Grupp ta\u2019 ri\u010berkaturi Rumeni \u0127olqu mudell lingwistiku (LLM) dedikat g\u0127all-lingwa Rumena, li jista\u2019 jintu\u017ca biex ji\u0121i \u017cviluppat g\u0127odod u pjattaformi A.I. Il-mudell huwa \u201copen source\u201d, g\u0127alhekk jista\u2019 ji\u0121i a\u010b\u010bessat u u\u017cat minn kul\u0127add li jixtieq jibni g\u0127odod ibba\u017cati fuq l-intelli\u0121enza artifi\u010bjali. Bil-pubblikazzjoni ta\u2019 dan il-LLM, l-inizjaturi tal-pro\u0121ett wkoll i\u0127e\u0121\u0121u l-komunit\u00e0 OpenLLM -Ro, li trid t\u0121ib flimkien lil dawk kollha li jixtiequ jikkontribwixxu b\u2019modijiet varji g\u0127all-i\u017cvilupp tat-teknolo\u0121iji A.I. g\u0127all-lingwa Rumena. It-tnejn pro\u0121etti huma inizzjati u implimentati minn POLITEHNICA Bucharest, l-Universit\u00e0 ta\u2019 Bucharest u l-Istitut tal-Logika u x-Xjenza tad-Dejta, b\u2019appo\u0121\u0121 mill-BRD Groupe Soci\u00e9t\u00e9 G\u00e9n\u00e9rale.<\/p>\n
G\u0127alkemm hija teknolo\u0121ija li g\u0127andna a\u010b\u010bess g\u0127alha fuq skala kbira g\u0127al ftit snin biss, \u0127afna minna di\u0121\u00e0 interagixxew spiss mal-robots konversazzjonali, b\u0127alma huma ChatGPT (produ\u010but minn OpenAI), Copilot (\u017cviluppat minn Microsoft), Gemini (\u017cviluppat minn Google) e\u010b\u010b. Madankollu, g\u0127all-lingwa Rumena r-ri\u017cultati huma xi kultant imprecisi, g\u0127ax il-mudelli li fuqhom huma ta\u0127ri\u0121 ma kinux esposti g\u0127al \u0127afna sorsi ta\u2019 dejta Rumena. Fl-istess \u0127in, dawn it-tipi ta\u2019 g\u0127odod ma jistg\u0127ux jintu\u017caw fil-kumpaniji g\u0127ax l-a\u010b\u010bess dirett jista\u2019 jkun ristrett min\u0127abba ra\u0121unijiet ta\u2019 sigurt\u00e0 u kunfidenzjalit\u00e0. Soluzzjoni wa\u0127da f\u2019dawn is-sitwazzjonijiet hija l-implimentazzjoni ta\u2019 mudell lokali lokalizzat fl-infrastruttura tal-kumpanija. Dawn il-mudelli pubbli\u010bi li jistg\u0127u jintu\u017caw lokalment huma, madankollu, \u0121eneralment it-t\u0127ejjija tag\u0127hom f\u2019lingwa Ingli\u017ca jew u\u017cando numru \u017cg\u0127ir ta\u2019 dokumenti f\u2019lingwi b\u2019cirkolazzjoni baxxa.<\/p>\n
Il-mudell Rumeni l-lum inawgurat huwa l-adattament ta\u2019 LLM pubbliku \u017cviluppat prin\u010bipalment g\u0127all-Ingli\u017c. I\u017cda \u0121ie espost g\u0127al diversi miljun dokument f\u2019Rumeni, sabiex jifhem a\u0127jar il-\u0127sieb tal-kliem. Dan huwa essenzjali g\u0127at-t\u0127addim ta\u2019 dawn il-mudelli f\u2019sitwazzjonijiet fejn l-espressjoni tal-\u0127ti\u0121ijiet jew mistoqsijiet tal-utent kif ukoll il-wie\u0121eb g\u0127andhom ikunu f\u2019Rumeni. Mill-lejla tas-sena 2023, grupp ta\u2019 ri\u010berkaturi minn POLITEHNICA Bucharest, l-Universit\u00e0 ta\u2019 Bucharest u l-Istitut tal-Logika u x-Xjenza tad-Dejta \u0127admu fuq l-i\u017cvilupp u t-ta\u0127ri\u0121 