{"id":27690,"date":"2025-03-05T11:20:25","date_gmt":"2025-03-05T03:20:25","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=27690"},"modified":"2025-07-14T10:36:36","modified_gmt":"2025-07-14T02:36:36","slug":"research","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/research\/","title":{"rendered":"ReSearch\uff1a\u5f3a\u5316\u641c\u7d22\u63a8\u7406\u80fd\u529b\u7684 Qwen2.5-7B \u6a21\u578b\uff08\u5b9e\u9a8c\uff09"},"content":{"rendered":"<p>ReSearch \u662f\u4e00\u4e2a\u5f00\u6e90\u7814\u7a76\u5de5\u5177\uff0c\u7531 Agent-RL \u56e2\u961f\u5f00\u53d1\uff0c\u65e8\u5728\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning, RL\uff09\u63d0\u5347\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u641c\u7d22\u548c\u63a8\u7406\u80fd\u529b\u3002\u9879\u76ee\u7075\u611f\u6765\u6e90\u4e8e Deepseek-R1-Zero \u548c OpenAI \u7684 Deep Research\uff0c\u57fa\u4e8e Qwen2.5-7B \u6a21\u578b\uff0c\u5229\u7528 GRPO\uff08Generalized Reward Policy Optimization\uff09\u65b9\u6cd5\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\uff0c\u4f7f\u6a21\u578b\u4ec5\u4f9d\u9760\u5956\u52b1\u4fe1\u53f7\u81ea\u4e3b\u8c03\u7528\u641c\u7d22\u5de5\u5177\uff0c\u65e0\u9700\u76d1\u7763\u6570\u636e\u3002ReSearch \u5728 HotpotQA \u6570\u636e\u96c6\u4e0a\u9a8c\u8bc1\u4e86\u6548\u679c\uff0c\u5e76\u80fd\u6cdb\u5316\u81f3 Bamboogle \u548c StrategyQA \u7b49\u6570\u636e\u96c6\u3002\u6258\u7ba1\u4e8e GitHub\uff0c\u63d0\u4f9b\u5b8c\u6574\u4ee3\u7801\u548c\u5b9e\u9a8c\u6587\u6863\uff0c\u9002\u5408\u7814\u7a76\u4eba\u5458\u590d\u73b0\u6216\u6269\u5c55 RL \u4e0e LLM \u7ed3\u5408\u7684\u63a2\u7d22\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-27691\" title=\"ReSearch\uff1a\u5f3a\u5316\u641c\u7d22\u63a8\u7406\u80fd\u529b\u7684 Qwen2.5-7B \u6a21\u578b\uff08\u5b9e\u9a8c\uff09-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/7a9d3e499cf932f.jpg\" alt=\"ReSearch\uff1a\u5f3a\u5316\u641c\u7d22\u63a8\u7406\u80fd\u529b\u7684 Qwen2.5-7B \u6a21\u578b\uff08\u5b9e\u9a8c\uff09-1\" width=\"3034\" height=\"908\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/7a9d3e499cf932f.jpg 3034w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/7a9d3e499cf932f-768x230.jpg 768w\" sizes=\"auto, (max-width: 3034px) 100vw, 3034px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u7ba1\u9053<\/strong>\uff1a\u652f\u6301\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\u5927\u6a21\u578b\uff0c\u63d0\u4f9b\u5b8c\u6574\u7684\u53c2\u6570\u914d\u7f6e\u548c\u5956\u52b1\u4fe1\u53f7\u8bbe\u8ba1\u3002<\/li>\n<li><strong>\u641c\u7d22\u5de5\u5177\u8c03\u7528<\/strong>\uff1a\u6a21\u578b\u53ef\u6839\u636e\u95ee\u9898\u81ea\u52a8\u8c03\u7528\u641c\u7d22\u5de5\u5177\uff0c\u63d0\u5347\u590d\u6742\u63a8\u7406\u4efb\u52a1\u7684\u51c6\u786e\u6027\u3002<\/li>\n<li><strong>\u591a\u6570\u636e\u96c6\u9002\u914d<\/strong>\uff1a\u5728 HotpotQA \u8bad\u7ec3\u540e\uff0c\u80fd\u6269\u5c55\u81f3 