{"id":284,"date":"2026-01-18T18:07:12","date_gmt":"2026-01-18T10:07:12","guid":{"rendered":"https:\/\/blog.zeronely.cn\/?p=284"},"modified":"2026-01-18T18:07:13","modified_gmt":"2026-01-18T10:07:13","slug":"%e5%85%b3%e4%ba%8e%e5%be%ae%e8%b0%83%e7%9a%84%e6%b1%87%e6%8a%a5","status":"publish","type":"post","link":"https:\/\/blog.zeronely.cn\/index.php\/2026\/01\/18\/%e5%85%b3%e4%ba%8e%e5%be%ae%e8%b0%83%e7%9a%84%e6%b1%87%e6%8a%a5\/","title":{"rendered":"\u5173\u4e8e\u5fae\u8c03\u7684\u6c47\u62a5"},"content":{"rendered":"\n<p>CUDA_VISIBLE_DEVICES=2,3<br>python -m vllm.entrypoints.openai.api_server<br>&#8211;model \/netcache\/huggingface\/Qwen2.5-7B-Instruct<br>&#8211;tensor-parallel-size 2<br>&#8211;host 0.0.0.0<br>&#8211;port 18001<br>&#8211;gpu-memory-utilization 0.7<br>&#8211;dtype float16<\/p>\n\n\n\n<p>{&#8216;loss&#8217;: 0.8365, &#8216;grad_norm&#8217;: 0.0769263431429863, &#8216;learning_rate&#8217;: 4.241918085821547e-05, &#8216;epoch&#8217;: 0.71}<br>{&#8216;loss&#8217;: 0.8559, &#8216;grad_norm&#8217;: 0.08880975842475891, &#8216;learning_rate&#8217;: 2.9870182219917564e-05, &#8216;epoch&#8217;: 0.76}<br>{&#8216;loss&#8217;: 0.8649, &#8216;grad_norm&#8217;: 0.08375994861125946, &#8216;learning_rate&#8217;: 1.9194584955692706e-05, &#8216;epoch&#8217;: 0.81}<br>{&#8216;loss&#8217;: 0.8305, &#8216;grad_norm&#8217;: 0.09028902649879456, &#8216;learning_rate&#8217;: 1.067756988044848e-05, &#8216;epoch&#8217;: 0.86}<br>100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 396\/396 [3:17:24&lt;00:00, 29.91s\/it]<\/p>\n\n\n\n<p>LoRA saved to: \/home\/wanglu\/whz\/11\/lora_trajrel_qwen2p5_7b<\/p>\n\n\n\n<p>CUDA_VISIBLE_DEVICES=2,3<br>python -m vllm.entrypoints.openai.api_server<br>&#8211;model \/netcache\/huggingface\/Qwen2.5-7B-Instruct<br>&#8211;tensor-parallel-size 2<br>&#8211;host 0.0.0.0<br>&#8211;port 18001<br>&#8211;gpu-memory-utilization 0.7<br>&#8211;dtype float16<br>&#8211;enable-lora<br>&#8211;lora-modules trajrel=\/home\/wanglu\/whz\/11\/lora_trajrel_qwen2p5_7b<br>&#8211;max-loras 1<br>&#8211;max-lora-rank 16<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"\u5de5\u4f5c\u6c47\u62a5\">\u5de5\u4f5c\u6c47\u62a5<\/h1>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u57fa\u7ebf\u7ed3\u679c\">\u57fa\u7ebf\u7ed3\u679c<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"1-\u6570\u636e\u6784\u6210\">1. \u6570\u636e\u6784\u6210<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bad\u7ec3\u96c6\uff1a<code>train.jsonl<\/code>\uff0c\u5171\u00a0<strong>6334 \u6761<\/strong><\/li>\n\n\n\n<li>\u9a8c\u8bc1\u96c6\uff1a<code>val.jsonl<\/code>\uff0c\u5171\u00a0<strong>680 \u6761<\/strong><\/li>\n\n\n\n<li>\u6bcf\u6761\u6837\u672c\u683c\u5f0f\uff1a\n<ul class=\"wp-block-list\">\n<li><code>system<\/code>\uff1a\u4efb\u52a1\u8bf4\u660e\uff08\u5206\u7c7b\u4efb\u52a1\uff09<\/li>\n\n\n\n<li><code>user<\/code>\uff1a\u8f68\u8ff9\u6587\u672c\u8f93\u5165\uff08\u957f\u5e8f\u5217\uff0c\u5305\u542b\u591a\u5e27\u4fe1\u606f\uff09<\/li>\n\n\n\n<li><code>assistant<\/code>\uff1aJSON \u683c\u5f0f\u6807\u7b7e\uff0c\u4f8b\u5982\uff1a<code>{\"relation\": \"yields to\"}<\/code><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"2-\u521d\u59cb\u57fa\u7ebf\u8bc4\u6d4b\u672a\u5fae\u8c03\">2. \u521d\u59cb\u57fa\u7ebf\u8bc4\u6d4b\uff08\u672a\u5fae\u8c03\uff09<\/h3>\n\n\n\n<p>\u4f7f\u7528&nbsp;<strong>Qwen2.5-7B-Instruct<\/strong>\uff0c\u901a\u8fc7 vLLM API \u8fdb\u884c\u63a8\u7406\u8bc4\u6d4b\uff0c\u7ed3\u679c\u5982\u4e0b\uff08val \u96c6\uff09\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Accuracy \u2248\u00a0<strong>0.40<\/strong><\/li>\n\n\n\n<li>Macro Recall \u2248\u00a0<strong>0.38<\/strong><\/li>\n\n\n\n<li>Macro F1 \u2248\u00a0<strong>0.30<\/strong><\/li>\n\n\n\n<li>\u9884\u6d4b\u7ed3\u679c\u660e\u663e\u5927\u91cf\u9884\u6d4b\u4e3a\u67d0\u4e00\u7c7b<br><img decoding=\"async\" src=\"https:\/\/file+.vscode-resource.vscode-cdn.net\/d:\/research\/%E7%A7%91%E7%A0%94%E8%AE%AD%E7%BB%83\/13\/image.png\" alt=\"alt text\"><\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u5fae\u8c03\u65b9\u6848\u4e0e\u5df2\u5b8c\u6210\u5de5\u4f5c\">\u5fae\u8c03\u65b9\u6848\u4e0e\u5df2\u5b8c\u6210\u5de5\u4f5c<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"1-\u5fae\u8c03\u6280\u672f\u8def\u7ebf\">1. \u5fae\u8c03\u6280\u672f\u8def\u7ebf<\/h3>\n\n\n\n<p>\u91c7\u7528&nbsp;<strong>QLoRA\uff084bit\uff09+ LoRA Adapter<\/strong>&nbsp;\u7684\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u65b9\u6848\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u57fa\u6a21\u578b\uff1aQwen2.5-7B-Instruct<\/li>\n\n\n\n<li>\u91cf\u5316\uff1aNF4 4-bit<\/li>\n\n\n\n<li>LoRA \u63d2\u5165\u4f4d\u7f6e\uff1a\n<ul class=\"wp-block-list\">\n<li><code>q_proj<\/code>,\u00a0<code>k_proj<\/code>,\u00a0<code>v_proj<\/code>,\u00a0<code>o_proj<\/code><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u8bad\u7ec3\u53c2\u6570\u91cf\uff1a\n<ul class=\"wp-block-list\">\n<li>\u53ef\u8bad\u7ec3\u53c2\u6570\u7ea6\u00a0<strong>10M<\/strong><\/li>\n\n\n\n<li>\u5360\u539f\u6a21\u578b\u53c2\u6570\u7ea6\u00a0<strong>0.13%<\/strong><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"2-\u5fae\u8c03\u65b9\u5f0f\u5c1d\u8bd5\">2. \u5fae\u8c03\u65b9\u5f0f\u5c1d\u8bd5<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u65b9\u6848-a\u751f\u6210\u5f0f-sft\">\u65b9\u6848 A\uff1a\u751f\u6210\u5f0f SFT<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bad\u7ec3\u65b9\u5f0f\uff1a\n<ul class=\"wp-block-list\">\n<li>\u8f93\u5165\uff1asystem + user + assistant<\/li>\n\n\n\n<li>Loss\uff1aCausal LM token-level loss(\u672c\u6765\u60f3\u5148\u4e0d\u52a0mask\u5fae\u8c03)<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u7ed3\u679c\uff1a\u5931\u8d25<\/li>\n\n\n\n<li>\u731c\u6d4b\u53ef\u80fd\u7684\u95ee\u9898\uff1a\n<ul class=\"wp-block-list\">\n<li>loss \u5bf9\u00a0<strong>system \/ user \/ assistant<\/strong>\u00a0\u90fd\u751f\u6548<\/li>\n\n\n\n<li>\u5b9e\u9645\u8bad\u7ec3\u76ee\u6807\u662f\u201c\u590d\u8ff0 prompt + \u8f93\u51fa