\n
\u6d77\u91cf\u6570\u636e\u5230\u624b\u4e86 \u522b\u7ba1\u600e\u4e48\u6765\u7684\u7b2c\u4e00\u6b65\u5c31\u662f\u5199\u6b63\u5219\u53bb\u6807\u7b7e，\u7528 MinHash \u7b97\u6cd5\u53bb\u91cd\u3001\u7528\u5206\u7c7b\u6a21\u578b\u8fc7\u6ee4\u4f4e\u8d28\u91cf\u548c\u6709\u5bb3\u5185\u5bb9\u3001\u6309\u8bed\u8a00\u5206\u7c7b\u3002<\/p>\n
\u8fd9\u4e2a\u7b97\u6cd5 \u5927\u6982\u5c31\u662f\u628a\u4e00\u672c\u4e66 \u5207\u7247 (Shingling) \u6210\u5927\u91cf\u8fde\u7eed\u8bcd\u7ec4，\u6bcf\u4e2a\u8bcd\u7ec4\u90fd\u8ba1\u7b97\u51fa\u5bf9\u5e94\u7684\u54c8\u5e0c\u6620\u5c04（Hashing） \u914d\u7f6e \u591a\u5c11\u4e2a\u4e2a\u54c8\u5e0c\u51fd\u6570，\u6bcf\u7ec4\u54c8\u5e0c\u8fd0\u7b97\u90fd\u9009\u53d6\u54c8\u5e0c\u7ed3\u679c\u4e2d\u7684\u6700\u5c0f\u503c (Min) \u9009\u51fa\u591a\u5c11Hash\u51fd\u6570 \u5c31\u4f1a\u751f\u6210\u591a\u5c11\u6570\u5b57\u7ec4\u6210\u7684\u6570\u7ec4 \u751f\u6210\u6307\u7eb9 (Signature)<\/p>\n
\u4e24\u7bc7\u6587\u7ae0\u7684 MinHash \u6307\u7eb9\u91cd\u5408\u5ea6\u8d8a\u9ad8，\u5b83\u4eec\u539f\u6587\u7684\u76f8\u4f3c\u5ea6\u5c31\u8d8a\u9ad8\u3002<\/p>\n
\u4eca\u5e74\u8fc7\u8282\u4e0d\u6536\u793c，\u6536\u793c\u53ea\u6536\u8111\u767d\u91d1<\/p>\n
\u7c7b\u4f3c\u5e7f\u544a\u6295\u7684\u5230\u5904\u90fd\u662f \u6240\u4ee5\u4f1a\u5148\u8bad\u7ec3\u4e00\u4e2a\u975e\u5e38\u5c0f\u3001\u901f\u5ea6\u975e\u5e38\u5feb\u7684\u4f20\u7edf\u795e\u7ecf\u7f51\u7edc\u6216\u673a\u5668\u5b66\u4e60\u6a21\u578b<\/p>\n
\u5148\u8fdb\u884c\u4eba\u5de5\u6253\u6807（\u9ad8\u8d28\u91cf\u6587\u7ae0（\u7ef4\u57fa\u767e\u79d1） \u662f1 \u9ec4\u8d4c\u6bd2\u5185\u5bb9\u662f0）\u7136\u540e\u7ed9\u6a21\u578b \u8ba9\u4ed6\u533a\u5206\u4ec0\u4e48\u662f\u597d\u7684<\/p>\n
\u7b49\u771f\u6b63\u53d6\u6e05\u6d17\u6570\u636e\u7684\u65f6\u5019，\u6bcf\u6bb5\u6587\u672c\u90fd\u4f1a\u6253\u4e2a\u5206 \u5206\u4f4e\u7684\u90fd\u76f4\u63a5\u6d41\u653e\u4e86<\/p>\n
\u628a\u4e0a\u4e00\u6b65\u6d77\u91cf\u63d0\u7eaf\u8fc7\u7684\u6570\u636e\u653e\u8fdb\u53bb<\/p>\n
\u7edf\u8ba1\u6587\u672c\u91cc\u54ea\u4e9b\u5b57\/\u8bcd\u7ec4\u6700\u5e38\u8fde\u5728\u4e00\u8d77， \u7ed9\u6700\u5e38\u7528\u7684\u8bcd\u7ec4\u5206\u914d\u4e00\u4e2a\u552f\u4e00\u7684 ID<\/p>\n
\u6bd4\u5982 \u4eba \u540e\u9762\u7ecf\u5e38\u51fa\u73b0 \u5de5<\/p>\n
\u667a \u540e\u9762\u7ecf\u5e38\u51fa\u73b0 \u80fd<\/p>\n
\u5c31\u4f1a\u7ed9\u4eba\u5de5\u5206\u914d ID=1024；\u667a\u80fd\u5206\u914d ID =1025；<\/p>\n