Bamboogle\u3001StrategyQA \u7b49\u6570\u636e\u96c6\u3002<\/li>\n<li><strong>\u6027\u80fd\u8bc4\u4f30\u652f\u6301<\/strong>\uff1a\u96c6\u6210 FlashRAG \u73af\u5883\uff0c\u5feb\u901f\u6d4b\u8bd5\u6a21\u578b\u5728\u5f00\u53d1\u96c6\u4e0a\u7684\u8868\u73b0\u3002<\/li>\n<li><strong>\u5f00\u6e90\u5b9e\u73b0<\/strong>\uff1a\u63d0\u4f9b\u8be6\u7ec6\u4ee3\u7801\u548c\u5b9e\u9a8c\u914d\u7f6e\uff0c\u4fbf\u4e8e\u7814\u7a76\u590d\u73b0\u548c\u4e8c\u6b21\u5f00\u53d1\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<h3>\u5b89\u88c5\u6d41\u7a0b<\/h3>\n<p>ReSearch \u9700\u8981\u5728 GPU \u73af\u5883\u4e0b\u8fd0\u884c\uff0c\u4f9d\u8d56 verl \u548c FlashRAG \u6846\u67b6\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u5b89\u88c5\u6b65\u9aa4\uff1a<\/p>\n<h4>1. \u73af\u5883\u51c6\u5907<\/h4>\n<ul>\n<li><strong>\u7cfb\u7edf\u8981\u6c42<\/strong>\uff1a\u63a8\u8350 Linux\uff08\u5982 Ubuntu\uff09\uff0cWindows \u53ef\u80fd\u5b58\u5728\u517c\u5bb9\u6027\u95ee\u9898\u3002<\/li>\n<li><strong>Python \u7248\u672c<\/strong>\uff1a\u9700 Python 3.11 \u6216\u4ee5\u4e0a\u3002<\/li>\n<li><strong>GPU \u914d\u7f6e<\/strong>\uff1a\u652f\u6301 NVIDIA GPU\uff0c\u5b89\u88c5 CUDA 12.4\uff08\u4e0e torch \u7248\u672c\u5339\u914d\uff09\u3002<\/li>\n<\/ul>\n<h4>2. \u514b\u9686\u4ed3\u5e93<\/h4>\n<p>\u5728\u7ec8\u7aef\u8f93\u5165\u4ee5\u4e0b\u547d\u4ee4\uff1a<\/p>\n<pre><code>git clone https:\/\/github.com\/Agent-RL\/ReSearch.git\r\ncd ReSearch\r\n<\/code><\/pre>\n<h4>3. \u5b89\u88c5 verl \u73af\u5883<\/h4>\n<p>ReSearch \u57fa\u4e8e verl \u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\uff0c\u5b89\u88c5\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n<pre><code>cd verl\r\npip3 install -e .\r\ncd ..\r\n<\/code><\/pre>\n<ul>\n<li><strong>\u4f9d\u8d56\u7248\u672c<\/strong>\uff1atorch==2.4.0+cu124\uff0cvllm==0.6.3\uff0cray==2.10.0\u3002\u82e5\u6709\u51b2\u7a81\uff0c\u624b\u52a8\u5b89\u88c5\uff1a\n<pre><code>pip install torch==2.4.0+cu124 vllm==0.6.3 ray==2.10.0\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<h4>4. \u5b89\u88c5 FlashRAG \u73af\u5883<\/h4>\n<p>FlashRAG \u7528\u4e8e\u8bc4\u4f30\u548c <a href=\"https:\/\/www.kdjingpai.com\/pt\/rag\/\">RAG<\/a> \u670d\u52a1\uff0c\u5b89\u88c5\u65b9\u6cd5\uff1a<\/p>\n<pre><code>git clone https:\/\/github.com\/RUC-AIBox\/FlashRAG.git\r\ncd FlashRAG\r\npip3 install -e .\r\ncd ..\/ReSearch\r\n<\/code><\/pre>\n<h4>5. \u4e0b\u8f7d\u9884\u8bad\u7ec3\u6a21\u578b<\/h4>\n<p>\u9ed8\u8ba4\u4f7f\u7528 Qwen2.5-7B\uff0c\u4ece Hugging Face \u4e0b\u8f7d\uff1a<\/p>\n<pre><code>git lfs install\r\ngit clone https:\/\/huggingface.co\/Qwen\/Qwen2.5-7B\r\n<\/code><\/pre>\n<p>\u8bb0\u5f55\u6a21\u578b\u8def\u5f84\uff0c\u7a0d\u540e\u914d\u7f6e\u4f7f\u7528\u3002<\/p>\n<h3>\u4f7f\u7528\u65b9\u6cd5<\/h3>\n<h4>\u542f\u52a8 RAG \u670d\u52a1<\/h4>\n<ol>\n<li><strong>\u4e0b\u8f7d\u9884\u7d22\u5f15\u6570\u636e<\/strong>\uff1a\u4ece