ta\u2019 dan il-LLM. Il-partners akkademi\u010bi kkontribwixxew bi ri\u010berkaturi li \u0127admu pro-bono u, barra minn hekk, POLITEHNICA Bucharest ipprovdet ukoll il-qawwa tal-kompjuter me\u0127tie\u0121a biex it-ta\u0127ri\u0121 tal-mudell. Il-partner prin\u010bipali tal-pro\u0121ett huwa BRD Groupe Soci\u00e9t\u00e9 G\u00e9n\u00e9rale, li jappo\u0121\u0121ja l-innovazzjoni u t-teknolo\u0121iji tal-futur f\u2019Rumeni f\u2019kull forma tag\u0127hom.<\/p>\n
\u201cBiex l-ambjent ekonomiku u\/jew istituzzjonali f\u2019Rumeni jkun jista\u2019 ju\u017ca din it-teknolo\u0121ija promettenti \u0121dida, g\u0127andna b\u017conn mudelli spe\u010bjalizzati li jiltaqg\u0127u ma\u2019 \u0127afna konversazzjonijiet u dokumenti f\u2019Rumeni. Ir-ra\u0121uni hija sempli\u010bi: biex tkun kapa\u010bi tipprovdi l-informazzjoni li g\u0127andna b\u017conn. F\u2019BRD a\u0127na dejjem qed na\u0127dmu fuq soluzzjonijiet li jitjieb il-pro\u010bessi tax-xog\u0127ol tag\u0127na, billi nu\u017caw it-teknolo\u0121iji l-aktar \u0121odda li jistg\u0127u jag\u0127tu valur mi\u017cjud lill-klijenti tag\u0127na fl-ewwel post. Imma nifhmu wkoll li l-b\u017connijiet tag\u0127na huma maqsuma ma\u2019 \u0127afna atturi istituzzjonali o\u0127ra, u a\u0127na impenjati li nappo\u0121\u0121jaw l-innovazzjoni fl-intelli\u0121enza artifi\u010bjali kmieni. Billi ninvolvu ru\u0127na fil-pajsa\u0121\u0121 animat tieg\u0127u, nistg\u0127u ng\u0127inu biex it-teknolo\u0121iji l-aktar \u0121odda jkollhom impatt po\u017cittiv fil-so\u010bjet\u00e0 Rumena fil-prattikament fl-istess pass b\u0127at-tkabbir fil-qasam fuq livell internazzjonali,\u201d qal Horia Velicu, Kap tal-Laboratorju tal-I\u017cvilupp f\u2019BRD Groupe Soci\u00e9t\u00e9 G\u00e9n\u00e9rale.<\/p>\n
\u201cXi w\u0127ud mill-e\u017cempji ta\u2019 u\u017cu tal-mudell Rumeni huma: tfittxija ta\u2019 informazzjoni fil-ba\u017ci tad-dejta ta\u2019 organizzazzjoni, b\u2019gwidi u pro\u010beduri tax-xog\u0127ol, jew robots konversazzjonali g\u0127al klijenti ta\u2019 kumpaniji jew istituzzjonijiet biex jiggwidawhom fil-passi me\u0127tie\u0121a biex ju\u017caw prodott jew servizz. Fil-lakuni ta\u2019 dawn il-ka\u017cijiet, l-impjegati u\/jew il-klijenti jisalvaw \u0127in fl-a\u010b\u010bess g\u0127all-informazzjoni, benefikanti f\u2019\u0127afna situazzjonijiet u titjieb il-kwalit\u00e0 tag\u0127ha,\u201d qal Alin Stefanescu, Direttur tad-Dipartiment tal-Informazzjoni tal-Kompjuter fl-Universit\u00e0 ta\u2019 Bucharest u vi\u010bi president tal-Istitut tal-Logika u x-Xjenza tad-Dejta.<\/p>\n