\u4eba\u5de5\u8fd9\u4e2a\u8bcd\u7ec4\u540e\u9762\u7ecf\u5e38\u8ddf\u667a\u80fd\u8fd9\u4e2a\u8bcd\u7ec4 \u7ed9 \u4eba\u5de5\u667a\u80fd \u5206\u914d ID = 8899<\/p>\n
\n
\u7136\u540e\u5f97\u5230\u4e00\u4e2a\u8bcd\u8868\u6587\u4ef6（vocab.json，\u901a\u5e38\u51e0\u4e07\u5230\u5341\u51e0\u4e07\u4e2a\u8bcd\u5143 Token）\u3002<\/p>\n
{
\n "\u4eba": 1230,
\n "\u5de5": 1231,
\n "\u4eba\u5de5": 4521,
\n "\u667a": 1288,
\n "\u80fd":1289,
\n "\u4eba\u5de5\u667a\u80fd": 8899,
\n}<\/p>\n
\u7528 PyTorch \u4ee3\u7801\u5199\u51fa Embedding \u5c42<\/p>\n
\u628a\u8f93\u5165 token \u6574\u6570 ID（\u6bd4\u5982 0~100000 \u8bcd\u8868）\u6620\u5c04\u6210\u56fa\u5b9a\u957f\u5ea6\u5411\u91cf\u3002 \u4f8b\u5982\u8bcd\u8868 10 \u4e07，\u9690\u85cf\u7ef4\u5ea6 512：Embedding \u77e9\u9635\u5f62\u72b6 [100000,512] ，\u8fd9\u5c31\u662f\u7b2c\u4e00\u4e2a\u5de8\u5927\u6743\u91cd\u77e9\u9635\u3002 \u6bcf\u4e2a\u5355\u8bcd ID \u67e5\u8868\u53d6\u51fa\u4e00\u884c\u5411\u91cf，\u4f5c\u4e3a\u8f93\u5165\u3002<\/p>\n
\u8fd8\u6ca1\u6709\u8bad\u7ec3\u6570\u636e\u7684\u65f6\u5019，\u6240\u6709\u6743\u91cd\u4e0d\u80fd\u5168 0！\u5168 0 \u4f1a\u5bfc\u81f4\u6240\u6709\u795e\u7ecf\u5143\u8f93\u51fa\u4e00\u6a21\u4e00\u6837，\u5b66\u4e0d\u5230\u4e1c\u897f\u3002<\/p>\n
\u6240\u4ee5\u5168\u90e8\u521d\u59cb\u5316\u4e3a\u5c0f\u65b9\u5dee\u6b63\u6001\u5206\u5e03\u968f\u673a\u6570（\u6807\u51c6\u505a\u6cd5：N(0,\u03c3\u00b2)，\u03c3 \u53d6\u5f88\u5c0f\u7684\u503c）\u3002<\/p>\n
\u628a\u91cc\u9762\u51e0\u767e\u4ebf\u4e2a\u6743\u91cd\u53c2\u6570\u5168\u90e8\u521d\u59cb\u5316\u4e3a\u7b26\u5408\u6b63\u6001\u5206\u5e03\u7684\u968f\u673a\u6570\u3002<\/p>\n
\u628a\u968f\u673a\u521d\u59cb\u5316\u7684\u6a21\u578b\u67b6\u6784 \u548c \u6d17\u597d\u7684\u6d77\u91cf Token \u6570\u636e\u6d41 \u8f93\u5165<\/p>\n
\u81ea\u56de\u5f52\u8bed\u8a00\u5efa\u6a21，\u7ed9\u524d\u9762 N \u4e2a token，\u9884\u6d4b\u4e0b\u4e00\u4e2a token，\u5c31\u662f\u8bcd\u8bed\u63a5\u9f99\u3002<\/p>\n
\u91cf\u5316\u6a21\u578b\u9884\u6d4b\u548c\u6807\u51c6\u7b54\u6848\u4e4b\u95f4\u5dee\u8ddd\u3002loss \u8d8a\u5927，\u9519\u5f97\u8d8a\u79bb\u8c31\u3002<\/p>\n
\u5f97\u5230 loss \u4e4b\u540e，PyTorch \u81ea\u52a8\u505a\u53cd\u5411\u4f20\u64ad： \u6cbf\u7740\u7f51\u7edc\u4ece\u8f93\u51fa\u5f80\u56de\u8d70，\u8ba1\u7b97\u51e0\u767e\u4ebf\u53c2\u6570\u6bcf\u4e00\u4e2a\u6743\u91cd，\u5bf9\u8fd9\u4e2a\u8bef\u5dee\u8d21\u732e\u591a\u5927（\u68af\u5ea6 gradient）\u3002<\/p>\n
