{"id":29726,"date":"2025-04-01T19:48:55","date_gmt":"2025-04-01T11:48:55","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=29726"},"modified":"2025-07-01T02:40:59","modified_gmt":"2025-06-30T18:40:59","slug":"search-r1","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/search-r1\/","title":{"rendered":"Search-R1\uff1a\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u5927\u6a21\u578b\u641c\u7d22\u4e0e\u63a8\u7406\u7684\u5de5\u5177"},"content":{"rendered":"<p>Search-R1 \u662f\u4e00\u4e2a\u5f00\u6e90\u9879\u76ee\uff0c\u7531 PeterGriffinJin \u5728 GitHub \u4e0a\u5f00\u53d1\uff0c\u57fa\u4e8e veRL \u6846\u67b6\u6784\u5efa\u3002\u5b83\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6280\u672f\u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\uff0c\u8ba9\u6a21\u578b\u81ea\u4e3b\u5b66\u4f1a\u63a8\u7406\u548c\u8c03\u7528\u641c\u7d22\u5f15\u64ce\u89e3\u51b3\u95ee\u9898\u3002\u9879\u76ee\u652f\u6301 Qwen2.5-3B \u548c Llama3.2-3B \u7b49\u57fa\u7840\u6a21\u578b\uff0c\u6269\u5c55\u4e86 <a href=\"https:\/\/www.kdjingpai.com\/de\/deepseek-r1nenglixiang\/\">DeepSeek-R1<\/a> \u548c TinyZero \u7684\u65b9\u6cd5\u3002\u7528\u6237\u53ef\u4ee5\u7528\u5b83\u8bad\u7ec3\u6a21\u578b\u5904\u7406\u5355\u8f6e\u6216\u591a\u8f6e\u4efb\u52a1\uff0c\u63d0\u4f9b\u4ee3\u7801\u3001\u6570\u636e\u96c6\u548c\u5b9e\u9a8c\u65e5\u5fd7\u3002\u5b98\u65b9<a href=\"https:\/\/arxiv.org\/pdf\/2503.09516\">\u8bba\u6587<\/a>\u53d1\u5e03\u4e8e 2025 \u5e74 3 \u6708\uff0c\u9879\u76ee\u6a21\u578b\u548c\u6570\u636e\u53ef\u5728 Hugging Face \u4e0b\u8f7d\uff0c\u9002\u5408\u7814\u7a76\u4eba\u5458\u548c\u5f00\u53d1\u8005\u4f7f\u7528\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-29728\" title=\"Search-R1\uff1a\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u5927\u6a21\u578b\u641c\u7d22\u4e0e\u63a8\u7406\u7684\u5de5\u5177-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/b5b661bea9fc530.jpg\" alt=\"Search-R1\uff1a\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u5927\u6a21\u578b\u641c\u7d22\u4e0e\u63a8\u7406\u7684\u5de5\u5177-1\" width=\"687\" height=\"750\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/b5b661bea9fc530.jpg 824w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/b5b661bea9fc530-768x839.