{"id":28076,"date":"2025-03-10T18:50:23","date_gmt":"2025-03-10T10:50:23","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=28076"},"modified":"2025-03-10T18:56:44","modified_gmt":"2025-03-10T10:56:44","slug":"openmanus-rl","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/openmanus-rl\/","title":{"rendered":"OpenManus-RL\uff1a\u5fae\u8c03\u5927\u6a21\u578b\u5f3a\u5316\u667a\u80fd\u4f53\u63a8\u7406\u4e0e\u51b3\u7b56\u80fd\u529b"},"content":{"rendered":"<p>OpenManus-RL\u662f\u7531UIUC-Ulab\u4e0e <a href=\"https:\/\/www.kdjingpai.com\/de\/metagpt\/\">MetaGPT<\/a> \u793e\u533a\u7684OpenManus\u56e2\u961f\u8054\u5408\u5f00\u53d1\u7684\u5f00\u6e90\u9879\u76ee\uff0c\u6258\u7ba1\u4e8eGitHub\u3002\u8be5\u9879\u76ee\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6280\u672f\u63d0\u5347\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u667a\u80fd\u4f53\u7684\u63a8\u7406\u4e0e\u51b3\u7b56\u80fd\u529b\uff0c\u57fa\u4e8eDeepseek-R1\u3001QwQ-32B\u7b49\u6a21\u578b\u7684\u7ecf\u9a8c\uff0c\u63a2\u7d22\u65b0\u7684\u8c03\u4f18\u65b9\u6cd5\u3002\u56e2\u961f\u5b9a\u671f\u516c\u5f00\u8fdb\u5c55\uff0c\u4ee3\u7801\u3001\u6570\u636e\u96c6\u548c\u6d4b\u8bd5\u7ed3\u679c\u5b8c\u5168\u900f\u660e\uff0c\u652f\u6301\u5728GAIA\u3001AgentBench\u3001WebShop\u3001OSWorld\u7b49\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u9a8c\u8bc1\u6548\u679c\u3002\u9879\u76ee\u9f13\u52b1\u5168\u7403\u5f00\u53d1\u8005\u8d21\u732e\u4ee3\u7801\u3001\u6570\u636e\u96c6\u6216\u8ba1\u7b97\u8d44\u6e90\uff0c\u5171\u540c\u6784\u5efa\u4e00\u4e2a\u9ad8\u6548\u7684\u667a\u80fd\u4f53\u5f00\u53d1\u751f\u6001\u3002<\/p>\n<blockquote><p>\u81f3\u6b64\uff0c\u6253\u9020\u5f00\u6e90 <a href=\"https:\/\/www.kdjingpai.com\/manus\/\">Manus<\/a> \u6700\u540e\u4e00\u5757\u77ed\u677f\u4e5f\u8865\u5168\u4e86\uff0cMetaGPT \u8fd8\u771f\u662f\u4e0b\u6b7b\u624b&#8230;\u4e0d\u8fc7 <a href=\"https:\/\/www.kdjingpai.com\/de\/mgx\/\">MGX<\/a> \u5929\u7136\u53ef\u4ee5\u8986\u76d6 <a href=\"https:\/\/www.kdjingpai.com\/de\/manus\/\">Manus<\/a> \u6240\u6709\u80fd\u529b\uff0c\u5f00\u6e90\u590d\u73b0\u7684\u786e\u662f\u634e\u5e26\u624b\u7684\u4e8b\u3002<\/p><\/blockquote>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-28077\" title=\"OpenManus-RL\uff1a\u5fae\u8c03\u6a21\u578b\u4ee5\u5f3a\u5316\u667a\u80fd\u4f53\u4efb\u52a1\u5b8c\u6210\u80fd\u529b-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/3a3e508159f87fa.jpg\" alt=\"OpenManus-RL\uff1a\u5fae\u8c03\u6a21\u578b\u4ee5\u5f3a\u5316\u667a\u80fd\u4f53\u4efb\u52a1\u5b8c\u6210\u80fd\u529b-1\" width=\"741\" height=\"292\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/3a3e508159f87fa.jpg 