FlashRAG \u6587\u6863\u83b7\u53d6 Wikipedia \u7d22\u5f15\u3001\u8bed\u6599\u5e93\u548c\u68c0\u7d22\u6a21\u578b\u3002<\/li>\n<li><strong>\u914d\u7f6e\u670d\u52a1<\/strong>\uff1a\u7f16\u8f91\u00a0<code>rag_serving\/serving_config.yaml<\/code>\uff0c\u586b\u5165\u68c0\u7d22\u6a21\u578b\u3001\u7d22\u5f15\u3001\u8bed\u6599\u5e93\u8def\u5f84\u53ca\u53ef\u7528 GPU ID\u3002<\/li>\n<li><strong>\u8fd0\u884c\u670d\u52a1<\/strong>\uff1a\n<pre><code>conda activate flashrag\r\npython rag_serving\/serving.py --config rag_serving\/serving_config.yaml --num_retriever 1 --port 8000\r\n<\/code><\/pre>\n<p>\u670d\u52a1\u8fd0\u884c\u540e\uff0c\u63d0\u4f9b\u641c\u7d22\u652f\u6301\u3002<\/li>\n<\/ol>\n<h4>\u8bad\u7ec3\u6a21\u578b<\/h4>\n<ol>\n<li><strong>\u51c6\u5907\u6570\u636e<\/strong>\uff1a\u4e0b\u8f7d HotpotQA \u6570\u636e\u96c6\uff0c\u8fd0\u884c\u9884\u5904\u7406\u811a\u672c\uff1a\n<pre><code>python training\/data_preprocess_hpqa.py\r\n<\/code><\/pre>\n<p>\u751f\u6210\u7684\u8bad\u7ec3\u548c\u5f00\u53d1\u6570\u636e\u4fdd\u5b58\u4e3a parquet \u683c\u5f0f\u3002<\/li>\n<li><strong>\u914d\u7f6e\u53c2\u6570<\/strong>\uff1a\u4fee\u6539\u00a0<code>training\/run.sh<\/code>\uff0c\u8bbe\u7f6e\u6a21\u578b\u8def\u5f84\u3001\u641c\u7d22 URL\u3001\u6570\u636e\u8def\u5f84\u7b49\u3002<\/li>\n<li><strong>\u542f\u52a8\u8bad\u7ec3<\/strong>\uff1a\n<pre><code>conda activate verl\r\nbash training\/run.sh --actor_model_path \/path\/to\/Qwen2.5-7B --search_url http:\/\/localhost:8000 --train_data_path data\/train.parquet --dev_data_path data\/dev.parquet --save_path runs\/\r\n<\/code><\/pre>\n<ul>\n<li>\u5355\u8282\u70b9 8 GPU \u8bad\u7ec3\uff0c\u591a\u8282\u70b9\u9700\u4f7f\u7528 ray \u8c03\u6574\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<h4>\u8bc4\u4f30\u6a21\u578b<\/h4>\n<ol>\n<li><strong>\u542f\u52a8\u6a21\u578b\u670d\u52a1<\/strong>\uff1a\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c\u4f7f\u7528 SGLang \u90e8\u7f72\u6a21\u578b\uff1a\n<pre><code>python3 -m sglang.launch_server --served-model-name research --model-path runs\/trained_model --tp 2 --context-length 8192 --port 80\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u8fd0\u884c\u8bc4\u4f30<\/strong>\uff1a\n<pre><code>python evaluation\/run_eval.py --config_path evaluation\/eval_config.yaml --method_name research --split dev --dataset_name hotpotqa\r\n<\/code><\/pre>\n<ul>\n<li>\u7ed3\u679c\u4fdd\u5b58\u81f3\u00a0<code>evaluation\/results\/<\/code>\uff0c\u652f\u6301\u5207\u6362\u6570\u636e\u96c6\uff08\u5982 Bamboogle\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<h4>\u7279\u8272\u529f\u80fd\u64cd\u4f5c<\/h4>\n<ul>\n<li><strong>\u641c\u7d22\u5de5\u5177\u8c03\u7528<\/strong>\uff1a\n<ul>\n<li>\u8bad\u7ec3\u540e\uff0c\u6a21\u578b\u53ef\u81ea\u52a8\u5224\u65ad\u662f\u5426\u8c03\u7528\u641c\u7d22\u5de5\u5177\u3002\u4f8b\u5982\uff0c\u8f93\u5165\u201cHow many moons does Jupiter have?