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/b5b661bea9fc530-11x12.jpg 11w\" sizes=\"auto, (max-width: 687px) 100vw, 687px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li>\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u5927\u6a21\u578b\uff0c\u63d0\u5347\u63a8\u7406\u548c\u641c\u7d22\u80fd\u529b\u3002<\/li>\n<li>\u652f\u6301\u8c03\u7528\u8c37\u6b4c\u3001\u5fc5\u5e94\u3001Brave \u7b49\u641c\u7d22\u5f15\u64ce API\u3002<\/li>\n<li>\u63d0\u4f9b LoRA \u8c03\u4f18\u548c\u76d1\u7763\u5fae\u8c03\u529f\u80fd\uff0c\u4f18\u5316\u6a21\u578b\u8868\u73b0\u3002<\/li>\n<li>\u5185\u7f6e\u73b0\u6210\u91cd\u6392\u5e8f\u5668\uff0c\u63d0\u9ad8\u641c\u7d22\u7ed3\u679c\u51c6\u786e\u6027\u3002<\/li>\n<li>\u5305\u542b\u8be6\u7ec6\u5b9e\u9a8c\u65e5\u5fd7\u548c\u8bba\u6587\uff0c\u652f\u6301\u7ed3\u679c\u590d\u73b0\u3002<\/li>\n<li>\u63d0\u4f9b\u672c\u5730\u68c0\u7d22\u670d\u52a1\u5668\u529f\u80fd\uff0c\u65b9\u4fbf\u81ea\u5b9a\u4e49\u641c\u7d22\u3002<\/li>\n<li>\u652f\u6301\u7528\u6237\u4e0a\u4f20\u81ea\u5b9a\u4e49\u6570\u636e\u96c6\u548c\u8bed\u6599\u5e93\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<p>Search-R1 \u9762\u5411\u6709\u7f16\u7a0b\u548c\u673a\u5668\u5b66\u4e60\u57fa\u7840\u7684\u7528\u6237\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u5b89\u88c5\u548c\u4f7f\u7528\u6307\u5357\uff0c\u8ba9\u4f60\u5feb\u901f\u4e0a\u624b\u3002<\/p>\n<h3>\u5b89\u88c5\u6d41\u7a0b<\/h3>\n<p>\u4f7f\u7528 Search-R1 \u9700\u8981\u5148\u642d\u5efa\u73af\u5883\u3002\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li><strong>\u521b\u5efa Search-R1 \u73af\u5883<\/strong><br \/>\n\u5728\u7ec8\u7aef\u8fd0\u884c\uff1a<\/li>\n<\/ol>\n<pre><code>conda create -n searchr1 python=3.9\r\nconda activate searchr1\r\n<\/code><\/pre>\n<p>\u8fd9\u4f1a\u521b\u5efa\u4e00\u4e2a Python 3.9 \u7684\u865a\u62df\u73af\u5883\u3002<\/p>\n<ol start=\"2\">\n<li><strong>\u5b89\u88c5 PyTorch<\/strong><br \/>\n\u8f93\u5165\u4ee5\u4e0b\u547d\u4ee4\u5b89\u88c5 PyTorch 2.4.0\uff08\u652f\u6301 CUDA 12.1\uff09\uff1a<\/li>\n<\/ol>\n<pre><code>pip install torch==2.4.0 --index-url https:\/\/download.pytorch.org\/whl\/cu121\r\n<\/code><\/pre>\n<ol start=\"3\">\n<li><strong>\u5b89\u88c5 vLLM<\/strong><br \/>\n<a href=\"https:\/\/www.kdjingpai.com\/de\/vllm\/\">vLLM<\/a> \u662f\u8fd0\u884c\u5927\u6a21\u578b\u7684\u5173\u952e\u5e93\uff0c\u5b89\u88c5 0.6.3 \u7248\u672c\uff1a<\/li>\n<\/ol>\n<pre><code>pip3 install vllm==0.6.3\r\n<\/code><\/pre>\n<p>\u4e5f\u53ef\u9009\u62e9 0.5.4\u30010.4.2 \u6216 0.3.1 \u7248\u672c\u3002<\/p>\n<ol start=\"4\">\n<li><strong>\u5b89\u88c5 veRL<\/strong><br \/>\n\u5728\u9879\u76ee\u6839\u76ee\u5f55\u8fd0\u884c\uff1a<\/li>\n<\/ol>\n<pre><code>pip install -e .