741w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/3a3e508159f87fa-18x7.jpg 18w\" sizes=\"auto, (max-width: 741px) 100vw, 741px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u667a\u80fd\u4f53\u73af\u5883\u642d\u5efa<\/strong>\uff1a\u63d0\u4f9b\u5728\u7ebfRL\u8c03\u4f18\u7684\u667a\u80fd\u4f53\u73af\u5883\u914d\u7f6e\u5de5\u5177\u3002<\/li>\n<li><strong>\u8f68\u8ff9\u6570\u636e\u6536\u96c6<\/strong>\uff1a\u8fde\u63a5Deepseek-R1\u3001QwQ-32B\u7b49\u6a21\u578b\uff0c\u6536\u96c6\u590d\u6742\u4efb\u52a1\u7684\u884c\u4e3a\u6570\u636e\u3002<\/li>\n<li><strong>RL\u8c03\u4f18\u652f\u6301<\/strong>\uff1a\u652f\u6301\u5b9a\u5236\u5316\u667a\u80fd\u4f53\u884c\u4e3a\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u3002<\/li>\n<li><strong>\u57fa\u51c6\u6d4b\u8bd5\u96c6\u6210<\/strong>\uff1a\u5185\u7f6eWebShop\u3001GAIA\u3001OSWorld\u3001AgentBench\u7b49\u6d4b\u8bd5\u73af\u5883\u3002<\/li>\n<li><strong>\u591a\u6837\u5316\u7b56\u7565<\/strong>\uff1a\u96c6\u6210Tree-of-Thoughts\u3001Monte Carlo Tree Search\u7b49RL\u7b56\u7565\u3002<\/li>\n<li><strong>\u793e\u533a\u534f\u4f5c<\/strong>\uff1a\u5141\u8bb8\u63d0\u4ea4\u4ee3\u7801\u3001\u6570\u636e\u96c6\u7b49\uff0c\u91cd\u8981\u8d21\u732e\u8005\u53ef\u6210\u4e3a\u8bba\u6587\u5171\u540c\u4f5c\u8005\u3002<\/li>\n<li><strong>\u5b9e\u65f6\u8fdb\u5c55\u5206\u4eab<\/strong>\uff1a\u901a\u8fc7\u52a8\u6001\u66f4\u65b0\u5c55\u793aRL\u8c03\u4f18\u8fc7\u7a0b\u4e0e\u6210\u679c\u3002<\/li>\n<\/ul>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<h3>\u5b89\u88c5\u6d41\u7a0b<\/h3>\n<p>OpenManus-RL\u7684\u5b89\u88c5\u7b80\u5355\uff0c\u9002\u5408\u5177\u5907Python\u57fa\u7840\u7684\u7528\u6237\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u6b65\u9aa4\uff1a<\/p>\n<h4>1. \u521b\u5efaConda\u73af\u5883<\/h4>\n<p>\u4e3a\u907f\u514d\u4f9d\u8d56\u51b2\u7a81\uff0c\u5efa\u8bae\u4f7f\u7528Conda\uff1a<\/p>\n<pre><code>conda create -n openmanus-rl python=3.10  \r\nconda activate openmanus-rl\r\n<\/code><\/pre>\n<ul>\n<li>\u524d\u7f6e\u6761\u4ef6\uff1a\u9700\u5b89\u88c5Conda\uff0c\u53ef\u4ece<a href=\"https:\/\/www.anaconda.com\/\">Anaconda\u5b98\u7f51<\/a>\u4e0b\u8f7d\u3002<\/li>\n<li>\u6fc0\u6d3b\u540e\uff0c\u7ec8\u7aef\u663e\u793a<code>(openmanus-rl)<\/code>\u3002<\/li>\n<\/ul>\n<h4>2. \u514b\u9686\u9879\u76ee<\/h4>\n<p>\u786e\u4fdd\u5df2\u5b89\u88c5Git\uff08\u68c0\u67e5\uff1a<code>git --version<\/code>\uff0c\u672a\u5b89\u88c5\u53ef\u4ece<a href=\"https:\/\/git-scm.com\/\">git-scm.com<\/a>\u4e0b\u8f7d\uff09\uff1a<\/p>\n<pre><code>git clone https:\/\/github.com\/OpenManus\/OpenManus-RL.git  \r\ncd