\u8fd9\u90e8\u5206\u4e0d\u4fee\u6539\u6743\u91cd，\u53ea\u8ba1\u7b97\u68af\u5ea6<\/p>\n
\u62ff\u5230\u5168\u90e8\u53c2\u6570\u68af\u5ea6\u540e，AdamW\u4f18\u5316\u5668\u6267\u884c\u4e00\u6b65\u66f4\u65b0： \u6bcf\u4e2a\u6743\u91cd = \u65e7\u6743\u91cd \u2212 \u5b66\u4e60\u7387 \u00d7 \u68af\u5ea6（AdamW \u8fd8\u81ea\u5e26\u6743\u91cd\u8870\u51cf）\u3002 \u5c0f\u5e45\u5ea6\u632a\u52a8\u6bcf\u4e00\u4e2a\u968f\u673a\u521d\u59cb\u7684\u6743\u91cd <\/p>\n
\n
\u663e\u5361\u4f1a\u628a\u6574\u4e2a\u6d41\u7a0b\u8fde\u7eed\u8dd1\u6570\u6708<\/p>\n
\u628a \u57fa\u5ea7\u6a21\u578b + \u51e0\u5341\u4e07\u6761\u9ad8\u8d28\u91cf\u7684\u201c\u4eba\u5de5\u7f16\u5199\u95ee\u7b54\u5bf9\u201d（Q&A）\u8f93\u5165\u8fdb\u53bb<\/p>\n
\u8fd9\u90e8\u5206\u662f\u8ba9 AI \u80fd\u542c\u61c2\u4eba\u8bdd\u7684\u91cd\u8981\u90e8\u5206<\/p>\n
\u8fd9\u91cc\u5fc5\u987b\u6df7\u5165\u5927\u91cf\u201c\u5de5\u5177\u8c03\u7528\u201d\u683c\u5f0f\u7684\u6570\u636e\u3002<\/p>\n
\u6bd4\u5982 \u4f60\u95ee\u4eca\u5929\u5929\u6c14\u600e\u4e48\u6837ai\u8fd9\u65f6\u5019\u4f1a\u7ed9\u4f60\u778e\u7f16\u4e00\u4e2a\u6570\u636e<\/p>\n
\u6240\u4ee5 \u4eba\u7c7b\u95ee\u7b54\u5c31\u4f1a\u7ed9\u51fa\u6807\u51c6json {"action": "call_weather_api", "location": "\u5317\u4eac"} \u8fdb\u884c\u6743\u91cd\u8c03\u8282<\/p>\n
\u628a SFT \u6a21\u578b + \u4eba\u7c7b\u504f\u597d\u6570\u636e\u96c6 \u8f93\u5165\u8fdb\u53bb<\/p>\n
AI\u751f\u6210\u4e24\u4e2a\u7b54\u6848\u7531\u8bad\u7ec3\u4eba\u6765\u9009\u51fa\u4eba\u7c7b\u89c9\u5f97\u597d\u7684\u7b54\u6848<\/p>\n
\n
\u628a SFT \u6a21\u578b\u7684\u6700\u540e\u4e00\u5c42\u201c\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u201d\u7684\u8f93\u51fa\u5934\u780d\u6389，\u6362\u6210\u4e00\u4e2a\u8f93\u51fa\u201c\u6807\u91cf\u5206\u6570\u201d\u7684\u5934\u3002<\/p>\n
\u7528\u76d1\u7763\u5b66\u4e60\u8bad\u7ec3\u5b83，\u8ba9\u5b83\u7ed9\u4eba\u7c7b\u559c\u6b22\u7684\u56de\u7b54\u6253\u9ad8\u5206，\u4e0d\u559c\u6b22\u7684\u6253\u4f4e\u5206\u3002 <\/p>\n
\u8fd9\u6b65\u662f\u4ee3\u66ff\u4eba\u5de5\u6253\u5206，\u65b9\u4fbf\u8fdb\u884c\u4e0b\u4e00\u6b65\u5f3a\u5316\u5b66\u4e60<\/p>\n
\u628a SFT \u6a21\u578b + RM \u6a21\u578b + \u65e0\u7b54\u6848\u7684\u63d0\u793a\u8bcd\u96c6 \u8f93\u5165\u8fdb\u53bb<\/p>\n