JSON\u201d<\/li>\n\n\n\n<li>\u4e0e\u8bc4\u6d4b\u7684<strong>\u5206\u7c7b<\/strong>\u4e0d\u5bf9\u9f50<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u7ed3\u679c\uff1a<br>\u8bad\u7ec3\u65f6loss\u4e0b\u964d\uff0c\u4f46\u662f\u5fae\u8c03\u540e\u6b63\u786e\u7387\u4e0b\u964d\uff0c\u9884\u6d4b\u5206\u5e03\u66f4\u52a0\u504f\u7f6e\u3002<br><img decoding=\"async\" src=\"https:\/\/file+.vscode-resource.vscode-cdn.net\/d:\/research\/%E7%A7%91%E7%A0%94%E8%AE%AD%E7%BB%83\/13\/image-1.png\" alt=\"alt text\"><\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"\u65b9\u6848-bmasked-sft\">\u65b9\u6848 B\uff1aMasked SFT<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6539\u8fdb\u70b9\uff1a\n<ul class=\"wp-block-list\">\n<li><strong>\u53ea\u5728 assistant \u8f93\u51fa\u90e8\u5206\u8ba1\u7b97 loss\uff0c\u8ba1\u7b97token \u7ea7\u522b\u7684\u4ea4\u53c9\u71b5\u635f\u5931<\/strong><\/li>\n\n\n\n<li>system \/ user token \u7684 label \u8bbe\u4e3a\u00a0<code>-100<\/code>\uff08mask \u6389\uff09<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u6280\u672f\u5b9e\u73b0\uff1a\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528 tokenizer chat_template \u5b9a\u4f4d assistant token \u533a\u95f4<\/li>\n\n\n\n<li>\u6784\u9020\u00a0<code>labels<\/code>\uff0c\u786e\u4fdd loss \u53ea\u53cd\u5411\u4f20\u64ad\u5230\u7b54\u6848\u90e8\u5206<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u8bad\u7ec3\u8d44\u6e90\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5355\u5f20 RTX 3090<\/li>\n\n\n\n<li>\u8bad\u7ec3 2 \u4e2a epoch<\/li>\n\n\n\n<li>\u8bad\u7ec3\u8017\u65f6\u7ea6\u00a0<strong>6.5 \u5c0f\u65f6<\/strong><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u8bad\u7ec3\u65e5\u5fd7\">\u8bad\u7ec3\u65e5\u5fd7<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bad\u7ec3 loss\uff1a\n<ul class=\"wp-block-list\">\n<li>\u521d\u59cb \u2248 1.14<\/li>\n\n\n\n<li>\u6536\u655b\u81f3 \u2248\u00a0<strong>0.83<\/strong><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u68af\u5ea6\u7a33\u5b9a\uff0c\u65e0\u6570\u503c\u5f02\u5e38<\/li>\n\n\n\n<li>LoRA \u6587\u4ef6\u6210\u529f\u4fdd\u5b58\uff1a<code>\/home\/wanglu\/whz\/11\/lora_trajrel_qwen2p5_7b_masked<\/code><\/li>\n<\/ul>\n\n\n\n<p>\u4ece\u8bad\u7ec3\u89d2\u5ea6\u770b\uff0c<strong>\u6a21\u578b\u662f\u6536\u655b\u7684\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u6b63\u5e38<\/strong>\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u5fae\u8c03\u6a21\u578b\u8bc4\u6d4b\">\u5fae\u8c03\u6a21\u578b\u8bc4\u6d4b<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"1-\u6302\u8f7d\u65b9\u5f0f\">1. \u6302\u8f7d\u65b9\u5f0f<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528 vLLM API Server<\/li>\n\n\n\n<li>\u57fa\u6a21\u578b\uff1aQwen2.5-7B-Instruct<\/li>\n\n\n\n<li>LoRA Adapter\uff1a<code>trajrel<\/code><\/li>\n\n\n\n<li>\u652f\u6301\u901a\u8fc7\u00a0<code>model=trajrel<\/code>\u00a0\u6216\u663e\u5f0f\u6307\u5b9a LoRA \u8fdb\u884c\u63a8\u7406<\/li>\n<\/ul>\n\n\n\n<p>CUDA_VISIBLE_DEVICES=2,3<br>python -m vllm.entrypoints.openai.api_server<br>&#8211;model \/netcache\/huggingface\/Qwen2.5-7B-Instruct<br>&#8211;tensor-parallel-size 2<br>&#8211;host 0.0.0.0<br>&#8211;port 