\u0130nce ayar (fine-tuning), \u00f6nceden e\u011fitilmi\u015f bir temel modeli<\/a> bir g\u00f6reve, alana veya stile \u00f6zel olarak k\u00fcrat\u00f6rlenmi\u015f bir veri k\u00fcmesinde daha fazla e\u011fitme s\u00fcrecidir; sonu\u00e7 olarak temel modele k\u0131yasla o hedefte daha iyi performans g\u00f6steren bir model elde edilir. Modern ince ayar, minimal hesaplama ile modelleri uyarlayan LoRA gibi parametre verimli y\u00f6ntemlerin egemenli\u011findedir.<\/p>\n Birka\u00e7 ince ayar yakla\u015f\u0131m\u0131 vard\u0131r: tam ince ayar (t\u00fcm model parametrelerini g\u00fcnceller\u2014b\u00fcy\u00fck modeller i\u00e7in pahal\u0131), LoRA (D\u00fc\u015f\u00fck S\u0131ral\u0131 Adaptasyon, k\u00fc\u00e7\u00fck e\u011fitilebilir matrisler ekler\u2014modern ince ayar i\u00e7in tipik y\u00f6ntem), QLoRA (4-bit niceleme ile LoRA\u2014t\u00fcketici donan\u0131m\u0131nda \u00e7al\u0131\u015f\u0131r), prompt ayarlama (a\u011f\u0131rl\u0131klar\u0131 de\u011fi\u015ftirmek yerine yumu\u015fak prompt’lar\u0131 \u00f6\u011frenir) ve talimat ayarlama (talimat-takip verisi \u00fczerinde ince ayar). Do\u011frudan tercih optimizasyonu (DPO<\/a>) ve benzer y\u00f6ntemler model davran\u0131\u015f\u0131n\u0131 insan tercihlerine do\u011fru daha da iyile\u015ftirir.<\/p>\n \u0130nce ayar \u015fu durumlarda anlaml\u0131d\u0131r: temel model g\u00fc\u00e7l\u00fc prompt’lama ile bile belirli g\u00f6reve zorluk \u00e7ekiyorsa, en az 100-1.000+ y\u00fcksek kaliteli e\u011fitim \u00f6rne\u011finiz varsa, tutarl\u0131 \u00e7\u0131kt\u0131 bi\u00e7imlendirme veya stil ihtiyac\u0131n\u0131z varsa, uzun prompt’lar yerine davran\u0131\u015f\u0131 model a\u011f\u0131rl\u0131klar\u0131na kodlayarak token maliyetlerini azaltmak istiyorsan\u0131z veya daha b\u00fcy\u00fck bir modelin g\u00f6rev performans\u0131yla e\u015fle\u015fen daha k\u00fc\u00e7\u00fck, daha h\u0131zl\u0131 bir modeli da\u011f\u0131tman\u0131z gerekiyorsa. \u00c7o\u011fu jenerik g\u00f6rev i\u00e7in, \u00f6nce RAG<\/a> ve prompt m\u00fchendisli\u011fi<\/a> denenmelidir.<\/p>\n \u0130nce ayar veri kalitesi miktardan \u00e7ok daha \u00f6nemlidir\u2014k\u00fcrat\u00f6rlenmi\u015f, \u00e7e\u015fitli, hatas\u0131z \u00f6rnekler daha b\u00fcy\u00fck g\u00fcr\u00fclt\u00fcl\u00fc veri k\u00fcmelerinden daha iyi modeller \u00fcretir. Maliyetler b\u00fcy\u00fck \u00f6l\u00e7\u00fcde de\u011fi\u015fir: API ince ayar (OpenAI, Anthropic) milyon token ba\u015f\u0131na sentlerden dolarlara kadar de\u011fi\u015fir; a\u00e7\u0131k modellerin kendi kendine bar\u0131nd\u0131r\u0131lan LoRA ince ayar\u0131 tek bir GPU’da 100 dolar\u0131n alt\u0131nda yap\u0131labilir. Hukuki d\u00fc\u015f\u00fcnceler e\u011fitim verisi lisanslamas\u0131, \u00e7\u0131kt\u0131 sahipli\u011fi ve platform hizmet ko\u015fullar\u0131na uyumu i\u00e7erir.<\/p>\n","protected":false},"excerpt":{"rendered":" TLDR: \u0130nce ayar (fine-tuning), \u00f6nceden e\u011fitilmi\u015f bir temel modeli bir g\u00f6reve, alana veya stile \u00f6zel olarak k\u00fcrat\u00f6rlenmi\u015f bir veri k\u00fcmesinde daha fazla e\u011fitme s\u00fcrecidir; sonu\u00e7 olarak temel modele k\u0131yasla o hedefte daha iyi performans g\u00f6steren bir model elde edilir. Modern ince ayar, minimal hesaplama ile modelleri uyarlayan LoRA gibi parametre verimli y\u00f6ntemlerin egemenli\u011findedir. \u0130nce Ayar […]<\/p>\n","protected":false},"author":4,"featured_media":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","format":"standard","meta":{"footnotes":""},"categories":[],"class_list":["post-11911","term","type-term","status-publish","format-standard","hentry"],"_links":{"self":[{"href":"https:\/\/virconlegal.com\/tr\/wp-json\/wp\/v2\/term\/11911","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/virconlegal.com\/tr\/wp-json\/wp\/v2\/term"}],"about":[{"href":"https:\/\/virconlegal.com\/tr\/wp-json\/wp\/v2\/types\/term"}],"author":[{"embeddable":true,"href":"https:\/\/virconlegal.com\/tr\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/virconlegal.com\/tr\/wp-json\/wp\/v2\/comments?post=11911"}],"version-history":[{"count":1,"href":"https:\/\/virconlegal.com\/tr\/wp-json\/wp\/v2\/term\/11911\/revisions"}],"predecessor-version":[{"id":12319,"href":"https:\/\/virconlegal.com\/tr\/wp-json\/wp\/v2\/term\/11911\/revisions\/12319"}],"wp:attachment":[{"href":"https:\/\/virconlegal.com\/tr\/wp-json\/wp\/v2\/media?parent=11911"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/virconlegal.com\/tr\/wp-json\/wp\/v2\/categories?post=11911"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}\u0130nce Ayar Y\u00f6ntemleri<\/h3>\n
Ne Zaman \u0130nce Ayar Yap\u0131l\u0131r<\/h3>\n
Veri ve Maliyet D\u00fc\u015f\u00fcnceleri<\/h3>\n