OpenManus-RL\r\n<\/code><\/pre>\n<ul>\n<li>\u4e0b\u8f7d\u4ee3\u7801\u5e76\u8fdb\u5165\u9879\u76ee\u76ee\u5f55\u3002<\/li>\n<\/ul>\n<h4>3. \u5b89\u88c5\u4f9d\u8d56<\/h4>\n<p>\u5728\u9879\u76ee\u6839\u76ee\u5f55\u6267\u884c\uff1a<\/p>\n<pre><code>pip install -r requirements.txt\r\n<\/code><\/pre>\n<ul>\n<li>\u82e5\u4e0b\u8f7d\u6162\uff0c\u53ef\u7528\u56fd\u5185\u955c\u50cf\uff1a<\/li>\n<\/ul>\n<pre><code>pip install -r requirements.txt -i https:\/\/pypi.tuna.tsinghua.edu.cn\/simple\r\n<\/code><\/pre>\n<ul>\n<li>\u53ef\u89c6\u5316\u5de5\u5177\u9700\u989d\u5916\u5b89\u88c5\uff1a<\/li>\n<\/ul>\n<pre><code>pip install matplotlib numpy\r\n<\/code><\/pre>\n<h4>4. \u914d\u7f6e\u6a21\u578b\u4e0e\u6570\u636e\u96c6<\/h4>\n<ul>\n<li><strong>\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09<\/strong>\uff1a\u6307\u5b9a\u6a21\u578b\u548c\u6570\u636e\u96c6\uff1a<\/li>\n<\/ul>\n<pre><code>python -m openmanus_rl.sft --model_name_or_path Qwen\/Qwen2.5-1.5B-Instruct --dataset_name CharlieDreemur\/OpenManus-RL\r\n<\/code><\/pre>\n<ul>\n<li><strong>\u5f3a\u5316\u5b66\u4e60\u8c03\u4f18\uff08GRPO\uff09<\/strong>\uff1a\u914d\u7f6e\u5956\u52b1\u51fd\u6570\uff1a<\/li>\n<\/ul>\n<pre><code>python -m openmanus_rl.grpo --model_name_or_path Qwen\/Qwen2.5-1.5B-Instruct --dataset_name CharlieDreemur\/OpenManus-RL-GRPO --reward_funcs accuracy format tag_count\r\n<\/code><\/pre>\n<ul>\n<li>\u6570\u636e\u96c6\u53ef\u4ece<a href=\"https:\/\/huggingface.co\/datasets\/CharlieDreemur\/OpenManus-RL\">Huggingface<\/a>\u83b7\u53d6\u3002<\/li>\n<\/ul>\n<h4>5. \u8fd0\u884c\u9879\u76ee<\/h4>\n<ul>\n<li>\u5355GPU\u8fd0\u884cSFT\uff1a<\/li>\n<\/ul>\n<pre><code>python -m openmanus_rl.sft --output_dir data\/sft-output\r\n<\/code><\/pre>\n<ul>\n<li>\u591aGPU\u8fd0\u884cGRPO\uff08\u9700\u914d\u7f6e<code>zero3.yaml<\/code>\uff09\uff1a<\/li>\n<\/ul>\n<pre><code>accelerate launch --config_file=configs\/accelerate_configs\/zero3.yaml openmanus_rl\/grpo.py --output_dir data\/grpo-output\r\n<\/code><\/pre>\n<h3>\u4e3b\u8981\u529f\u80fd\u64cd\u4f5c\u6d41\u7a0b<\/h3>\n<h4>\u667a\u80fd\u4f53\u73af\u5883\u642d\u5efa<\/h4>\n<ul>\n<li><strong>\u64cd\u4f5c\u6b65\u9aa4<\/strong>\uff1a\n<ol>\n<li>\u8fd0\u884c<code>python -m openmanus_rl.sft<\/code>\u751f\u6210\u57fa\u7840\u73af\u5883\u3002<\/li>\n<li>\u4fee\u6539\u914d\u7f6e\u6587\u4ef6\uff08\u5982\u4efb\u52a1\u76ee\u6807\u6216\u5956\u52b1\u51fd\u6570\uff09\u3002<\/li>\n<li>\u6267\u884c<code>python -m