\u4e13\u4e1a\u6027：\u9700\u8981\u6a21\u578b\u6df1\u5ea6\u638c\u63e1\u5782\u76f4\u9886\u57df\u77e5\u8bc6<\/p>\n<\/li>\n
\u901a\u7528\u6027：\u5e0c\u671b\u4fdd\u7559\u57fa\u7840\u5bf9\u8bdd\u548c\u5e38\u8bc6\u80fd\u529b<\/p>\n<\/li>\n
\u7a33\u5b9a\u6027：\u8981\u6c42\u8f93\u51fa\u683c\u5f0f\u89c4\u8303\u3001\u8868\u8fbe\u591a\u6837<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
\u4e13\u4e1a\u7cbe\u51c6，\u57fa\u7840\u5d29\u6e83：\u6a21\u578b\u80fd\u51c6\u786e\u56de\u7b54"\u52b3\u52a8\u5408\u540c\u89e3\u9664\u6761\u4ef6"，\u5374\u65e0\u6cd5\u6b63\u5e38\u56de\u5e94"\u4f60\u597d"，\u751a\u81f3\u5c06\u5929\u6c14\u67e5\u8be2\u56de\u7b54\u6210\u6cd5\u5f8b\u6761\u6587<\/p>\n<\/li>\n
- \n
\u8bad\u7ec3\u65f6\u5b8c\u7f8e，\u4e0a\u7ebf\u540e\u590d\u8bfb：\u8bad\u7ec3loss\u4e00\u8def\u4e0b\u964d，\u4e0a\u7ebf\u540e\u5374\u53cd\u590d\u4f7f\u7528\u76f8\u540c\u53e5\u5f0f，\u751f\u6210\u5185\u5bb9\u7f3a\u4e4f\u591a\u6837\u6027，\u50cf\u4e00\u53f0\u590d\u8bfb\u673a<\/p>\n<\/li>\n
- \n
\u6570\u636e\u8d28\u91cf\u654f\u611f：\u4e00\u6761\u9519\u8bef\u6807\u6ce8\u6216\u683c\u5f0f\u6df7\u4e71\u7684\u6570\u636e，\u53ef\u80fd\u5bfc\u81f4\u6574\u4e2a\u6a21\u578b\u80fd\u529b\u5d29\u584c<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
\u964d\u4f4e\u5b66\u4e60\u7387？\u2192 \u4e13\u4e1a\u80fd\u529b\u63d0\u5347\u53d7\u9650<\/p>\n<\/li>\n
- \n
\u589e\u52a0\u901a\u7528\u8bed\u6599？\u2192 \u8bad\u7ec3\u65f6\u95f4\u7ffb\u500d，\u4e13\u4e1a\u6027\u7a00\u91ca<\/p>\n<\/li>\n
- \n
\u8c03\u6574\u8bad\u7ec3\u8f6e\u6b21？\u2192 \u4eca\u5929\u597d\u660e\u5929\u574f，\u6548\u679c\u4e0d\u7a33\u5b9a<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
\u751f\u6210\u5185\u5bb9\u6a21\u677f\u5316\u3001\u7f3a\u4e4f\u591a\u6837\u6027<\/p>\n<\/li>\n
- \n
\u540c\u4e49\u6539\u5199\u3001\u5b57\u6bb5\u987a\u5e8f\u53d8\u5316\u7b49\u5c0f\u6539\u52a8\u5bfc\u81f4\u8f93\u51fa\u6f02\u79fb<\/p>\n<\/li>\n
- \n
\u4fe1\u606f\u5bc6\u5ea6\u4e0b\u964d，\u56de\u7b54\u53d8\u957f\u4f46\u5185\u5bb9\u7a7a\u6d1e<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
\u539f\u6709\u901a\u7528\u80fd\u529b\u88ab\u6324\u538b<\/p>\n<\/li>\n
- \n
\u8bad\u7ec3\u8fc7\u7a0b\u4e0d\u7a33\u5b9a，loss\u66f2\u7ebf\u6ce2\u52a8\u5267\u70c8<\/p>\n<\/li>\n
- \n
\u6a21\u578b\u53d8\u5f97"\u504f\u79d1"，\u5931\u53bb\u57fa\u7840\u5bf9\u8bdd\u80fd\u529b<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
\u5355\u5361\u8bad\u7ec3：\u65e0\u9700\u5b89\u88c5deepspeed，\u8282\u7701\u914d\u7f6e\u65f6\u95f4<\/p>\n<\/li>\n
- \n
\u663e\u5b58<24GB：\u52a1\u5fc5\u5f00\u542fLoRA，\u5426\u5219\u53ef\u80fdOOM<\/p>\n<\/li>\n
- \n
\u68af\u5ea6\u68c0\u67e5\u70b9：\u663e\u5b58\u7d27\u5f20\u65f6\u6dfb\u52a0–enable_gradit_checkpoing 'true'\u53c2\u6570<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
–token_batch：SFT\u8bad\u7ec3\u7684\u6838\u5fc3\u53c2\u6570，\u63a7\u5236\u6bcf\u6b21\u53cd\u5411\u4f20\u64ad\u7684token\u6570\u91cf\u3002\u503c\u8d8a\u5c0f，token loss\u8ba1\u7b97\u8d8a\u7cbe\u786e，NLIRG\u6548\u679c\u8d8a\u597d，\u4f46\u8bad\u7ec3\u901f\u5ea6\u8d8a\u6162\u3002\u5b98\u65b9\u63a8\u8350\u9ed8\u8ba4\u503c10<\/p>\n<\/li>\n
- \n
–use_NLIRG：\u9ed8\u8ba4\u4e3atrue，\u505a\u5bf9\u7167\u5b9e\u9a8c\u65f6\u52a1\u5fc5\u663e\u5f0f\u8bbe\u7f6e\u4e3afalse<\/p>\n<\/li>\n
- \n
–lora_target_modules：LoRA\u9002\u914d\u7684\u6a21\u5757，\u5bf9Qwen\u7cfb\u5217\u6a21\u578b\u4f7f\u7528\u4e0a\u8ff0\u914d\u7f6e<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
DeepSpeed\u914d\u7f6e：\u5b98\u65b9\u652f\u6301Stage3，\u8282\u7701\u663e\u5b58<\/p>\n<\/li>\n
- \n
\u6587\u672c\u6253\u5305：\u901a\u8fc7–pack_length\u5c06\u77ed\u6587\u672c\u5408\u5e76，\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387<\/p>\n<\/li>\n
- \n
\u4e3b\u8282\u70b9\u7aef\u53e3：–master_port\u9700\u786e\u4fdd\u672a\u88ab\u5360\u7528<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
\u91cd\u590d\u7387\u8ba1\u7b97：nltk.translate.bleu_score\u6216\u81ea\u5b9a\u4e49n-gram\u91cd\u590d\u7387<\/p>\n<\/li>\n
- \n
\u683c\u5f0f\u9a8c\u8bc1：\u7f16\u5199\u6b63\u5219\u8868\u8fbe\u5f0f\u9a8c\u8bc1\u8f93\u51fa\u683c\u5f0f<\/p>\n<\/li>\n
- \n
\u4eba\u5de5\u8bc4\u4f30：\u4f7f\u7528\u6807\u6ce8\u5e73\u53f0，\u8ba93\u4f4d\u6807\u6ce8\u5458\u5bf9\u8f93\u51fa\u8d28\u91cf\u6253\u5206(1-5\u5206)<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
loss\u66f2\u7ebf：\u89c2\u5bdfNLIRG\u662f\u5426\u4f7f\u8bad\u7ec3\u66f4\u7a33\u5b9a<\/p>\n<\/li>\n