18001<br>&#8211;gpu-memory-utilization 0.7<br>&#8211;dtype float16<br>&#8211;enable-lora<br>&#8211;lora-modules trajrel=\/home\/wanglu\/whz\/11\/lora_trajrel_qwen2p5_7b<br>&#8211;max-loras 1<br>&#8211;max-lora-rank 16<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"2-\u5fae\u8c03\u540e\u8bc4\u6d4b\u7ed3\u679cval-\u96c6--\u5b50\u96c6\">2. \u5fae\u8c03\u540e\u8bc4\u6d4b\u7ed3\u679c\uff08val \u96c6 \/ \u5b50\u96c6\uff09<\/h3>\n\n\n\n<p>\u7ed3\u679c\u5982\u4e0b\uff1a<br>metrics: {&#8216;total_scenarios&#8217;: 680, &#8216;valid_count&#8217;: 680, &#8216;invalid_count&#8217;: 0, &#8216;accuracy&#8217;: 0.325, &#8216;recall_macro&#8217;: 0.3572, &#8216;f1_macro&#8217;: 0.2591}<br><img decoding=\"async\" src=\"https:\/\/file+.vscode-resource.vscode-cdn.net\/d:\/research\/%E7%A7%91%E7%A0%94%E8%AE%AD%E7%BB%83\/13\/image-2.png\" alt=\"alt text\"><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"3-\u7ed3\u8bba\">3. \u7ed3\u8bba<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5fae\u8c03\u672a\u5e26\u6765\u7a33\u5b9a\u7684\u51c6\u786e\u7387\u63d0\u5347<\/strong><\/li>\n\n\n\n<li>\u9884\u6d4b\u4f9d\u7136\u4e25\u91cd\u504f\u5411\u67d0\u4e00\u7c7b\u522b<\/li>\n\n\n\n<li>\u6a21\u578b\u672a\u5b66\u5230\u6e05\u6670\u7684\u4e09\u5206\u7c7b\u51b3\u7b56\u8fb9\u754c<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u7591\u95ee\u4e0e\u8fdb\u4e00\u6b65\u8ba1\u5212\">\u7591\u95ee\u4e0e\u8fdb\u4e00\u6b65\u8ba1\u5212\uff1a<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e3a\u4ec0\u4e48\u4f1a\u5168\u90e8\u504f\u7f6e\u5230\u540c\u4e00\u7c7b\uff0c\u8bad\u7ec3\u96c66000\u6761\uff0c\u4e09\u7c7b\u54042000\u6761\uff0814b\u6a21\u578b\u57fa\u7ebf\u6b63\u786e\u738750%\u65e0\u660e\u663e\u504f\u7f6e\uff09<\/li>\n\n\n\n<li>\u8ba1\u5212\uff0c\u76ee\u524d\u6682\u65f6\u662f\u7b80\u5355\u7684\u4e09\u5206\u7c7b\uff0c\u672a\u6269\u5c55\u5230\u66f4\u591a\u7684\u60c5\u51b5\uff0c\u6240\u4ee5\u5c1d\u8bd5\u673a\u5668\u5b66\u4e60\u5206\u7c7b\uff0c\u770b\u770b\u662f\u4e0d\u662f\u6570\u636e\u96c6\u6709\u95ee\u9898\uff0c\u5982\u679c\u6210\u529f\u68c0\u67e5\u5fae\u8c03\u4ee3\u7801\uff0c\u770b\u770b\u662f\u4e0d\u662f\u6709\u4ec0\u4e48\u95ee\u9898\u6216\u8005\u8bd5\u8bd5\u6846\u67b6llama-factory\u6765\u505asft<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>CUDA_VISIBLE_DEVICES=2,3python -m vllm.entrypoints.open [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-284","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/posts\/284","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/comments?post=284"}],"version-history":[{"count":1,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/posts\/284\/revisions"}],"predecessor-version":[{"id":285,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/posts\/284\/revisions\/285"}],"wp:attachment":[{"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/media?parent=284"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/categories?post=284"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/tags?post=284"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}