\r\n<\/code><\/pre>\n<p>\u8fd9\u4f1a\u5b89\u88c5 veRL \u6846\u67b6\u3002<\/p>\n<ol start=\"5\">\n<li><strong>\u5b89\u88c5\u53ef\u9009\u4f9d\u8d56<\/strong><br \/>\n\u4e3a\u63d0\u5347\u6027\u80fd\uff0c\u5b89\u88c5 Flash Attention \u548c Wandb\uff1a<\/li>\n<\/ol>\n<pre><code>pip3 install flash-attn --no-build-isolation\r\npip install wandb\r\n<\/code><\/pre>\n<ol start=\"6\">\n<li><strong>\u5b89\u88c5\u68c0\u7d22\u5668\u73af\u5883\uff08\u53ef\u9009\uff09<\/strong><br \/>\n\u5982\u679c\u9700\u8981\u672c\u5730\u68c0\u7d22\u670d\u52a1\u5668\uff0c\u53e6\u5efa\u73af\u5883\uff1a<\/li>\n<\/ol>\n<pre><code>conda create -n retriever python=3.10\r\nconda activate retriever\r\nconda install pytorch==2.4.0 pytorch-cuda=12.1 -c pytorch -c nvidia\r\npip install transformers datasets\r\nconda install -c pytorch -c nvidia faiss-gpu=1.8.0\r\npip install uvicorn fastapi\r\n<\/code><\/pre>\n<h3>\u5feb\u901f\u5f00\u59cb<\/h3>\n<p>\u4ee5\u4e0b\u662f\u57fa\u4e8e NQ \u6570\u636e\u96c6\u8bad\u7ec3\u6a21\u578b\u7684\u6b65\u9aa4\uff1a<\/p>\n<ol>\n<li><strong>\u4e0b\u8f7d\u7d22\u5f15\u548c\u8bed\u6599<\/strong><br \/>\n\u8bbe\u7f6e\u4fdd\u5b58\u8def\u5f84\u5e76\u8fd0\u884c\uff1a<\/li>\n<\/ol>\n<pre><code>save_path=\/\u4f60\u7684\u4fdd\u5b58\u8def\u5f84\r\npython scripts\/download.py --save_path $save_path\r\ncat $save_path\/part_* &gt; $save_path\/e5_Flat.index\r\ngzip -d $save_path\/wiki-18.jsonl.gz\r\n<\/code><\/pre>\n<ol start=\"2\">\n<li><strong>\u5904\u7406 NQ \u6570\u636e<\/strong><br \/>\n\u8fd0\u884c\u811a\u672c\u751f\u6210\u8bad\u7ec3\u6570\u636e\uff1a<\/li>\n<\/ol>\n<pre><code>python scripts\/data_process\/nq_search.py\r\n<\/code><\/pre>\n<ol start=\"3\">\n<li><strong>\u542f\u52a8\u68c0\u7d22\u670d\u52a1\u5668<\/strong><br \/>\n\u5728\u68c0\u7d22\u5668\u73af\u5883\u8fd0\u884c\uff1a<\/li>\n<\/ol>\n<pre><code>conda activate retriever\r\nbash retrieval_launch.sh\r\n<\/code><\/pre>\n<ol start=\"4\">\n<li><strong>\u8fd0\u884c RL \u8bad\u7ec3<\/strong><br \/>\n\u5728 Search-R1 \u73af\u5883\u8fd0\u884c\uff1a<\/li>\n<\/ol>\n<pre><code>conda activate searchr1\r\nbash train_ppo.sh\r\n<\/code><\/pre>\n<p>\u8fd9\u4f1a\u7528 Llama-3.2-3B \u57fa\u7840\u6a21\u578b\u8fdb\u884c PPO \u8bad\u7ec3\u3002<\/p>\n<h3>\u4f7f\u7528\u81ea\u5b9a\u4e49\u6570\u636e\u96c6<\/h3>\n<ol>\n<li><strong>\u51c6\u5907 QA \u6570\u636e<\/strong><br \/>\n\u6570\u636e\u9700\u4e3a JSONL \u683c\u5f0f\uff0c\u6bcf\u884c\u5305\u542b\u4ee5\u4e0b\u5b57\u6bb5\uff1a<\/li>\n<\/ol>\n<pre><code>{\r\n\"data_source\": \"web\",\r\n\"prompt\": [{\"role\": \"user\", \"content\": \"\u95ee\u9898\"}],\r\n\"ability\": \"fact-reasoning\",\r\n\"reward_model\": {\"style\": \"rule\", \"ground_truth\": \"\u7b54\u6848\"},\r\n\"extra_info\": {\"split\": \"train\", \"index\": 1}\r\n}\r\n<\/code><\/pre>\n<p>\u53c2\u8003\u00a0<code>&lt;scripts\/data_process\/nq_search.py&gt;<\/code>\u3002<\/p>\n<ol