- \n
\u68af\u5ea6\u5206\u5e03：\u9a8c\u8bc1\u4e0d\u540closs\u533a\u95f4token\u7684\u68af\u5ea6\u53d8\u5316<\/p>\n<\/li>\n
- \n
\u9a8c\u8bc1\u96c6\u6307\u6807：\u53ca\u65f6\u53d1\u73b0\u8fc7\u62df\u5408\u8ff9\u8c61<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
\u63d0\u524d\u53d1\u73b0\u810f\u6570\u636e：\u4f4e\u5206\u8bed\u6599\u5f80\u5f80\u662f\u683c\u5f0f\u9519\u8bef\u6216\u5185\u5bb9\u9519\u8bef，\u53ef\u4f18\u5148\u4fee\u590d<\/p>\n<\/li>\n
- \n
\u4f18\u5316\u8bad\u7ec3\u987a\u5e8f：\u975e\u7b80\u5355"\u7531\u6613\u5230\u96be"，\u800c\u662f\u667a\u80fd\u6df7\u5408，\u8bad\u7ec3\u66f4\u7a33\u5b9a<\/p>\n<\/li>\n
- \n
\u8282\u7701\u8bad\u7ec3\u8d44\u6e90：\u8fc7\u6ee4\u6389\u660e\u663e\u9519\u8bef\u7684\u6837\u672c，\u51cf\u5c11\u65e0\u6548\u8bad\u7ec3<\/p>\n<\/p>\n<\/li>\n
\u6b63\u5982\u5b98\u65b9\u6587\u6863\u6307\u51fa：<\/p>\n
"\u518d\u8bad\u7ec3\u4e4b\u524d\u6392\u67e5\u8bed\u6599\u95ee\u9898，\u68c0\u67e5\u5f97\u5206\u8f83\u4f4e\u7684\u8bed\u6599\u662f\u5426\u7f16\u5199\u9519\u8bef\u3002\u5e73\u8861\u8bed\u6599\u96be\u6613\u987a\u5e8f，\u9010\u6b65\u8bad\u7ec3\u3002\u6ce8\u610f：\u7b80\u5355\u7684\u5c06\u8bed\u6599\u4ece\u6613\u5230\u96be\u8fdb\u884c\u8bad\u7ec3，\u5b9e\u9645\u6d4b\u8bd5\u6548\u679c\u5e76\u4e0d\u5982\u6df7\u5408\u8fdb\u884c\u3002"<\/p>\n<\/p>\n
7.2 \u8d44\u6e90\u4f18\u5316\u7b56\u7565（\u9762\u5411\u4e0d\u540c\u786c\u4ef6\u73af\u5883）<\/h4>\n<\/p>\n
\u5355\u5361<24GB\u663e\u5b58\u914d\u7f6e：<\/p>\n
–use_lora 'true' \\\\
\n–enable_gradit_checkpoing 'true' \\\\
\n–batch_size 1 \\\\
\n–token_batch 5 \\\\
\n–model_path_to_load Qwen\/Qwen3-1.5B <\/p>\n\u5355\u536124GB+\u663e\u5b58\u914d\u7f6e：<\/p>\n
–use_lora 'true' \\\\
\n–batch_size 1 \\\\
\n–token_batch 10 \\\\
\n–model_path_to_load Qwen\/Qwen3-7B <\/p>\n\u591a\u5361\u8bad\u7ec3(4xA100)\u914d\u7f6e：<\/p>\n
–use_deepspeed 'true' \\\\
\n–batch_size 2 \\\\
\n–pack_length 2048 \\\\
\n–model_path_to_load Qwen\/Qwen3-8B <\/p>\n\u5173\u952e\u539f\u5219：<\/p>\n
- \n
- \n
token_batch\u503c\u8d8a\u5c0f，NLIRG\u6548\u679c\u8d8a\u597d，\u4f46\u901f\u5ea6\u8d8a\u6162<\/p>\n<\/li>\n
- \n
\u663e\u5b58\u4e0d\u8db3\u65f6，\u4f18\u5148\u4fdd\u8bc1token_batch\u4e0d\u4f4e\u4e8e5，\u518d\u8003\u8651\u964d\u4f4ebatch_size<\/p>\n<\/li>\n
- \n
LoRA\u79e9(rank)\u8bbe\u7f6e：\u5782\u76f4\u9886\u57df\u5efa\u8bae8-16，\u901a\u7528\u9886\u57df\u5efa\u8bae4-8<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
- \n
\u5148\u9a8c\u8bc1\u518d\u6295\u5165：\u7528\u4f60\u7684\u5782\u76f4\u6570\u636e\u8dd1\u4e00\u6b21\u6700\u7b80\u5bf9\u7167\u5b9e\u9a8c（\u5f00\u542f\/\u5173\u95edNLIRG）<\/p>\n<\/li>\n
- \n
\u5efa\u7acb\u8bc4\u4f30\u4f53\u7cfb：\u521b\u5efa\u901a\u7528\u80fd\u529b\u56de\u5f52\u6d4b\u8bd5\u96c6，\u76d1\u63a7\u6a21\u578b\u6574\u4f53\u5065\u5eb7\u5ea6<\/p>\n<\/li>\n
- \n
\u9010\u6b65\u5e94\u7528\u9ad8\u7ea7\u529f\u80fd：\u4ece\u57fa\u7840NLIRG\u5230\u8bed\u6599\u6392\u5e8f，\u518d\u5230\u8d44\u6e90\u4f18\u5316<\/p>\n<\/li>\n
- \n
\u8bb0\u5f55\u8fed\u4ee3\u8fc7\u7a0b：\u4fdd\u5b58\u6bcf\u6b21\u5b9e\u9a8c\u7684\u914d\u7f6e\u548c\u7ed3\u679c，\u5f62\u6210\u56e2\u961f\u77e5\u8bc6\u5e93<\/p>\n<\/p>\n<\/li>\n
\u6b63\u5982Y-Trainer\u5b98\u65b9\u6587\u6863\u6240\u5f3a\u8c03：<\/p>\n
"Y-Trainer\u7b97\u6cd5\u901a\u8fc7\u6a21\u578b\u5185\u90e8\u4fe1\u53f7，\u53ef\u4ee5\u5bf9\u8bed\u6599\u8fdb\u884c\u8d28\u91cf\u8bc4\u5206，\u63d0\u65e9\u6392\u67e5\u9519\u8bef\u3002\u4f20\u7edf\u7684SFT\u901a\u5e38\u9700\u8981\u6df7\u5408\u4e00\u5b9a\u6bd4\u4f8b\u901a\u7528\u8bed\u6599，\u9632\u6b62\u6a21\u578b\u80fd\u529b\u9000\u5316，Y-Trainer\u7b97\u6cd5\u53ef\u5728\u53ea\u4f7f\u7528\u5782\u76f4\u9886\u57df\u8bed\u6599\u7684\u60c5\u51b5\u4e0b\u8bad\u7ec3，\u5e76\u53d6\u5f97\u66f4\u597d\u7684\u6548\u679c\u3002"<\/p>\n<\/p>\n
\u544a\u522b\u6570\u636e\u5e73\u8861\u7684\u5669\u68a6，\u91cd\u83b7\u8bad\u7ec3\u638c\u63a7\u6743\u3002\u5728\u8fd9\u4e2a\u5927\u6a21\u578b\u5e94\u7528\u7206\u53d1\u7684\u65f6\u4ee3，\u638c\u63e1\u9ad8\u6548\u7684\u8bad\u7ec3\u65b9\u6cd5\u6bd4\u76f2\u76ee\u5806\u780c\u7b97\u529b\u66f4\u91cd\u8981\u3002\u4f60\u51c6\u5907\u597d\u544a\u522b"\u590d\u8bfb\u673a"\u6a21\u578b\u4e86\u5417？<\/p>\n<\/p>\n
\u8d44\u6e90\u63a8\u8350<\/h3>\n<\/p>\n
- \n
- \n
\u5b98\u65b9\u6587\u6863：\u6700\u6743\u5a01\u7684\u4f7f\u7528\u6307\u5357<\/p>\n<\/li>\n
- \n
\u6838\u5fc3\u7b97\u6cd5\u8be6\u89e3：NLIRG\u6570\u5b66\u539f\u7406<\/p>\n<\/li>\n
- \n
