{"id":9952,"date":"2025-02-12T18:10:57","date_gmt":"2025-02-12T09:10:57","guid":{"rendered":"https:\/\/www.skyer9.pe.kr\/wordpress\/?p=9952"},"modified":"2025-02-12T19:09:06","modified_gmt":"2025-02-12T10:09:06","slug":"deepseek-%eb%85%bc%eb%ac%b8-%ec%a0%95%eb%a6%ac","status":"publish","type":"post","link":"https:\/\/www.skyer9.pe.kr\/wordpress\/?p=9952","title":{"rendered":"DeepSeek \ub17c\ubb38 \uc815\ub9ac"},"content":{"rendered":"<h1>DeepSeek \ub17c\ubb38 \uc815\ub9ac<\/h1>\n<p>\ud2c0\ub9b0 \ubd80\ubd84\uc774 \uc788\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<h2>DeepSeek-R1-Zero<\/h2>\n<ul>\n<li>\n<p>DeepSeek-V3-Base \ub97c \uae30\ubc18\uc73c\ub85c RL \ub9cc\uc744 \uc774\uc6a9\ud574 \ud559\uc2b5<\/p>\n<p>\ucc98\uc74c\ubd80\ud130 \uc2dc\uc791\ud558\uba74 \uc2dc\uac04\uacfc \ub3c8\uc774 \ub9ce\uc774 \ub4e4\uae30 \ub54c\ubb38\uc5d0 \uad50\uc0ac \ubaa8\ub378\ub85c\ubd80\ud130 \uc9c0\uc2dd \uc99d\ub958(Knowledge Distillation)\ud568<\/p>\n<p>[\ucd94\uce21] \uc5ec\ub7ec \ubc29\ubc95\uc774 \uc788\uc9c0\ub9cc \uc218\uc2ed\ub9cc\uac1c\uc758 \uc9c8\ubb38\uc744 \ub358\uc9c0\uace0 \ub2f5\ubcc0\uc744 \ubc1b\uc544, \uadf8 \ub370\uc774\ud0c0\ub85c \ud559\uc2b5\ud588\uc744 \ub4ef.<\/p>\n<\/li>\n<li>\n<p>\uc9c0\ub3c4\ud559\uc2b5 \uc0dd\ub7b5<\/p>\n<p>\uc9c0\ub3c4\ud559\uc2b5(\uc0ac\ub78c\uc774 \uc0dd\uc131\ud55c \uc9c8\ubb38\uacfc \uc815\ub2f5 \uc81c\uacf5) \uc740 \uc2dc\uac04\uacfc \ube44\uc6a9\uc774 \ub9ce\uc774 \ub4e4\uae30 \ub54c\ubb38\uc5d0 \uc0dd\ub825<\/p>\n<\/li>\n<li>\n<p>RL \ub9cc\uc744 \uc774\uc6a9\ud574 \ud559\uc2b5<\/p>\n<ul>\n<li>\uae30\uc874 \ubcf4\uc0c1\ubaa8\ub378 \uc0dd\ub7b5<\/li>\n<\/ul>\n<p>\ubaa8\ub378\uc774 \ub2f5\uc744 \ub0b4\uba74 \uc0ac\ub78c\uc774 \ub2f5\uc744 \ud3c9\uac00\ud558\uace0,<br \/>\n\uadf8\uac83(\uc9c8\ubb38+\ub2f5+\ud3c9\uac00)\uc744 \uac00\uc9c0\uace0 \ud3c9\uac00\ub97c \uc790\ub3d9\ud654\ud558\ub294 \ubcf4\uc0c1 \ubaa8\ub378\uc744 \uc0c8\ub85c \uad6c\uce21\ud558\uc9c0\ub9cc,<br \/>\n\uc774\uac83\uc744 \uc0dd\ub7b5\ud568<\/p>\n<ul>\n<li>\ub2f5\ubcc0 \ud3c9\uac00 \uc790\ub3d9\ud654<\/li>\n<\/ul>\n<p>GRPO (Group Relative Policy Optimization) \ubc29\ubc95 \uac1c\ubc1c<\/p>\n<ul>\n<li>\n<p>\ub3d9\uc77c\ud55c \uc9c8\ubb38\uc5d0 \ub2f5\ubcc0\uc744 \uc5ec\ub7ec\uac1c \uc0dd\uc131\ud558\ub3c4\ub85d \ud568<\/p>\n<\/li>\n<li>\n<p>\ubbf8\ub9ac \uc815\uc758\ub41c \ub2f5\ubcc0 \ud3c9\uac00 \uaddc\uce59\uc5d0 \ub530\ub77c \uc810\uc218\ud654 (\uc608\uc2dc: \uc218\ud559\uc758 \uacbd\uc6b0 \uc815\ub2f5\uc774 \uc788\uc74c)<\/p>\n<\/li>\n<li>\n<p>\uac00\uc7a5 \uc88b\uc740 \uc810\uc218\uac00 \ub098\uc628 \uc9c8\ubb38-\ub2f5\ubcc0\uc73c\ub85c \ud559\uc2b5<\/p>\n<\/li>\n<li>\n<p>\uc815\ub2f5\uc774 \uc5c6\ub294 \uc9c8\ubb38\uc758 \uacbd\uc6b0 \uac01 \ub2f5\ubcc0\ub4e4\uc744 \uc810\uc218\ud654\ud574\uc11c \ubaa8\ub450 \ud559\uc2b5<\/p>\n<\/li>\n<li>\n<p>\ub2f5\ubcc0\uc774 \ub09c\ud574, \uc5b8\uc5b4 \ud63c\ud569 \ubb38\uc81c(\uc601\uc5b4,\uc911\uad6d\uc5b4 \ud63c\uc7ac)<\/p>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h2>DeepSeek-R1 (\uc774\ubc88 \uc5f0\uad6c \ubaa8\ub378)<\/h2>\n<ul>\n<li>\n<p>\uc21c\uc11c : \uc9c0\uc2dd \uc99d\ub958(Knowledge Distillation) =&gt; SFT =&gt; RL<\/p>\n<\/li>\n<li>\n<p>\uc5ed\uc2dc \uc9c0\uc2dd\uc99d\ub958\ub85c \uc2dc\uc791<\/p>\n<\/li>\n<li>\n<p>\uc18c\ub7c9\uc758 \uc9c0\ub3c4 \ud559\uc2b5 \ub370\uc774\ud0c0 \uc81c\uacf5 (\uc0ac\ub78c\uc774 \uc9c8\ubb38 \ub2f5\ubcc0 \uc0dd\uc131)<\/p>\n<\/li>\n<li>\n<p>\uc774\ud6c4 \uac15\ud654\ud559\uc2b5 \uc2dc\uc791<\/p>\n<ul>\n<li>\ub2f5\ubcc0 \ud3c9\uac00 \uc790\ub3d9\ud654<\/li>\n<\/ul>\n<p>GRPO (Group Relative Policy Optimization)<\/p>\n<ul>\n<li>\ubcf4\uc0c1\ubaa8\ub378<\/li>\n<\/ul>\n<p>\ubaa8\ub378\uc774 \ud2b9\uc815 \ud615\uc2dd(\uc608: <code>&lt;think&gt;<\/code> \uc0ac\uace0 \uacfc\uc815 <code>&lt;\/think&gt;<\/code> <code>&lt;answer&gt;<\/code> \ucd5c\uc885 \ub2f5\ubcc0 <code>&lt;\/answer&gt;<\/code>)\uc5d0 \ub530\ub77c \ucd94\ub860\ud558\uace0 \ub2f5\ubcc0\ud558\ub3c4\ub85d \uc720\ub3c4<\/p>\n<p>\ubaa8\ub378\uc774 \uc790\uae30 \ud3c9\uac00\ub97c \ud558\uae30 \uc2dc\uc791\ud568<\/p>\n<ul>\n<li>\ub2e4\ub2e8\uacc4\ud559\uc2b5<\/li>\n<\/ul>\n<p>\ucd94\ub860\uacfc\uc815\uc5d0\uc11c \ub098\uc628 \uc9c8\ubb38\ub2f5\ubcc0\uc744 \uc774\uc6a9\ud574 \uc7ac\ud559\uc2b5<\/p>\n<ul>\n<li>\uc5b8\uc5b4 \uc77c\uad00\uc131<\/li>\n<\/ul>\n<p><code>&lt;think&gt;<\/code> \uc0ac\uace0 \uacfc\uc815 <code>&lt;\/think&gt;<\/code> \uc601\uc5ed\uc5d0 \ub2e8\uc77c \uc5b8\uc5b4\ub9cc \uc0ac\uc6a9\ud558\ub3c4\ub85d \uc720\ub3c4<\/p>\n<p>\ubaa9\ud45c \uc5b8\uc5b4 \ub2e8\uc5b4 \ube44\uc728\uc744 \ubcf4\uc0c1\uc5d0 \ubc18\uc601<\/p>\n<ul>\n<li>\uc9c1\uc811 \uc99d\ub958<\/li>\n<\/ul>\n<p>\uad50\uc0ac \ubaa8\ub378\uc5d0 \ubaa8\ub378\uc774 \uc9c1\uc811 \uc9c8\ubb38\ud558\uac8c \ud558\uace0 \ub2f5\ubcc0\uc744 \ud559\uc2b5<\/p>\n<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>DeepSeek \ub17c\ubb38 \uc815\ub9ac \ud2c0\ub9b0 \ubd80\ubd84\uc774 \uc788\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4. DeepSeek-R1-Zero DeepSeek-V3-Base \ub97c \uae30\ubc18\uc73c\ub85c RL \ub9cc\uc744 \uc774\uc6a9\ud574 \ud559\uc2b5 \ucc98\uc74c\ubd80\ud130 \uc2dc\uc791\ud558\uba74 \uc2dc\uac04\uacfc \ub3c8\uc774 \ub9ce\uc774 \ub4e4\uae30 \ub54c\ubb38\uc5d0 \uad50\uc0ac \ubaa8\ub378\ub85c\ubd80\ud130 \uc9c0\uc2dd \uc99d\ub958(Knowledge Distillation)\ud568 [\ucd94\uce21] \uc5ec\ub7ec \ubc29\ubc95\uc774 \uc788\uc9c0\ub9cc \uc218\uc2ed\ub9cc\uac1c\uc758 \uc9c8\ubb38\uc744 \ub358\uc9c0\uace0 \ub2f5\ubcc0\uc744 \ubc1b\uc544, \uadf8 \ub370\uc774\ud0c0\ub85c \ud559\uc2b5\ud588\uc744 \ub4ef. \uc9c0\ub3c4\ud559\uc2b5 \uc0dd\ub7b5 \uc9c0\ub3c4\ud559\uc2b5(\uc0ac\ub78c\uc774 \uc0dd\uc131\ud55c \uc9c8\ubb38\uacfc \uc815\ub2f5 \uc81c\uacf5) \uc740 \uc2dc\uac04\uacfc \ube44\uc6a9\uc774 \ub9ce\uc774 \ub4e4\uae30 \ub54c\ubb38\uc5d0 \uc0dd\ub825\u2026 <span class=\"read-more\"><a href=\"https:\/\/www.skyer9.pe.kr\/wordpress\/?p=9952\">Read More &raquo;<\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12],"tags":[],"class_list":["post-9952","post","type-post","status-publish","format-standard","hentry","category-devops"],"_links":{"self":[{"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/9952","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=9952"}],"version-history":[{"count":3,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/9952\/revisions"}],"predecessor-version":[{"id":9955,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/9952\/revisions\/9955"}],"wp:attachment":[{"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=9952"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=9952"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=9952"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}