Ishicilelwe i-OpenChatKit, ikhithi yamathuluzi yokwakha ama-chatbots

Kwethulwa ikhithi yamathuluzi evulekile ye-OpenChatKit, okuhloswe ngayo ukwenza lula ukudalwa kwama-chatbots ezinhlelo zokusebenza ezikhethekile nezijwayelekile. Uhlelo luguqulelwe ukwenza imisebenzi efana nokuphendula imibuzo, ukuqhuba izingxoxo zezigaba eziningi, ukufingqa, ukukhipha ulwazi, ukuhlukanisa umbhalo. Ikhodi ibhalwe ngePython futhi isatshalaliswa ngaphansi kwelayisensi ye-Apache 2.0. Le phrojekthi ihlanganisa imodeli eseyenziwe ngomumo, ikhodi yokuqeqesha imodeli yakho, izinsiza zokuhlola imiphumela yemodeli, amathuluzi okwengeza imodeli ngomongo ovela kunkomba yangaphandle kanye nokulungisa imodeli eyisisekelo ukuze uxazulule izinkinga zakho.

I-bot isuselwe kumodeli yokufunda yomshini eyisisekelo (i-GPT-NeoXT-Chat-Base-20B), eyakhelwe kusetshenziswa imodeli yolimi emboza amapharamitha angaba yizigidi eziyizinkulungwane ezingama-20 futhi yathuthukiselwa ukuxhumana kwengxoxo. Imodeli yaqeqeshwa kusetshenziswa idatha etholwe ekuqoqweni kwephrojekthi ye-LAION, Ndawonye kanye ne-Ontocord.ai.

Ukuze kwandiswe isisekelo solwazi esikhona, kuhlongozwa isistimu ekwazi ukukhipha ulwazi olwengeziwe kumakhosombe angaphandle, ama-API neminye imithombo. Isibonelo, kungenzeka ukuthi ubuyekeze ulwazi usebenzisa idatha evela ku-Wikipedia kanye nezifunzo zezindaba. Ukwengeza, imodeli yokulinganisa iyatholakala, eqeqeshwe ngamapharamitha ayizigidi eziyizinkulungwane ezingu-6, ngokusekelwe kumodeli ye-GPT-JT, futhi yakhelwe ukuhlunga imibuzo engafanele noma ikhawulele izingxoxo ezihlokweni ezithile.

Ngokwehlukana, singaqaphela iphrojekthi ye-ChatLLaMA, enikezela ngomtapo wolwazi wokudala abasizi abahlakaniphile abafana ne-ChatGPT. Le phrojekthi ithuthuka ngeso lokuthi kungenzeka isebenze ngemishini yayo futhi idale izixazululo eziqondene nomuntu eziklanyelwe ukuhlanganisa izindawo eziwumngcingo zolwazi (isibonelo, imithi, umthetho, imidlalo, ucwaningo lwesayensi, njll.). Ikhodi ye-ChatLLaMA ilayisensi ngaphansi kwe-GPLv3.

Le phrojekthi isekela ukusetshenziswa kwamamodeli asekelwe esakhiweni se-LLaMA (Large Language Model Meta AI) esihlongozwe yi-Meta. Imodeli egcwele ye-LLaMA ihlanganisa amapharamitha ayizigidi eziyizinkulungwane ezingu-65, kodwa ku-ChatLLaMA kunconywa ukuthi kusetshenziswe okuhlukile okunamapharamitha ayizigidi eziyizinkulungwane ezingu-7 nezingu-13 noma i-GPTJ (amabhiliyoni angu-6), i-GPTNeoX (ibhiliyoni elingu-1.3), 20BOPT (ibhiliyoni elingu-13), i-BLOOM (ibhiliyoni elingu-7.1) kanye namamodeli e-Galactica (amabhiliyoni angu-6.7) ). Ekuqaleni, amamodeli e-LLaMA anikezwa abacwaningi kuphela ngesicelo esikhethekile, kodwa njengoba kwasetshenziswa izifufula ukuletha idatha, abashisekayo balungiselele umbhalo ovumela noma ubani ukulanda imodeli.

Source: opennet.ru

Engeza amazwana