\u5feb\u901f\u5f00\u59cb\u793a\u4f8b：5\u5206\u949f\u4e0a\u624b<\/p>\n<\/li>\n
- \n
GitHub\u4ed3\u5e93：\u6e90\u4ee3\u7801\u4e0eissue\u8ba8\u8bba<\/p>\n<\/p>\n
- <\/ul>\n<\/li>\n<\/ul>\n
\u4e92\u52a8\u8ba8\u8bba<\/h3>\n<\/p>\n
\u4f60\u5728\u5782\u76f4SFT\u8bad\u7ec3\u4e2d\u9047\u5230\u8fc7\u54ea\u4e9b\u95ee\u9898？NLIRG\u7b97\u6cd5\u5bf9\u4f60\u6709\u54ea\u4e9b\u542f\u53d1？\u6b22\u8fce\u5728\u8bc4\u8bba\u533a\u5206\u4eab\u4f60\u7684\u7ecf\u9a8c\u548c\u7591\u95ee，\u4e00\u8d77\u63a2\u8ba8\u5927\u6a21\u578b\u8bad\u7ec3\u7684\u6700\u4f73\u5b9e\u8df5！<\/p>\n<\/p>\n
#LLM #\u6a21\u578b\u8bad\u7ec3 #SFT #AI\u5de5\u7a0b\u5316 #YTrainer<\/p>\n","protected":false},"excerpt":{"rendered":"
\u6458\u8981：\u672c\u6587\u6df1\u5165\u5256\u6790\u5782\u76f4\u9886\u57dfSFT\u8bad\u7ec3\u4e2d\u7684\u4e09\u5927\u987d\u75be：\u6a21\u578b\u590d\u8bfb\u3001\u707e\u96be\u6027\u9057\u5fd8\u548c\u8fc7\u62df\u5408\u3002\u901a\u8fc7Y-Trainer\u6846\u67b6\u53ca\u5176\u6838\u5fc3NLIRG\u7b97\u6cd5，\u6784\u5efa\u4e00\u5957\u53ef\u590d\u73b0\u3001\u53ef\u5bf9\u7167\u7684\u8bad\u7ec3\u6d41\u7a0b\u3002\u6587\u672b\u63d0\u4f9b\u5b8c\u6574\u4ee3\u7801\u793a\u4f8b\u3001\u6548\u679c\u9a8c\u8bc1\u65b9\u6cd5\u53ca\u5b9e\u7528\u6280\u5de7，\u52a9\u4f60\u63d0\u5347\u5782\u76f4\u9886\u57df\u6a21\u578b\u8bad\u7ec3\u6548\u679c\u3002 \u4e00\u3001\u75db\u70b9\u76f4\u51fb：\u5782\u76f4SFT\u8bad\u7ec3\u4e2d\u7684\\”\u4e0d\u53ef\u80fd\u4e09\u89d2\\” \u5728\u5782\u76f4\u9886\u57df\u5927\u6a21\u578b\u5fae\u8c03\u8fc7\u7a0b\u4e2d，\u6211\u4eec\u5e38\u5e38\u9762\u4e34\u4e00\u4e2a\u5c34\u5c2c\u7684\\”\u4e0d\u53ef\u80fd\u4e09\u89d2\\”： \u4e13\u4e1a\u6027：\u9700<\/p>\n","protected":false},"author":2,"featured_media":66334,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[587,175,50,207,86,427,51],"topic":[],"class_list":["post-66339","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-587","tag-chatgpt","tag-50","tag-207","tag-86","tag-427","tag-51"],"yoast_head":"\n
\u4e09\u62db\u89e3\u51b3\u590d\u8bfb\u3001\u8fc7\u62df\u5408\u548c\u5931\u5fc6\u95ee\u9898\uff0cY-Trainer\u5b9e\u6218\u6307\u5357 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/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:\/\/www.wsisp.com\/helps\/66339.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u4e09\u62db\u89e3\u51b3\u590d\u8bfb\u3001\u8fc7\u62df\u5408\u548c\u5931\u5fc6\u95ee\u9898\uff0cY-Trainer\u5b9e\u6218\u6307\u5357 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6458\u8981：\u672c\u6587\u6df1\u5165\u5256\u6790\u5782\u76f4\u9886\u57dfSFT\u8bad\u7ec3\u4e2d\u7684\u4e09\u5927\u987d\u75be：\u6a21\u578b\u590d\u8bfb\u3001\u707e\u96be\u6027\u9057\u5fd8\u548c\u8fc7\u62df\u5408\u3002\u901a\u8fc7Y-Trainer\u6846\u67b6\u53ca\u5176\u6838\u5fc3NLIRG\u7b97\u6cd5，\u6784\u5efa\u4e00\u5957\u53ef\u590d\u73b0\u3001\u53ef\u5bf9\u7167\u7684\u8bad\u7ec3\u6d41\u7a0b\u3002\u6587\u672b\u63d0\u4f9b\u5b8c\u6574\u4ee3\u7801\u793a\u4f8b\u3001\u6548\u679c\u9a8c\u8bc1\u65b9\u6cd5\u53ca\u5b9e\u7528\u6280\u5de7，\u52a9\u4f60\u63d0\u5347\u5782\u76f4\u9886\u57df\u6a21\u578b\u8bad\u7ec3\u6548\u679c\u3002 \u4e00\u3001\u75db\u70b9\u76f4\u51fb：\u5782\u76f4SFT\u8bad\u7ec3\u4e2d\u7684"\u4e0d\u53ef\u80fd\u4e09\u89d2" \u5728\u5782\u76f4\u9886\u57df\u5927\u6a21\u578b\u5fae\u8c03\u8fc7\u7a0b\u4e2d，\u6211\u4eec\u5e38\u5e38\u9762\u4e34\u4e00\u4e2a\u5c34\u5c2c\u7684"\u4e0d\u53ef\u80fd\u4e09\u89d2"： \u4e13\u4e1a\u6027：\u9700\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/66339.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-01-26T10:05:04+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260126100501-69773c4d53c23.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"5 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/66339.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/66339.html\",\"name\":\"\u4e09\u62db\u89e3\u51b3\u590d\u8bfb\u3001\u8fc7\u62df\u5408\u548c\u5931\u5fc6\u95ee\u9898\uff0cY-Trainer\u5b9e\u6218\u6307\u5357 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-01-26T10:05:04+00:00\",\"dateModified\":\"2026-01-26T10:05:04+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/66339.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/66339.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/66339.