\u201d\uff1a\n<pre><code>python inference.py --model_path runs\/trained_model --question \"How many moons does Jupiter have?\"\r\n<\/code><\/pre>\n<p>\u8f93\u51fa\u793a\u4f8b\uff1a<code>Jupiter has 95 known moons as of 2025.<\/code>\u3002<\/li>\n<li>\u6d41\u7a0b\uff1a\u6a21\u578b\u6839\u636e\u95ee\u9898\u751f\u6210\u641c\u7d22\u67e5\u8be2\uff0c\u8c03\u7528 RAG \u670d\u52a1\u83b7\u53d6\u4fe1\u606f\u540e\u63a8\u7406\u56de\u7b54\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u8de8\u6570\u636e\u96c6\u6cdb\u5316<\/strong>\uff1a\n<ul>\n<li>\u6d4b\u8bd5\u6a21\u578b\u5728 StrategyQA \u4e0a\u7684\u8868\u73b0\uff1a\n<pre><code>python evaluation\/run_eval.py --config_path evaluation\/eval_config.yaml --method_name research --split dev --dataset_name strategyqa\r\n<\/code><\/pre>\n<p>\u8f93\u51fa\u5305\u542b\u63a8\u7406\u8fc7\u7a0b\u548c\u7b54\u6848\uff0c\u9a8c\u8bc1\u6cdb\u5316\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>\u6ce8\u610f\u4e8b\u9879<\/h4>\n<ul>\n<li><strong>\u786c\u4ef6\u9700\u6c42<\/strong>\uff1a\u8bad\u7ec3\u9700 24GB \u4ee5\u4e0a\u663e\u5b58\uff0c\u8bc4\u4f30\u9700 16GB\u3002<\/li>\n<li><strong>\u65e5\u5fd7\u76d1\u63a7<\/strong>\uff1a\u4f7f\u7528 TensorBoard \u67e5\u770b\u8bad\u7ec3\u8fdb\u5c55\uff1a\n<pre><code>tensorboard --logdir runs\/\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u6545\u969c\u6392\u67e5<\/strong>\uff1a\u82e5\u62a5\u9519\uff0c\u68c0\u67e5\u4f9d\u8d56\u7248\u672c\u6216\u67e5\u770b GitHub Issues\u3002<\/li>\n<\/ul>\n<p>\u901a\u8fc7\u4ee5\u4e0a\u64cd\u4f5c\uff0c\u7528\u6237\u53ef\u5b8c\u6574\u590d\u73b0 ReSearch \u5b9e\u9a8c\uff0c\u63a2\u7d22\u5f3a\u5316\u5b66\u4e60\u4e0e\u5927\u6a21\u578b\u7684\u7ed3\u5408\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>ReSearch \u662f\u4e00\u4e2a\u5f00\u6e90\u7814\u7a76\u5de5\u5177\uff0c\u7531 Agent-RL \u56e2\u961f\u5f00\u53d1\uff0c\u65e8\u5728\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning, RL\uff09\u63d0\u5347\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u641c\u7d22\u548c\u63a8\u7406\u80fd\u529b\u3002\u9879\u76ee\u7075\u611f\u6765\u6e90\u4e8e Deepseek-R1-Zero \u548c &#8230;<\/p>\n","protected":false},"author":1,"featured_media":61971,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230,243],"class_list":["post-27690","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu","tag-aizhishikuyukefu"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/27690","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=27690"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/27690\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media\/61971"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=27690"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=27690"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=27690"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}