\u8ba9 SFT \u6a21\u578b\u5bf9\u63d0\u793a\u8bcd\u751f\u6210\u56de\u7b54，RM \u7ed9\u56de\u7b54\u6253\u5206\u3002<\/p>\n
SFT \u6a21\u578b\u6839\u636e\u5206\u6570，\u5229\u7528 PPO \u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u8c03\u6574\u81ea\u5df1\u7684\u53c2\u6570，\u529b\u6c42\u4ee5\u540e\u751f\u6210\u80fd\u62ff\u66f4\u9ad8\u5206\u7684\u56de\u7b54，\u540c\u65f6\u9650\u5236\u5b83\u4e0d\u80fd\u504f\u79bb SFT \u6a21\u578b\u592a\u8fdc<\/p>\n
\u5bf9\u9f50\u6a21\u578b (Aligned \/ Chat Model） \u5c31\u662f\u5b83\u6700\u7ec8\u4ea4\u4ed8\u7684\u6743\u91cd<\/p>\n
\u5de5\u7a0b\u5e08\u4f7f\u7528 vLLM \u7b49\u63a8\u7406\u6846\u67b6，\u5c06\u6743\u91cd\u52a0\u8f7d\u5230 GPU \u663e\u5b58\u3002\u5b9e\u73b0 KV Cache（\u7f13\u5b58\u8ba1\u7b97\u8fc7\u7684\u4e0a\u4e0b\u6587，\u6781\u5927\u5730\u52a0\u901f\u63a8\u65ad）\u548c Continuous Batching（\u9ad8\u5e76\u53d1\u5904\u7406\u8bf7\u6c42）\u3002\u6700\u540e\u5957\u4e0a\u4e00\u4e2a HTTP \u63a5\u53e3（\u5982 FastAPI）\u3002<\/p>\n
\u8fd9\u90e8\u5206\u4fdd\u8bc1\u9ad8\u541e\u5410\u91cf<\/p>\n
\u8fd9\u90e8\u5206\u662f\u670d\u52a1\u5546\u8981\u5b8c\u6210\u7684\u5e95\u5c42\u90e8\u7f72\u5de5\u4f5c <\/p>\n
\u6211\u4eec\u7528API Key \u8bf7\u6c42\u7684\u6a21\u578b\u4e00\u822c\u5c31\u662f\u5bf9\u9f50\u6a21\u578b（\u5bf9\u9f50\u597d\u7684\u5bf9\u8bdd\u6a21\u578b）<\/p>\n
\u5c11\u6570\u4f1a\u63d0\u4f9b\u57fa\u5ea7\u6a21\u578b<\/p>\n
\n
\n
\n","protected":false},"excerpt":{"rendered":"
\u539f\u59cb\u6570\u636e\u6e05\u6d17（Data Cleaning）\u6d77\u91cf\u6570\u636e\u5230\u624b\u4e86 \u522b\u7ba1\u600e\u4e48\u6765\u7684\u7b2c\u4e00\u6b65\u5c31\u662f\u5199\u6b63\u5219\u53bb\u6807\u7b7e，\u7528 MinHash \u7b97\u6cd5\u53bb\u91cd\u3001\u7528\u5206\u7c7b\u6a21\u578b\u8fc7\u6ee4\u4f4e\u8d28\u91cf\u548c\u6709\u5bb3\u5185\u5bb9\u3001\u6309\u8bed\u8a00\u5206\u7c7b\u3002MinHash\u8fd9\u4e2a\u7b97\u6cd5 \u5927\u6982\u5c31\u662f\u628a\u4e00\u672c\u4e66 \u5207\u7247 (Shingling) \u6210\u5927\u91cf\u8fde\u7eed\u8bcd\u7ec4，\u6bcf\u4e2a\u8bcd\u7ec4\u90fd\u8ba1\u7b97\u51fa\u5bf9\u5e94\u7684\u54c8\u5e0c\u6620\u5c04（Hashing） \u914d\u7f6e \u591a\u5c11\u4e2a\u4e2a\u54c8\u5e0c\u51fd\u6570，\u6bcf\u7ec4\u54c8\u5e0c\u8fd0\u7b97\u90fd\u9009\u53d6\u54c8\u5e0c\u7ed3\u679c\u4e2d\u7684\u6700\u5c0f\u503c (Min) <\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[66],"topic":[],"class_list":["post-100653","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai"],"yoast_head":"\n