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4e09\u62db\u89e3\u51b3\u590d\u8bfb\u3001\u8fc7\u62df\u5408\u548c\u5931\u5fc6\u95ee\u9898\uff0cY-Trainer\u5b9e\u6218\u6307\u5357\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u4e09\u62db\u89e3\u51b3\u590d\u8bfb\u3001\u8fc7\u62df\u5408\u548c\u5931\u5fc6\u95ee\u9898\uff0cY-Trainer\u5b9e\u6218\u6307\u5357 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","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:\/\/www.wsisp.com\/helps\/66339.html","og_locale":"zh_CN","og_type":"article","og_title":"\u4e09\u62db\u89e3\u51b3\u590d\u8bfb\u3001\u8fc7\u62df\u5408\u548c\u5931\u5fc6\u95ee\u9898\uff0cY-Trainer\u5b9e\u6218\u6307\u5357 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6458\u8981：\u672c\u6587\u6df1\u5165\u5256\u6790\u5782\u76f4\u9886\u57dfSFT\u8bad\u7ec3\u4e2d\u7684\u4e09\u5927\u987d\u75be：\u6a21\u578b\u590d\u8bfb\u3001\u707e\u96be\u6027\u9057\u5fd8\u548c\u8fc7\u62df\u5408\u3002\u901a\u8fc7Y-Trainer\u6846\u67b6\u53ca\u5176\u6838\u5fc3NLIRG\u7b97\u6cd5，\u6784\u5efa\u4e00\u5957\u53ef\u590d\u73b0\u3001\u53ef\u5bf9\u7167\u7684\u8bad\u7ec3\u6d41\u7a0b\u3002\u6587\u672b\u63d0\u4f9b\u5b8c\u6574\u4ee3\u7801\u793a\u4f8b\u3001\u6548\u679c\u9a8c\u8bc1\u65b9\u6cd5\u53ca\u5b9e\u7528\u6280\u5de7，\u52a9\u4f60\u63d0\u5347\u5782\u76f4\u9886\u57df\u6a21\u578b\u8bad\u7ec3\u6548\u679c\u3002 \u4e00\u3001\u75db\u70b9\u76f4\u51fb：\u5782\u76f4SFT\u8bad\u7ec3\u4e2d\u7684\"\u4e0d\u53ef\u80fd\u4e09\u89d2\" \u5728\u5782\u76f4\u9886\u57df\u5927\u6a21\u578b\u5fae\u8c03\u8fc7\u7a0b\u4e2d，\u6211\u4eec\u5e38\u5e38\u9762\u4e34\u4e00\u4e2a\u5c34\u5c2c\u7684\"\u4e0d\u53ef\u80fd\u4e09\u89d2\"： \u4e13\u4e1a\u6027：\u9700","og_url":"https:\/\/www.wsisp.com\/helps\/66339.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-01-26T10:05:04+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260126100501-69773c4d53c23.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"5 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/66339.html","url":"https:\/\/www.wsisp.com\/helps\/66339.html","name":"\u4e09\u62db\u89e3\u51b3\u590d\u8bfb\u3001\u8fc7\u62df\u5408\u548c\u5931\u5fc6\u95ee\u9898\uff0cY-Trainer\u5b9e\u6218\u6307\u5357 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-01-26T10:05:04+00:00","dateModified":"2026-01-26T10:05:04+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/66339.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/66339.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/66339.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u4e09\u62db\u89e3\u51b3\u590d\u8bfb\u3001\u8fc7\u62df\u5408\u548c\u5931\u5fc6\u95ee\u9898\uff0cY-Trainer\u5b9e\u6218\u6307\u5357"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/66339","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=66339"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/66339\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/66334"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=66339"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=66339"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=66339"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=66339"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
- \n
\u516b\u3001\u9002\u7528\u573a\u666f\u5206\u6790：\u4ec0\u4e48\u60c5\u51b5\u4e0b\u5e94\u8be5\u9009\u62e9Y-Trainer？<\/h3>\n<\/p>\n
8.1 \u9002\u5408\u573a\u666f（\u6548\u679c\u663e\u8457）<\/h4>\n<\/p>\n
\u2705 \u5f3a\u5782\u76f4\u9886\u57df\u6a21\u578b：\u6cd5\u5f8b\u3001\u533b\u7597\u3001\u91d1\u878d\u7b49\u9700\u8981\u4e13\u4e1a\u6027，\u540c\u65f6\u4fdd\u7559\u57fa\u7840\u80fd\u529b\u7684\u573a\u666f<\/p>\n
\u2705 \u6570\u636e\u8d28\u91cf\u53c2\u5dee\u4e0d\u9f50：NLIRG\u7684\u9ad8loss\u5f52\u96f6\u673a\u5236\u80fd\u6709\u6548\u9694\u79bb\u566a\u58f0\u6837\u672c<\/p>\n
\u2705 \u8d44\u6e90\u6709\u9650\u73af\u5883：\u5355\u5361\/\u5c0f\u96c6\u7fa4\u8bad\u7ec3，\u9700\u8981LoRA\u548c\u663e\u5b58\u4f18\u5316\u652f\u6301<\/p>\n
\u2705 \u5feb\u901f\u8fed\u4ee3\u9700\u6c42：\u65e0\u9700\u53cd\u590d\u8c03\u6574\u6570\u636e\u6bd4\u4f8b，\u8282\u7701\u8c03\u53c2\u65f6\u95f4<\/p>\n<\/p>\n
8.2 \u4e0d\u9002\u5408\u573a\u666f（\u6548\u679c\u6709\u9650）<\/h4>\n<\/p>\n
\u274c \u901a\u7528\u804a\u5929\u673a\u5668\u4eba\u8bad\u7ec3：\u672c\u8eab\u5c31\u662f\u6df7\u5408\u8bed\u6599，NLIRG\u4f18\u52bf\u4e0d\u660e\u663e<\/p>\n
\u274c \u6781\u5927\u89c4\u6a21\u96c6\u7fa4\u8bad\u7ec3（>64\u5361）：\u53ef\u80fd\u9700\u8981\u5b9a\u5236\u5206\u5e03\u5f0f\u7b56\u7565<\/p>\n