openmanus_rl.grpo<\/code>\u5f00\u59cb\u8c03\u4f18\u3002<\/li>\n<\/ol>\n<\/li>\n<li><strong>\u4f7f\u7528\u573a\u666f<\/strong>\uff1a\u4e3a\u7279\u5b9a\u4efb\u52a1\uff08\u5982\u8d2d\u7269\u51b3\u7b56\uff09\u5b9a\u5236\u667a\u80fd\u4f53\u73af\u5883\u3002<\/li>\n<\/ul>\n<h4>\u6570\u636e\u6536\u96c6\u4e0e\u6d4b\u8bd5<\/h4>\n<ul>\n<li><strong>\u64cd\u4f5c\u6b65\u9aa4<\/strong>\uff1a\n<ol>\n<li>\u914d\u7f6e\u6a21\u578b\uff08\u5982Deepseek-R1\uff09\uff1a<\/li>\n<\/ol>\n<\/li>\n<\/ul>\n<pre><code>python -m openmanus_rl.grpo --model_name_or_path <a href=\"https:\/\/www.kdjingpai.com\/de\/deepseek-r1nenglixiang\/\">Deepseek-R1<\/a>\r\n<\/code><\/pre>\n<ol start=\"2\">\n<li>\u8fd0\u884c\u6d4b\u8bd5\uff1a<code>--benchmark GAIA<\/code>\uff0c\u7ed3\u679c\u4fdd\u5b58\u81f3<code>data\/<\/code>\u76ee\u5f55\u3002<\/li>\n<\/ol>\n<ul>\n<li><strong>\u4f7f\u7528\u573a\u666f<\/strong>\uff1a\u5206\u6790\u667a\u80fd\u4f53\u5728\u590d\u6742\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002<\/li>\n<\/ul>\n<h4>RL\u8c03\u4f18\u64cd\u4f5c<\/h4>\n<ul>\n<li><strong>\u64cd\u4f5c\u6b65\u9aa4<\/strong>\uff1a\n<ol>\n<li>\u8fd0\u884cGRPO\u6a21\u5f0f\uff1a<\/li>\n<\/ol>\n<\/li>\n<\/ul>\n<pre><code>python -m openmanus_rl.grpo --reward_funcs accuracy\r\n<\/code><\/pre>\n<ol start=\"2\">\n<li>\u67e5\u770b\u8bad\u7ec3\u65e5\u5fd7\uff0c\u6a21\u578b\u4fdd\u5b58\u81f3<code>data\/grpo-output<\/code>\u3002<\/li>\n<\/ol>\n<ul>\n<li><strong>\u4f7f\u7528\u573a\u666f<\/strong>\uff1a\u4f18\u5316\u667a\u80fd\u4f53\u884c\u4e3a\uff0c\u5982\u63d0\u5347WebShop\u8d2d\u4e70\u6210\u529f\u7387\u3002<\/li>\n<\/ul>\n<h4>\u793e\u533a\u8d21\u732e<\/h4>\n<ul>\n<li><strong>\u64cd\u4f5c\u6b65\u9aa4<\/strong>\uff1a\n<ol>\n<li>Fork\u9879\u76ee\u81f3\u4e2a\u4ebaGitHub\u8d26\u6237\u3002<\/li>\n<li>\u672c\u5730\u4fee\u6539\u5e76\u63d0\u4ea4\uff1a<\/li>\n<\/ol>\n<\/li>\n<\/ul>\n<pre><code>git add .  \r\ngit commit -m \"\u4f18\u5316RL\u7b56\u7565\"  \r\ngit push origin main\r\n<\/code><\/pre>\n<ol start=\"3\">\n<li>\u63d0\u4ea4Pull Request\uff0c\u6216\u90ae\u4ef6\u8054\u7cfb<code>kunlunz2@illinois.edu<\/code>\u3002<\/li>\n<\/ol>\n<ul>\n<li><strong>\u4f7f\u7528\u573a\u666f<\/strong>\uff1a\u8d21\u732e\u65b0\u7b97\u6cd5\u6216\u6570\u636e\u96c6\uff0c\u53c2\u4e0e\u6838\u5fc3\u5f00\u53d1\u3002<\/li>\n<\/ul>\n<h3>\u7279\u8272\u529f\u80fd\u8be6\u89e3<\/h3>\n<h4>RL\u8c03\u4f18\u652f\u6301<\/h4>\n<ul>\n<li><strong>\u5982\u4f55\u64cd\u4f5c<\/strong>\uff1a\u8fd0\u884cGRPO\uff0c\u6307\u5b9a\u5956\u52b1\u51fd\u6570\uff08\u5982<code>accuracy<\/code>\uff09\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u663e\u793a\u5b9e\u65f6\u65e5\u5fd7\uff0c\u5b8c\u6210\u540e\u6a21\u578b\u4fdd\u5b58\u81f3\u6307\u5b9a\u76ee\u5f55\u3002<\/li>\n<li><strong>\u6548\u679c<\/strong>\uff1a\u667a\u80fd\u4f53\u80fd\u6839\u636e\u4efb\u52a1\u8c03\u6574\u884c\u4e3a\uff0c\u4f8b\u5982\u5728OSWorld\u4e2d\u4f18\u5316\u591a\u6a21\u6001\u4efb\u52a1\u8868\u73b0\u3002<\/li>\n<\/ul>\n<h4>\u57fa\u51c6\u6d4b\u8bd5\u96c6\u6210<\/h4>\n<ul>\n<li><strong>\u5982\u4f55\u64cd\u4f5c<\/strong>\uff1a\u8fd0\u884c<code>python -m openmanus_rl.grpo --benchmark AgentBench<\/code>\uff0c\u7cfb\u7edf\u81ea\u52a8\u751f\u6210\u6210\u529f\u7387\u3001\u54cd\u5e94\u65f6\u95f4\u7b49\u62a5\u544a\u3002<\/li>\n<li><strong>\u6548\u679c<\/strong>\uff1a\u63d0\u4f9b\u91cf\u5316\u6307\u6807\uff0c\u5e2e\u52a9\u5f00\u53d1\u8005\u5bf9\u6bd4\u6a21\u578b\u6027\u80fd\u3002<\/li>\n<\/ul>\n<h4>\u591a\u6837\u5316\u7b56\u7565<\/h4>\n<ul>\n<li><strong>\u5982\u4f55\u64cd\u4f5c<\/strong>\uff1a\u5728\u914d\u7f6e\u6587\u4ef6\u4e2d\u9009\u62e9\u7b56\u7565\uff08\u5982<code>Tree-of-Thoughts<\/code>\uff09\uff0c\u8fd0\u884c\u8c03\u4f18\u547d\u4ee4\u6d4b\u8bd5\u6548\u679c\u3002<\/li>\n<li><strong>\u6548\u679c<\/strong>\uff1a\u63d0\u5347\u667a\u80fd\u4f53\u5728\u957f\u7a0b\u89c4\u5212\u4efb\u52a1\u4e2d\u7684\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<\/ul>\n<p>OpenManus-RL\u901a\u8fc7\u4ee5\u4e0a\u529f\u80fd\u5e2e\u52a9\u7528\u6237\u5feb\u901f\u4e0a\u624b\u3002\u9879\u76ee\u8fd8\u63d0\u4f9b\u793e\u533a\u7fa4\u7ec4\uff08\u89c1GitHub\u201cCommunity Group\u201d\uff09\uff0c\u52a0\u5165\u540e\u53ef\u4e0e\u5f00\u53d1\u8005\u4ea4\u6d41\uff0c\u83b7\u53d6\u6700\u65b0\u8d44\u8baf\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>OpenManus-RL\u662f\u7531UIUC-Ulab\u4e0e MetaGPT \u793e\u533a\u7684OpenManus\u56e2\u961f\u8054\u5408\u5f00\u53d1\u7684\u5f00\u6e90\u9879\u76ee\uff0c\u6258\u7ba1\u4e8eGitHub\u3002\u8be5\u9879\u76ee\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6280\u672f\u63d0\u5347\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u667a\u80fd\u4f53\u7684\u63a8\u7406\u4e0e\u51b3\u7b56\u80fd\u529b\uff0c\u57fa\u4e8eDeepseek-R1&#8230;<\/p>\n","protected":false},"author":1,"featured_media":62006,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230,365],"class_list":["post-28076","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu","tag-damoxingweidiao"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/28076","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=28076"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/28076\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media\/62006"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=28076"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=28076"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=28076"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}