\u987b\u4f7f\u7528\u5177\u6709\u6700\u591a 3 \u4f4d\u6570\u7684 tokenizer\uff0c\u8bf7\u786e\u4fdd\u6570\u636e token \u65b9\u5411\u4e3a R2L\u3002<\/section>\n
<\/section>\n
\u5982\u679c\u4f60\u5df2\u7ecf\u6709\u4e00\u4e2a\u7ecf\u8fc7\u8bad\u7ec3\u7684\u6a21\u578b\uff0c\u6570\u636e token \u65b9\u5f0f\u4e3a L2R\uff0c\u90a3\u4e48\u4f60\u53ef\u4ee5\u901a\u8fc7\u4f7f\u7528 R2L \u6765\u83b7\u5f97\u66f4\u597d\u7684\u6570\u5b66\u6027\u80fd\u3002<\/section>\n
<\/section>\n
\u6700\u91cd\u8981\u7684\u662f\uff0c\u5bf9\u4e8e\u7b97\u672f\u8fd0\u7b97\uff0c\u5355\u4f4d\u6570 tokenization \u7684\u6027\u80fd\u660e\u663e\u4f18\u4e8e\u5176\u4ed6\u65b9\u6cd5\u3002<\/section>\n
<\/a><\/section>\n
\u603b\u7ed3\u800c\u8a00\uff0ctokenization \u5bf9\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u7b97\u672f\u6027\u80fd\u6709\u663e\u8457\u5f71\u54cd\u3002\u901a\u8fc7\u4ed4\u7ec6\u9009\u62e9\uff0c\u6211\u4eec\u53ef\u4ee5\u6839\u636e\u95ee\u9898\u7c7b\u578b\u4f18\u5316 tokenization \u7b56\u7565\uff0c\u4ece\u800c\u63d0\u9ad8 LLM \u5728\u6570\u5b66\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u3002<\/section>\n
<\/section>\n
\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/spaces\/huggingface\/number-tokenization-blog<\/sup><\/em><\/section>\n
\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"
\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f […]<\/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":[68],"class_list":["post-37588","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"\n\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86 - \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=37588\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86\",\"datePublished\":\"2025-01-31T06:01:10+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588\"},\"wordCount\":160,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/frc-3434e7be47307131711f3f7048d28d7b.png\",\"keywords\":[\"\u5927\u6a21\u578b\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=37588#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=37588\",\"name\":\"\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/frc-3434e7be47307131711f3f7048d28d7b.png\",\"datePublished\":\"2025-01-31T06:01:10+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=37588\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/frc-3434e7be47307131711f3f7048d28d7b.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/frc-3434e7be47307131711f3f7048d28d7b.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37588#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86\"}]},{\"@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\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"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":"\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86 - \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=37588","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=37588#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=37588"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86","datePublished":"2025-01-31T06:01:10+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=37588"},"wordCount":160,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=37588#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-3434e7be47307131711f3f7048d28d7b.png","keywords":["\u5927\u6a21\u578b"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=37588#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=37588","url":"https:\/\/17aitech.com\/?p=37588","name":"\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=37588#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=37588#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-3434e7be47307131711f3f7048d28d7b.png","datePublished":"2025-01-31T06:01:10+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=37588#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=37588"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=37588#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-3434e7be47307131711f3f7048d28d7b.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-3434e7be47307131711f3f7048d28d7b.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=37588#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u4ece2019\u5e74\u5230\u73b0\u5728\uff0c\u662f\u65f6\u5019\u91cd\u65b0\u5ba1\u89c6Tokenization\u4e86"}]},{"@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":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","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\/37588","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=37588"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/37588\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=37588"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=37588"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=37588"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}