\u274c \u6570\u636e\u5df2\u9ad8\u5ea6\u4f18\u5316：\u5982\u679c\u6570\u636e\u5df2\u7ecf\u7ecf\u8fc7\u4e25\u683c\u6e05\u6d17\u548c\u5e73\u8861，\u6536\u76ca\u53ef\u80fd\u6709\u9650<\/p>\n<\/p>\n
\u7ed3\u8bed<\/h3>\n<\/p>\n
\u5782\u76f4SFT\u8bad\u7ec3\u7684\u56f0\u5883，\u672c\u8d28\u662f\u68af\u5ea6\u5206\u914d\u7684\u7c97\u653e\u7ba1\u7406\u3002Y-Trainer\u901a\u8fc7NLIRG\u7b97\u6cd5，\u5c06\u8bad\u7ec3\u5f3a\u5ea6\u63a7\u5236\u4ece"\u6837\u672c\u7ea7"\u63a8\u8fdb\u5230"token\u7ea7"，\u5c31\u50cf\u4ece"\u5927\u6c34\u6f2b\u704c"\u5230"\u7cbe\u51c6\u6ef4\u704c"\u7684\u519c\u4e1a\u53d8\u9769\u3002<\/p>\n<\/p>\n
\u6211\u7684\u5efa\u8bae\u884c\u52a8\u8def\u5f84：<\/p>\n
- \n
- \n
\u4e03\u3001\u9ad8\u7ea7\u6280\u5de7：\u63d0\u5347\u8bad\u7ec3\u6548\u679c\u7684\u5b9e\u7528\u7b56\u7565<\/h3>\n<\/p>\n
7.1 \u8bed\u6599\u8d28\u91cf\u9884\u7b5b（\u8bad\u7ec3\u524d\u6392\u96f7）<\/h4>\n<\/p>\n
\u5728\u8bad\u7ec3\u524d，\u4f7f\u7528Y-Trainer\u7684\u8bed\u6599\u6392\u5e8f\u5de5\u5177\u8bc4\u4f30\u6570\u636e\u8d28\u91cf：<\/p>\n
python -m training_code.utils.schedule.sort \\\\
\n –data_path raw_data.json \\\\
\n –output_path sorted_data.json \\\\
\n –model_path Qwen\/Qwen3-8B \\\\
\n –mode "similarity_rank" # \u57fa\u4e8e\u66f2\u7ebf\u76f8\u4f3c\u5ea6\u7684\u6392\u5e8f\u7b56\u7565 <\/p>\n\u7b97\u6cd5\u539f\u7406：\u901a\u8fc7\u5206\u6790\u6a21\u578b\u5bf9\u6bcf\u6761\u8bed\u6599\u7684\u54cd\u5e94（\u635f\u5931\u503c\u548c\u71b5\u7684\u53d8\u5316），\u8ba1\u7b97\u8bed\u6599\u96be\u5ea6\u8bc4\u5206<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260126100502-69773c4ef1d39.png\")
<\/p>\n<\/p>\n\u5b9e\u7528\u4ef7\u503c：<\/p>\n
- \n
6.3 \u8bad\u7ec3\u8fc7\u7a0b\u53ef\u89c6\u5316（TensorBoard\u76d1\u63a7）<\/h4>\n
# \u8bad\u7ec3\u65f6\u542f\u7528TensorBoard
\npython -m training_code.start_training \\\\
\n –use_tensorboard 'true' \\\\
\n –tensorboard_path .\/logs \\\\
\n …\u5176\u4ed6\u53c2\u6570…<\/p>\n# \u542f\u52a8\u53ef\u89c6\u5316（\u65b0\u7ec8\u7aef）
\ntensorboard –logdir=.\/logs –port=6006 <\/p>\n![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260126100502-69773c4e721dc.png\")
<\/p>\n<\/p>\n\u5173\u952e\u76d1\u63a7\u6307\u6807：<\/p>\n
- \n
- \n
\u516d\u3001\u6548\u679c\u9a8c\u8bc1：\u4e0d\u8981\u53ea\u770bloss，\u5efa\u7acb\u79d1\u5b66\u8bc4\u4f30\u4f53\u7cfb<\/h3>\n<\/p>\n
6.1 \u5bf9\u7167\u5b9e\u9a8c\u8bbe\u8ba1<\/h4>\n
# Baseline（\u5173\u95edNLIRG）
\npython -m training_code.start_training \\\\
\n –use_NLIRG 'false' \\\\ # \u663e\u5f0f\u5173\u95edNLIRG
\n –model_path_to_load Qwen\/Qwen3-1.5B \\\\
\n –training_type 'sft' \\\\
\n –use_lora 'true' \\\\
\n –batch_size 1 \\\\
\n –token_batch 10 \\\\
\n –data_path your_data.json \\\\
\n –output_dir baseline_results<\/p>\n# \u5b9e\u9a8c\u7ec4（\u5f00\u542fNLIRG）
\npython -m training_code.start_training \\\\
\n –use_NLIRG 'true' \\\\ # \u663e\u5f0f\u5f00\u542fNLIRG
\n –model_path_to_load Qwen\/Qwen3-1.5B \\\\
\n –training_type 'sft' \\\\
\n –use_lora 'true' \\\\
\n –batch_size 1 \\\\
\n –token_batch 10 \\\\
\n –data_path your_data.json \\\\
\n –output_dir nlirg_results <\/p>\n\u5173\u952e\u63d0\u793a：Y-Trainer\u9ed8\u8ba4\u5f00\u542fNLIRG，\u505a\u5bf9\u7167\u5b9e\u9a8c\u65f6\u52a1\u5fc5\u663e\u5f0f\u8bbe\u7f6e–use_NLIRG 'false'，\u907f\u514d\u8bef\u5224\u3002\u6240\u6709\u5176\u4ed6\u53c2\u6570\u5fc5\u987b\u4fdd\u6301\u4e00\u81f4，\u786e\u4fdd\u5b9e\u9a8c\u516c\u5e73\u6027\u3002<\/p>\n<\/p>\n
6.2 \u8bc4\u4f30\u6307\u6807\u4f53\u7cfb（\u4e09\u7ef4\u5ea6\u9a8c\u8bc1）<\/h4>\n<\/p>\n
\u4e0d\u8981\u53ea\u770b\u8bad\u7ec3loss！\u5efa\u7acb\u4e09\u7ef4\u5ea6\u8bc4\u4f30\u4f53\u7cfb：<\/p>\n<\/p>\n
\n\n
\n \n \u8bc4\u4f30\u7ef4\u5ea6<\/p>\n<\/td>\n
\n \u5177\u4f53\u6307\u6807<\/p>\n<\/td>\n
\n \u8bc4\u4f30\u65b9\u6cd5<\/p>\n<\/td>\n
\n \u91cd\u8981\u6027<\/p>\n<\/td>\n<\/tr>\n
\n \n \u9886\u57df\u80fd\u529b<\/p>\n<\/td>\n
\n \u4efb\u52a1\u51c6\u786e\u7387\u3001\u4e1a\u52a1\u89c4\u5219\u9075\u5faa\u5ea6<\/p>\n<\/td>\n
\n \u5782\u76f4\u4efb\u52a1\u6d4b\u8bd5\u96c6（200+\u6837\u672c）<\/p>\n<\/td>\n
\n \u2605\u2605\u2605<\/p>\n<\/td>\n<\/tr>\n
\n \n \u901a\u7528\u80fd\u529b<\/p>\n<\/td>\n
\n \u6307\u4ee4\u9075\u5faa\u7387\u3001\u57fa\u7840\u5bf9\u8bdd\u8d28\u91cf\u3001\u5e38\u8bc6\u95ee\u7b54\u6b63\u786e\u7387<\/p>\n<\/td>\n
\n \u901a\u7528\u56de\u5f52\u6d4b\u8bd5\u96c6（100+\u6837\u672c）<\/p>\n<\/td>\n
\n \u2605\u2605\u2606<\/p>\n<\/td>\n<\/tr>\n