L-isforz biex ji\u0121u spe\u010bjalizzati mudelli lingwisti\u010bi huwa spiss kkoordinat mill-komunit\u00e0 akkademika asso\u010bjata ma\u2019 dik il-lingwa, b\u2019e\u017cempji re\u010benti minn pajji\u017ci b\u0127al Franza, \u0120ermanja, Spanja, il-Finlandja, il-Bulgarija. Madankollu, ir-ri\u017corsi me\u0127tie\u0121a huma kbar kemm fir-rekwi\u017citi teknolo\u0121i\u010bi me\u0127tie\u0121a (e\u017c. hardwer dedikat b\u0127al karti grafika b\u2019qawwa g\u0127olja) kif ukoll fir-ri\u010berkaturi u programmaturi esperti. G\u0127alhekk hemm b\u017conn ta\u2019 appo\u0121\u0121 wiesa\u2019, ta\u2019 medda u twila mill-atturi so\u010bjali ewlenin: ekonomi\u010bi, akkademi\u010bi u, fl-a\u0127\u0127ar i\u017cda mhux l-inqas, l-ambjent governattiv, permezz ta\u2019 programmi dedikati g\u0127all-i\u017cvilupp tat-teknolo\u0121iji ta\u2019 intelli\u0121enza artifi\u010bjali.<\/p>\n
Hekk kif i\u010b-\u010breaturi ta\u2019 dan il-mudell qed inawgura l-komunit\u00e0 OpenLLM.ro fl-istess \u0127in. Dan g\u0127andu l-g\u0127an li jinkora\u0121\u0121ixxi l-interazzjoni bejn atturi varji jew fa\u010bilitaturi li jixtiequ jikkontribwixxu g\u0127all-i\u017cvilupp ta\u2019 din it-teknolo\u0121ija g\u0127all-lingwa Rumena u li jinawguraw mudelli spe\u010bjalizzati g\u0127al oqsma \u010berti. L-ibliet ta\u2019 dan id-djalogu f\u2019ambjent open source se jippromwovi l-\u0127olqien ta\u2019 mudelli aktar effettivi, implimentati fil-kumpaniji jew istituzzjonijiet Rumeni, li jirri\u017cultaw fi tnaqqis fil-produttivit\u00e0 \u0121enerali tas-so\u010bjet\u00e0 kollha.<\/p>\n
\u201cNisperaw li l-inawgurar ta\u2019 dan il-mudell ikun biss l-bidu ta\u2019 sforz twil li jirri\u017culta fi LLMs a\u0127jar g\u0127all-lingwa Rumena. Di\u0121\u00e0 skoprejna metodu li rridu nippplikaw g\u0127al mudelli o\u0127ra li \u0121ew inawgurat re\u010bentement (Llama-3 u Mistral) u li \u0121eneralment juru prestazzjoni a\u0127jar minn dik li nibdew minnha (Llama-2). Madankollu, biex ikollna mudelli tajbin g\u0127all-lingwa Rumena g\u0127andna b\u017conn 2 tipi ta\u2019 ri\u017corsi: kollezzjonijiet ta\u2019 dejta kbar u kkurati, ta\u2019 kwalit\u00e0 tajba, kif ukoll ri\u017corsi ta\u2019 hardwer (b\u2019mod partikolari, GPUs g\u0127all-it-ta\u0127ri\u0121 tal-mudell). Nisperaw li kemm entitajiet privati u pubbli\u010bi jifhmu l-importanza li ji\u0121u \u017cviluppati mudelli lingwisti\u010bi kbar u multi-modal (test-imma\u0121ini) g\u0127all-lingwa Rumena. Nistennew lil kul\u0127add jissie\u0127eb mag\u0127na fl-inizjattiva OpenLLM-Ro u l-pro\u0121etti ta\u2019 ri\u010berka li jappo\u0121\u0121jawha,\u201d qal Traian Rebedea, lektur fil-POLITEHNICA Bucharest u ri\u010berkatur anzjan f\u2019NVIDIA, wie\u0127ed mill-koordinaturi tekni\u010bi tal-inizjativa OpenLLM-Ro.<\/p>\n
Il-rapport tekniku jista\u2019 jinstab hawn: https:\/\/arxiv.org\/abs\/2405.07703<\/a><\/p>\n Il-mudell LLM jista\u2019 jitni\u017c\u017cel mill-pjattaforma Hugging Face: https:\/\/huggingface.co\/OpenLLM-Ro<\/a><\/p>\n Il-kodi\u010bi asso\u010bjat mal-mudell jista\u2019 jitni\u017c\u017cel minn GitHub: https:\/\/github.com\/OpenLLM-Ro<\/a><\/p>\n