UJina unemithombo evulekile yemodeli yokufunda yomshini yokumelela umbhalo we-vector, i-jina-embeddings-v2.0, ngaphansi kwelayisensi ye-Apache 2. Imodeli ikuvumela ukuthi uguqule umbhalo ongasho lutho, okuhlanganisa kufika ezinhlamvini ezingu-8192, ube yinombolo encane elandelanayo yezinombolo zangempela ezakha i-vector eqhathaniswa nombhalo womthombo futhi ikhiqize kabusha i-semantics (incazelo). I-Jina Embedding bekuyimodeli yokuqala yokufunda yomshini ovulekile ukuba nokusebenza okufanayo njengemodeli yobunikazi be-vectorization yombhalo ovela kuphrojekthi ye-OpenAI (text-embedding-ada-002), futhi ekwazi ukucubungula umbhalo ngamathokheni afika kwangu-8192.
Ibanga eliphakathi kwama-vector amabili akhiqiziwe lingasetshenziswa ukunquma ubudlelwano be-semantic bemibhalo yomthombo. Empeleni, ama-vectors akhiqiziwe angasetshenziswa ukuhlaziya ukufana kwemibhalo, ukuhlela ukusesha kwezinto ezihlobene nesihloko (imiphumela yokukala ngokusondelana kwe-semantic), imibhalo yeqembu ngencazelo, ukukhiqiza izincomo (ukunikeza uhlu lwezintambo zombhalo ezifanayo), bona okudidayo, thola ukukopela bese uhlukanisa ukuhlolwa. Izibonelo zezindawo ezisetshenziswayo zifaka ukusetshenziswa kwemodeli yokuhlaziywa kwemibhalo yezomthetho, ukuhlaziya kwebhizinisi, ocwaningweni lwezokwelapha lokucubungula ama-athikili esayensi, ekugxekeni kwemibhalo, ukuhlaziya imibiko yezezimali kanye nokwenza ngcono ikhwalithi yokucubungula i-chatbot yezindaba eziyinkimbinkimbi.
Izinguqulo ezimbili zemodeli ye-jina-embeddings ziyatholakala ukuze zilandwe (okuyisisekelo - 0.27 GB futhi kwehlisiwe - 0.07 GB), eziqeqeshwe ngamapheya ayizigidi ezingu-400 zokulandelana kombhalo ngesiNgisi, okuhlanganisa imikhakha eyahlukene yolwazi. Ngesikhathi sokuqeqeshwa, ukulandelana okunobukhulu bamathokheni we-512 kusetshenziswe, okwadluliselwa kubukhulu be-8192 kusetshenziswa indlela ye-ALiBi (Ukunakwa nge-Linear Biases).
Imodeli eyisisekelo ihlanganisa amapharamitha ayizigidi ezingu-137 futhi yakhelwe ukusetshenziswa kumasistimu amile ane-GPU. Imodeli encishisiwe ihlanganisa amapharamitha ayizigidi ezingu-33, inikeza ukunemba okuncane futhi ihloselwe ukusetshenziswa kumadivayisi eselula kanye nezinhlelo ezinenkumbulo encane. Esikhathini esizayo esiseduze bahlela nokushicilela imodeli enkulu ezohlanganisa amapharamitha ayizigidi ezingu-435. Inguqulo yezilimi eziningi yemodeli nayo iyathuthukiswa, okwamanje igxile ekusekelweni kwesiJalimane neSpanishi. I-plugin ilungiselelwe ngokwehlukana ukuze kusetshenziswe imodeli yokushumeka igama ngekhithi yamathuluzi ye-LLM.
Source: opennet.ru