\n \n \u8f93\u51fa\u8d28\u91cf<\/p>\n<\/td>\n
\n \u91cd\u590d\u7387(3-gram)\u3001\u4fe1\u606f\u5bc6\u5ea6\u3001\u683c\u5f0f\u7a33\u5b9a\u6027<\/p>\n<\/td>\n
\n \u81ea\u52a8\u5316\u811a\u672c+\u4eba\u5de5\u62bd\u6837（50+\u6837\u672c）<\/p>\n<\/td>\n
\n \u2605\u2605\u2606<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\u8bc4\u4f30\u5de5\u5177\u63a8\u8350：<\/p>\n
- \n
- \n
5.3 \u5927\u578bCPT\u8bad\u7ec3（\u591a\u5361\u5168\u91cf\u65b9\u6848）<\/h4>\n
deepspeed –master_port 29501 –include localhost:0,1,2,3 \\\\
\n –module training_code.start_training \\\\
\n –model_path_to_load Qwen\/Qwen3-8B \\\\
\n –training_type 'cpt' \\\\
\n –use_NLIRG 'true' \\\\
\n –use_deepspeed 'true' \\\\ # \u542f\u7528DeepSpeed
\n –pack_length 2048 \\\\ # \u6587\u672c\u6253\u5305\u957f\u5ea6，\u77ed\u6587\u672c\u8bad\u7ec3\u5fc5\u8bbe
\n –batch_size 2 \\\\
\n –epoch 3 \\\\
\n –data_path domain_corpus.json \\\\
\n –output_dir results_cpt <\/p>\n\u591a\u5361\u8bad\u7ec3\u6ce8\u610f\u4e8b\u9879：<\/p>\n
- \n
- \n
5.2 \u5782\u76f4SFT\u8bad\u7ec3（\u5355\u5361LoRA\u65b9\u6848）<\/h4>\n
python -m training_code.start_training \\\\
\n –model_path_to_load Qwen\/Qwen3-1.5B \\\\ # \u57fa\u7840\u6a21\u578b\u8def\u5f84，\u53ef\u66ff\u6362\u4e3a\u672c\u5730\u8def\u5f84
\n –training_type 'sft' \\\\ # \u8bad\u7ec3\u7c7b\u578b：sft(\u6307\u4ee4\u5fae\u8c03)\/cpt(\u7ee7\u7eed\u9884\u8bad\u7ec3)
\n –use_NLIRG 'true' \\\\ # \u6838\u5fc3\u7b97\u6cd5\u5f00\u5173，true\u4e3a\u5f00\u542f
\n –use_lora 'true' \\\\ # LoRA\u5f00\u5173，\u663e\u5b58\u4e0d\u8db3\u5fc5\u5f00
\n –lora_target_modules "q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj" \\\\
\n –batch_size 1 \\\\ # \u5355\u5361\u901a\u5e38\u8bbe\u4e3a1
\n –token_batch 10 \\\\ # SFT\u5173\u952e\u53c2\u6570！\u63a7\u5236\u6bcf\u6b21\u53cd\u5411\u4f20\u64ad\u7684token\u6570
\n –epoch 3 \\\\ # \u8bad\u7ec3\u8f6e\u6b21，\u5efa\u8bae3-5\u8f6e
\n –data_path your_vertical_data.json \\\\ # \u4f60\u7684\u5782\u76f4\u6570\u636e\u8def\u5f84
\n –output_dir results_sft \\\\ # \u8f93\u51fa\u76ee\u5f55
\n –checkpoint_epoch '0,1' # \u4fdd\u5b58\u68c0\u67e5\u70b9\u7684\u8f6e\u6b21，\u5982'0,1,2' <\/p>\n\u53c2\u6570\u8be6\u89e3（\u9488\u5bf9SFT\u573a\u666f）：<\/p>\n
- \n
- \n
Y-Trainer\u5b98\u65b9\u6587\u6863\u4e00\u9488\u89c1\u8840\u5730\u6307\u51fa：<\/p>\n
"\u707e\u96be\u6027\u9057\u5fd8\u901a\u5e38\u662f\u7531\u8fc7\u96be\u8bed\u6599\u5bfc\u81f4，\u901a\u8fc7\u8bc6\u522b\u8fd9\u4e9btoken，\u8fdb\u884c\u52a8\u6001\u8c03\u6574，\u53ef\u6709\u6548\u907f\u514d\u3002\u8fc7\u62df\u5408\u662f\u7531\u76f8\u4f3c\u8bed\u6599\u6216\u8005\u6a21\u578b\u5df2\u7ecf\u638c\u63e1\u7684\u77e5\u8bc6\u5bfc\u81f4，\u901a\u8fc7\u8bc6\u522b\u8fd9\u4e9btoken，\u8fdb\u884c\u52a8\u6001\u8c03\u6574，\u53ef\u6709\u6548\u907f\u514d\u3002"<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260126100501-69773c4da03b4.png\")
<\/p>\n\u4e09\u3001\u4f20\u7edf\u89e3\u6cd5\u7684\u5c40\u9650\u6027：\u5728"\u6570\u636e\u4fa7"\u4fee\u4fee\u8865\u8865<\/h3>\n<\/p>\n
\u9762\u5bf9\u4e0a\u8ff0\u95ee\u9898，\u4e1a\u754c\u5e38\u7528\u89e3\u51b3\u65b9\u6848\u53ca\u5176\u5c40\u9650\u6027：<\/p>\n<\/p>\n
\n\n
\n \n \u89e3\u51b3\u65b9\u6848<\/p>\n<\/td>\n
\n \u4f18\u70b9<\/p>\n<\/td>\n
\n \u7f3a\u70b9<\/p>\n<\/td>\n
\n \u5b9e\u9645\u5e94\u7528\u75db\u70b9<\/p>\n<\/td>\n<\/tr>\n
\n \n \u6df7\u5408\u901a\u7528\u8bed\u6599<\/p>\n<\/td>\n
\n \u4fdd\u7559\u57fa\u7840\u80fd\u529b<\/p>\n<\/td>\n
\n \u9700\u8981\u5408\u89c4\u83b7\u53d6\/\u6e05\u6d17\/\u7ef4\u62a4\u901a\u7528\u6570\u636e；\u6bd4\u4f8b\u9700\u53cd\u590d\u8c03\u6574<\/p>\n<\/td>\n
\n \u6570\u636e\u5408\u89c4\u9650\u5236\u4e0b\u65e0\u6cd5\u4f7f\u7528\u5916\u90e8\u901a\u7528\u8bed\u6599<\/p>\n<\/td>\n<\/tr>\n
\n \n \u964d\u4f4e\u5b66\u4e60\u5f3a\u5ea6<\/p>\n<\/td>\n
\n \u7f13\u89e3\u8fc7\u62df\u5408<\/p>\n<\/td>\n
\n \u4e13\u9879\u80fd\u529b\u63d0\u5347\u53d7\u9650；"\u65e2\u4e0d\u591f\u4e13\u4e1a，\u4e5f\u4e0d\u591f\u7a33\u5b9a"<\/p>\n<\/td>\n
\n \u4e13\u4e1a\u95ee\u9898\u51c6\u786e\u7387\u663e\u8457\u4e0b\u964d<\/p>\n<\/td>\n<\/tr>\n
\n \n \u6570\u636e\u6e05\u6d17\u89c4\u8303<\/p>\n<\/td>\n
\n \u63d0\u9ad8\u6570\u636e\u8d28\u91cf<\/p>\n<\/td>\n
\n \u65e0\u6cd5\u89e3\u51b3"\u540c\u4e00\u6570\u636e\u4e2d\u54ea\u4e9btoken\u5e94\u8be5\u5b66\u5f97\u66f4\u91cd"<\/p>\n<\/td>\n