start=\"2\">\n<li><strong>\u51c6\u5907\u8bed\u6599\u5e93<\/strong><br \/>\n\u8bed\u6599\u9700\u4e3a JSONL \u683c\u5f0f\uff0c\u6bcf\u884c\u5305\u542b\u00a0<code>id<\/code>\u00a0\u548c\u00a0<code>contents<\/code>\uff0c\u5982\uff1a<\/p>\n<pre><code>{\"id\": \"0\", \"contents\": \"\u6587\u672c\u5185\u5bb9\"}\r\n<\/code><\/pre>\n<p>\u53ef\u53c2\u8003\u00a0<code>&lt;example\/corpus.jsonl&gt;<\/code>\u3002<\/li>\n<li><strong>\u7d22\u5f15\u8bed\u6599\uff08\u53ef\u9009\uff09<\/strong><br \/>\n\u5982\u679c\u4f7f\u7528\u672c\u5730\u68c0\u7d22\uff0c\u8fd0\u884c\uff1a<\/p>\n<pre><code>bash search_r1\/search\/build_index.sh\r\n<\/code><\/pre>\n<\/li>\n<\/ol>\n<h3>\u8c03\u7528\u81ea\u5b9a\u4e49\u641c\u7d22\u5f15\u64ce<\/h3>\n<ol>\n<li>\u4fee\u6539\u00a0<code>&lt;search_r1\/search\/retriever_server.py&gt;<\/code>\uff0c\u914d\u7f6e API\u3002<\/li>\n<li>\u542f\u52a8\u670d\u52a1\u5668\uff1a\n<pre><code>python search_r1\/search\/retriever_server.py\r\n<\/code><\/pre>\n<\/li>\n<li>\u6a21\u578b\u4f1a\u901a\u8fc7\u00a0<code>http:\/\/127.0.0.1:8000\/retrieve<\/code>\u00a0\u8c03\u7528\u641c\u7d22\u3002<\/li>\n<\/ol>\n<h3>\u63a8\u7406\u64cd\u4f5c<\/h3>\n<ol>\n<li>\u542f\u52a8\u68c0\u7d22\u670d\u52a1\u5668\uff1a\n<pre><code>bash retrieval_launch.sh\r\n<\/code><\/pre>\n<\/li>\n<li>\u8fd0\u884c\u63a8\u7406\uff1a\n<pre><code>python infer.py\r\n<\/code><\/pre>\n<\/li>\n<li>\u4fee\u6539\u00a0<code>&lt;infer.py&gt;<\/code>\u00a0\u7b2c 7 \u884c\u7684\u00a0<code>question<\/code>\uff0c\u8f93\u5165\u4f60\u60f3\u95ee\u7684\u95ee\u9898\u3002<\/li>\n<\/ol>\n<h3>\u6ce8\u610f\u4e8b\u9879<\/h3>\n<ul>\n<li>\u8bad\u7ec3\u9700\u8981 GPU\uff0c\u81f3\u5c11 24GB \u663e\u5b58\uff08\u5982 NVIDIA A100\uff09\u3002<\/li>\n<li>\u786e\u4fdd API \u5bc6\u94a5\u6709\u6548\uff0c\u7f51\u7edc\u8fde\u63a5\u7a33\u5b9a\u3002<\/li>\n<li>\u5b98\u65b9\u8bba\u6587\u548c\u5b9e\u9a8c\u65e5\u5fd7\uff08<code>&lt;Full experiment log 1&gt;<\/code>\u00a0\u548c\u00a0<code>&lt;Full experiment log 2&gt;<\/code>\uff09\u63d0\u4f9b\u66f4\u591a\u7ec6\u8282\u3002<\/li>\n<\/ul>\n<p>\u901a\u8fc7\u8fd9\u4e9b\u6b65\u9aa4\uff0c\u4f60\u53ef\u4ee5\u7528 Search-R1 \u8bad\u7ec3\u4e00\u4e2a\u80fd\u63a8\u7406\u548c\u641c\u7d22\u7684\u6a21\u578b\uff0c\u5904\u7406\u5404\u79cd\u4efb\u52a1\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u5e94\u7528\u573a\u666f<\/h2>\n<ol>\n<li><strong>\u79d1\u7814\u5b9e\u9a8c<\/strong><br \/>\n\u7814\u7a76\u4eba\u5458\u53ef\u4ee5\u7528 Search-R1 \u590d\u73b0\u8bba\u6587\u7ed3\u679c\uff0c\u63a2\u7d22\u5f3a\u5316\u5b66\u4e60\u5728\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u5e94\u7528\u3002<\/li>\n<li><strong>\u667a\u80fd\u52a9\u624b\u5f00\u53d1<\/strong><br \/>\n\u5f00\u53d1\u8005\u53ef\u4ee5\u8bad\u7ec3\u6a21\u578b\uff0c\u96c6\u6210\u5230\u804a\u5929\u5de5\u5177\u4e2d\uff0c\u63d0\u4f9b\u641c\u7d22\u548c\u63a8\u7406\u529f\u80fd\u3002<\/li>\n<li><strong>\u77e5\u8bc6\u67e5\u8be2<\/strong><br \/>\n\u7528\u6237\u53ef\u4ee5\u7528\u5b83\u5feb\u901f\u56de\u7b54\u590d\u6742\u95ee\u9898\uff0c\u901a\u8fc7\u641c\u7d22\u83b7\u53d6\u6700\u65b0\u4fe1\u606f\u3002<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h2>QA<\/h2>\n<ol>\n<li><strong>Search-R1 \u652f\u6301\u54ea\u4e9b\u6a21\u578b\uff1f<\/strong><br \/>\n\u76ee\u524d\u652f\u6301 Qwen2.5-3B \u548c Llama3.2-3B \u57fa\u7840\u6a21\u578b\uff0c\u5176\u4ed6\u6a21\u578b\u9700\u81ea\u884c\u9002\u914d\u3002<\/li>\n<li><strong>\u8bad\u7ec3\u9700\u8981\u591a\u957f\u65f6\u95f4\uff1f<\/strong><br \/>\n\u89c6\u6570\u636e\u96c6\u548c\u786c\u4ef6\u800c\u5b9a\uff0c\u5728 24GB \u663e\u5b58 GPU \u4e0a\uff0cNQ \u6570\u636e\u96c6\u8bad\u7ec3\u7ea6\u9700\u6570\u5c0f\u65f6\u3002<\/li>\n<li><strong>\u5982\u4f55\u9a8c\u8bc1\u8bad\u7ec3\u6548\u679c\uff1f<\/strong><br \/>\n\u67e5\u770b\u00a0<code>&lt;Preliminary results&gt;<\/code>\u00a0\u4e2d\u7684\u56fe\u8868\uff0c\u6216\u68c0\u67e5 Wandb \u65e5\u5fd7\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>Search-R1 \u662f\u4e00\u4e2a\u5f00\u6e90\u9879\u76ee\uff0c\u7531 PeterGriffinJin \u5728 GitHub \u4e0a\u5f00\u53d1\uff0c\u57fa\u4e8e veRL \u6846\u67b6\u6784\u5efa\u3002\u5b83\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6280\u672f\u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\uff0c\u8ba9\u6a21\u578b\u81ea\u4e3b\u5b66\u4f1a\u63a8\u7406\u548c\u8c03\u7528\u641c\u7d22\u5f15\u64ce\u89e3\u51b3\u95ee\u9898\u3002\u9879\u76ee\u652f\u6301 Qwen2&#8230;.<\/p>\n","protected":false},"author":1,"featured_media":62182,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20,419,427],"tags":[230,365],"class_list":["post-29726","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","category-ai-learning","category-research-assistant","tag-aikaiyuanxiangmu","tag-damoxingweidiao"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/29726","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/comments?post=29726"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/29726\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media\/62182"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media?parent=29726"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/categories?post=29726"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/tags?post=29726"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}