JetBrains ya buɗe wani samfuri Mellum2, an tsara shi don amfani a cikin kayan aikin AI don haɓaka software. An buga samfurin a ƙarƙashin lasisi Apache 2.0Ana samun nauyin a kan Hugging Face. JetBrains ya jaddada cewa an horar da Mellum2 tun daga farko kuma ba don ayyuka da yawa ba, amma don aiki tare da rubutu da lambar: hanyar neman hanya, bututun RAG, taƙaitawa, wakilai masu taimako, da kuma tura su cikin kayayyakin more rayuwa na kamfani.
An gina Mellum2 akan gine-gine Cakuda-da-KwararruTare da jimlar girman 12 biliyan sigogi Kimanin 1000 ne kawai aka kunna a kowace alama 2.5 biliyan sigogi, wanda zai rage farashin lissafi da jinkirin aiki yayin bincike. A cewar JetBrains, aikin ma'aunin samfurin yana kama da samfuran buɗaɗɗen tushe masu girman iri ɗaya, amma yana ba da fiye da ninki biyu na saurin bincike.
JetBrains ya bayyana Mellum2 a matsayin juyin halittar samfurin Mellum na asali, wanda aka ƙirƙira shi da farko don kammala lambar. Sabuwar sigar ta faɗaɗa zuwa babban aji na ayyuka waɗanda ke buƙatar aiki tare da lambar shirin da harshe na halitta. Kamfanin yana sanya Mellum2 a matsayin samfurin "mai da hankali" - ba maye gurbin manyan LLMs na gabaɗaya ba, amma wani abu mai sauri, na musamman don ayyukan matsakaici akai-akai a cikin tsarin AI mai rikitarwa.
Daga cikin shawarwarin da aka bayar na amfani da su akwai ake kira Rarrabawa da kuma daidaita buƙatu tsakanin samfura da kayan aiki, matsewa da sarrafa mahallin a cikin tsarin RAG, shirya bayanai ga wakilai, tsara jadawalin, tabbatar da sakamakon matsakaici, da aiwatar da su a cikin mahallin inda ba zai yiwu a aika lambar tushe ko bayanan ciki zuwa APIs na waje ba.
Akan Rufe Fuska aka buga tarin Mellum 2, wanda ya haɗa da nau'ikan samfura da yawa: Tunani, Koyarwa, Tunani-SFT, Koyarwa-SFT, Tushe, da Tushe-Pretrain. Ana rarraba samfuran a cikin tsarin Safetensors a ƙarƙashin lasisin Apache 2.0.
An bayar da misalan amfani ta hanyar Transformers, vLLM, SGLang, da Docker Model Runner don ƙaddamarwa.
Abin da ya fi ban sha'awa a fannin fasaha ba shine fitowar wani tsarin bude tushen ba, amma abin da JetBrains ya zaɓa. Kamfanin ba ya mai da hankali kan yin gogayya da manyan samfuran gama gari, amma akan kayan aiki masu rahusa da sauri waɗanda za a iya haɗa su kai tsaye cikin IDEs, mataimakan ciki, tsarin RAG na kamfanoni, da bututun mai. Ga masu haɓakawa da kamfanoni, wannan yana nufin ikon gudanar da wasu dabaru na AI a gida ko akan sabar su, yayin da suke riƙe da iko akan lambar, bayanai, da farashin ƙididdiga.
source: linux.org.ru