\n \u6e05\u6d17\u591a\u8f6e\u540e\u4ecd\u9700\u52a0\u5165\u901a\u7528\u8bed\u6599<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\u8fd9\u4e9b\u65b9\u6cd5\u672c\u8d28\u4e0a\u90fd\u662f\u5728"\u6570\u636e\u4fa7"\u8865\u6d1e，\u800c\u975e\u5728"\u4fe1\u53f7\u4fa7"\u4f18\u5316\u3002\u5c31\u50cf\u7ed9\u53d1\u70e7\u75c5\u4eba\u76d6\u88ab\u5b50，\u800c\u975e\u6cbb\u7597\u75c5\u56e0\u3002<\/p>\n<\/p>\n
Y-Trainer\u7684\u5b9a\u4f4d\u63d0\u4f9b\u65b0\u601d\u8def：<\/p>\n
"\u65e0\u9700\u4f9d\u8d56\u901a\u7528\u8bed\u6599，\u5373\u53ef\u5353\u8d8a\u5730\u4fdd\u7559\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b，\u5b88\u4f4f\u6838\u5fc3\u80fd\u529b\u7684\u540c\u65f6\u5b9e\u73b0\u4e13\u9879\u63d0\u5347！\u65e0\u9700\u8bed\u6599\u5e73\u8861，\u5373\u4f7f\u5728\u8bed\u6599\u5206\u5e03\u5f88\u4e0d\u5747\u5300\u7684\u60c5\u51b5\u4e0b，\u4f9d\u7136\u80fd\u591f\u7a33\u5b9a\u8bad\u7ec3\u3002"<\/p>\n<\/p>\n
\u56db\u3001NLIRG\u7b97\u6cd5\u8be6\u89e3：\u4e3a\u6bcf\u4e2atoken\u7cbe\u51c6\u5206\u914d"\u5b66\u4e60\u80fd\u91cf"<\/h3>\n<\/p>\n
4.1 \u6838\u5fc3\u601d\u60f3：\u7528loss\u8861\u91cf\u96be\u5ea6，\u52a8\u6001\u8c03\u6574\u68af\u5ea6<\/h4>\n<\/p>\n
NLIRG（Nonlinear Learning Intensity Regulation，\u57fa\u4e8e\u68af\u5ea6\u7684\u975e\u7ebf\u6027\u5b66\u4e60\u5f3a\u5ea6\u8c03\u8282）\u7684\u6838\u5fc3\u601d\u60f3\u975e\u5e38\u76f4\u89c2：\u6839\u636e\u6bcf\u4e2atoken\u7684loss\u503c，\u52a8\u6001\u8ba1\u7b97\u5176\u68af\u5ea6\u6743\u91cd，\u5b9e\u73b0"\u667a\u80fd\u56e0\u6750\u65bd\u6559"\u3002<\/p>\n
def dynamic_sigmoid_batch(losses, max_lr=1.0, x0=1.2, min_lr=5e-8,
\n k=1.7, loss_threshold=3.0, loss_deadline=15.0):
\n """
\n NLIRG\u6838\u5fc3\u7b97\u6cd5：\u57fa\u4e8e\u635f\u5931\u503c\u7684\u52a8\u6001\u6743\u91cd\u8ba1\u7b97<\/p>\n\u53c2\u6570\u8bf4\u660e：
\n – losses: \u635f\u5931\u503c\u5f20\u91cf (shape: [batch_size, seq_len])
\n – max_lr: \u6700\u5927\u5b66\u4e60\u7387\u6743\u91cd (\u9ed8\u8ba4: 1.0)，\u63a7\u5236\u68af\u5ea6\u589e\u5f3a\u4e0a\u9650
\n – x0: \u7b2c\u4e00\u4e2asigmoid\u51fd\u6570\u7684\u4e2d\u5fc3\u70b9 (\u9ed8\u8ba4: 1.2)，\u51b3\u5b9a\u4f4eloss\u533a\u95f4\u7684\u5206\u754c
\n – min_lr: \u6700\u5c0f\u5b66\u4e60\u7387\u6743\u91cd (\u9ed8\u8ba4: 5e-8)，\u9632\u6b62\u5b8c\u5168\u963b\u65ad\u68af\u5ea6
\n – k: sigmoid\u51fd\u6570\u7684\u659c\u7387 (\u9ed8\u8ba4: 1.7)，\u63a7\u5236\u66f2\u7ebf\u9661\u5ced\u7a0b\u5ea6
\n – loss_threshold: \u635f\u5931\u9608\u503c (\u9ed8\u8ba4: 3.0)，\u533a\u5206\u4e2d\u7b49\u548c\u56f0\u96be\u533a\u57df
\n – loss_deadline: \u635f\u5931\u4e0a\u9650 (\u9ed8\u8ba4: 15.0)，\u8d85\u8fc7\u6b64\u503c\u68af\u5ea6\u5f52\u96f6<\/p>\n\u8fd4\u56de：\u6743\u91cd\u5f20\u91cf，\u7528\u4e8e\u8c03\u6574\u53cd\u5411\u4f20\u64ad\u7684\u68af\u5ea6
\n """ <\/p>\n\u5173\u952e\u70b9\u5728\u4e8e：\u8fd9\u4e0d\u662f\u65b0\u7684\u635f\u5931\u51fd\u6570，\u800c\u662f\u5bf9\u539f\u6709\u68af\u5ea6\u8fdb\u884c\u7684\u667a\u80fd\u52a0\u6743，\u5b9e\u73b0\u8bad\u7ec3\u54f2\u5b66\u7684\u8f6c\u53d8\u2014\u2014\u4ece"\u5e73\u5747\u7528\u529b"\u5230"\u7cbe\u51c6\u6295\u653e"\u3002<\/p>\n<\/p>\n
4.2 \u56db\u6bb5\u5f0f\u68af\u5ea6\u5206\u914d\u7b56\u7565：\u79d1\u5b66\u5e94\u5bf9\u4e0d\u540c\u96be\u5ea6token<\/h4>\n<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260126100501-69773c4d53c23.png\")
<\/p>\n<\/p>\nY-Trainer\u5c06loss\u5206\u4e3a\u56db\u4e2a\u533a\u95f4，\u5b9e\u65bd\u4e0d\u540c\u7b56\u7565：<\/p>\n<\/p>\n
\n\n
\n \n \u635f\u5931\u533a\u95f4<\/p>\n<\/td>\n
\n \u68af\u5ea6\u7b56\u7565<\/p>\n<\/td>\n
\n \u8bad\u7ec3\u76ee\u7684<\/p>\n<\/td>\n
\n \u5b9e\u9645\u6548\u679c<\/p>\n<\/td>\n<\/tr>\n
\n \n loss \u2264 1.45<\/p>\n<\/td>\n
\n \u524a\u51cf\u68af\u5ea6<\/p>\n<\/td>\n
\n \u907f\u514d\u8fc7\u62df\u5408<\/p>\n<\/td>\n
\n \u9632\u6b62\u6a21\u578b\u8fc7\u5ea6\u8bb0\u5fc6\u7b80\u5355\u6837\u672c，\u51cf\u5c11\u590d\u8bfb\u73b0\u8c61<\/p>\n<\/td>\n<\/tr>\n
\n \n 1.45 < loss < 6.6<\/p>\n<\/td>\n
\n \u589e\u5f3a\u68af\u5ea6<\/p>\n<\/td>\n
\n \u9ad8\u6548\u5b66\u4e60<\/p>\n<\/td>\n
\n \u91cd\u70b9\u653b\u514b\u6709\u4ef7\u503c\u6837\u672c，\u63d0\u5347\u5b66\u4e60\u6548\u7387<\/p>\n<\/td>\n<\/tr>\n
\n \n 6.6 < loss < 15<\/p>\n<\/td>\n
\n \u524a\u51cf\u68af\u5ea6<\/p>\n<\/td>\n
\n \u7a33\u5b9a\u8bad\u7ec3<\/p>\n<\/td>\n
\n \u9632\u6b62\u56f0\u96be\u6837\u672c\u5e26\u504f\u6a21\u578b，\u4fdd\u62a4\u5df2\u6709\u80fd\u529b<\/p>\n<\/td>\n<\/tr>\n
\n \n loss \u2265 15.0<\/p>\n<\/td>\n
\n \u68af\u5ea6\u5f52\u96f6<\/p>\n<\/td>\n
\n \u9694\u79bb\u566a\u58f0<\/p>\n<\/td>\n
\n \u5ffd\u7565\u660e\u663e\u9519\u8bef\u6837\u672c，\u907f\u514d\u8bad\u7ec3\u5d29\u574f<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260126100502-69773c4e513cf.png\")
<\/p>\nNLIRG\u6743\u91cd\u8ba1\u7b97\u516c\u5f0f，L\u4e3a\u635f\u5931\u503c，W(L)\u4e3a\u5bf9\u5e94\u6743\u91cd<\/p>\n<\/p>\n
\u5173\u952e\u6d1e\u5bdf：\u4e2d\u7b49\u96be\u5ea6token（1.45 < loss < 6.6）\u624d\u662f\u771f\u6b63\u7684"\u4ef7\u503c\u6d3c\u5730"，\u503c\u5f97\u5206\u914d\u66f4\u591a"\u5b66\u4e60\u80fd\u91cf"\u3002\u8fd9\u4e0e\u4eba\u7c7b\u5b66\u4e60\u89c4\u5f8b\u4e00\u81f4：\u592a\u7b80\u5355\u7684\u5185\u5bb9\u65e0\u9700\u53cd\u590d\u7ec3\u4e60，\u592a\u96be\u7684\u5185\u5bb9\u6682\u65f6\u653e\u5f03，\u4e2d\u7b49\u96be\u5ea6\u7684\u5185\u5bb9\u6700\u80fd\u4fc3\u8fdb\u80fd\u529b\u63d0\u5347\u3002<\/p>\n<\/p>\n
\u4e94\u3001\u5b9e\u6218\u6307\u5357：\u4e09\u6b65\u642d\u5efa\u53ef\u5bf9\u7167\u7684\u8bad\u7ec3\u6d41\u7a0b<\/h3>\n<\/p>\n
5.1 \u73af\u5883\u51c6\u5907（5\u5206\u949f\u5feb\u901f\u4e0a\u624b）<\/h4>\n
# \u514b\u9686\u4ee3\u7801
\ngit clone https:\/\/github.com\/yafo-ai\/y-trainer.git
\ncd y-trainer<\/p>\n# \u5b89\u88c5\u4f9d\u8d56（\u5355\u5361\u73af\u5883\u53ef\u4e0d\u88c5deepspeed）
\n# \u6ce8\u610f：\u5b98\u65b9\u6587\u6863\u4e2dpeft\u5305\u540d\u6709\u65ad\u884c，\u5b9e\u9645\u5e94\u4e3apeft>=0.10.0
\npip install torch peft>=0.10.0 tensorboard matplotlib
\npip install -r requirements.txt<\/p>\n# \u5efa\u8baePython 3.8+\u73af\u5883
\npython –version <\/p>\n\u73af\u5883\u914d\u7f6e\u6280\u5de7：<\/p>\n
- \n
- \n
\u73b0\u8c612：\u5c11\u91cf\u9ad8loss token\u89e6\u53d1\u8fc7\u5f3a\u66f4\u65b0<\/p>\n
\u5782\u76f4\u6570\u636e\u4e2d\u4e0d\u53ef\u907f\u514d\u5b58\u5728\u56f0\u96be\u6837\u672c：\u4e13\u4e1a\u672f\u8bed\u62fc\u5199\u9519\u8bef\u3001\u957f\u5c3e\u8868\u8fbe\u3001\u566a\u58f0\u6807\u6ce8\u7b49\u3002\u5b83\u4eec\u5bf9\u5e94\u9ad8loss token，\u4ea7\u751f\u66f4\u5927\u68af\u5ea6，\u8feb\u4f7f\u6a21\u578b\u7528\u5927\u4ee3\u4ef7\u62df\u5408\u5c11\u91cf\u5f02\u5e38\u4fe1\u53f7，\u5bfc\u81f4：<\/p>\n<\/p>\n
- \n
- \n
\u4eca\u5929，\u6211\u60f3\u548c\u5927\u5bb6\u63a2\u8ba8\u4e00\u4e2a\u53ef\u80fd\u88ab\u5ffd\u89c6\u7684\u65b9\u5411：\u8bad\u7ec3\u4fe1\u53f7\u7684\u7cbe\u7ec6\u5206\u914d\u3002\u4e0d\u662f\u6240\u6709token\u90fd\u503c\u5f97\u540c\u7b49\u5bf9\u5f85，\u800cY-Trainer\u7684NLIRG\u7b97\u6cd5\u6b63\u662f\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\u7684\u5229\u5668\u3002<\/p>\n<\/p>\n
\u4e8c\u3001\u95ee\u9898\u6839\u6e90：\u68af\u5ea6\u5206\u914d\u4e0d\u5408\u7406\u662f\u5143\u51f6<\/h3>\n<\/p>\n
2.1 \u88ab\u5ffd\u89c6\u7684\u6838\u5fc3\u4e8b\u5b9e：SFT\u7684\u68af\u5ea6\u6765\u6e90\u662ftoken，\u4e0d\u662f\u6837\u672c<\/h4>\n<\/p>\n
\u5728\u6307\u4ee4\u5fae\u8c03\u4e2d，\u771f\u6b63\u9a71\u52a8\u6a21\u578b\u5b66\u4e60\u7684\u662f\u6bcf\u4e2atoken\u4ea7\u751f\u7684\u68af\u5ea6，\u800c\u975e\u6574\u4e2a\u6837\u672c\u3002\u4e00\u4e2a\u6837\u672c\u4e2d\u54ea\u4e9btoken\u8d21\u732e\u5927\u3001\u54ea\u4e9b\u68af\u5ea6\u66f4\u6fc0\u8fdb，\u76f4\u63a5\u51b3\u5b9a\u4e86"\u6a21\u578b\u5728\u8fd9\u4e00\u8f6e\u8bad\u7ec3\u4e2d\u5230\u5e95\u5b66\u4e86\u4ec0\u4e48"\u3002<\/p>\n<\/p>\n
2.2 \u4e24\u5927\u6781\u7aef\u73b0\u8c61\u5982\u4f55\u8054\u624b\u6bc1\u6389\u4f60\u7684\u6a21\u578b<\/h4>\n<\/p>\n
<\/p>\n<\/p>\n\u73b0\u8c611：\u5927\u91cf\u4f4eloss token\u53cd\u590d\u51fa\u73b0<\/p>\n
\u5782\u76f4\u6570\u636e\u5f80\u5f80\u5b58\u5728\u5f3a\u6a21\u677f：\u6cd5\u5f8b\u6587\u4e66\u7684"\u6839\u636e\u300aXX\u6cd5\u300b\u7b2cX\u6761"\u3001\u533b\u7597\u62a5\u544a\u7684"\u60a3\u8005\u4e3b\u8bc9…"\u3001\u5ba2\u670d\u8bdd\u672f\u7684"\u5f88\u62b1\u6b49\u7ed9\u60a8\u5e26\u6765\u4e0d\u4fbf"\u7b49\u3002\u8fd9\u4e9btoken\u5728\u8bad\u7ec3\u65e9\u671f\u5c31\u8fdb\u5165\u4f4eloss\u533a\u95f4\u3002\u5982\u679c\u7ee7\u7eed\u4ee5\u76f8\u540c\u5f3a\u5ea6\u8bad\u7ec3\u5b83\u4eec，\u6a21\u578b\u4f1a\u5728"\u8868\u9762\u6a21\u5f0f"\u4e0a\u8fc7\u62df\u5408，\u5bfc\u81f4：<\/p>\n<\/p>\n
- \n
- \n
\u6700\u4ee4\u4eba\u6cae\u4e27\u7684\u662f：\u8fd9\u4e9b\u95ee\u9898\u5f80\u5f80\u540c\u65f6\u51fa\u73b0，\u4e14\u5e38\u89c4\u89e3\u51b3\u65b9\u6848\u6548\u679c\u6709\u9650：<\/p>\n
- \n
- \n
\u5f53\u6211\u4eec\u5728\u4e00\u4e2a\u89d2\u4e0a\u53d1\u529b，\u53e6\u5916\u4e24\u4e2a\u89d2\u5f80\u5f80\u4f1a\u5d29\u6e83\u3002\u4f60\u662f\u5426\u4e5f\u7ecf\u5386\u8fc7\u8fd9\u4e9b\u4ee4\u4eba\u5d29\u6e83\u7684\u573a\u666f？<\/p>\n<\/p>\n
- \n