{"id":18159,"date":"2025-02-04T13:12:45","date_gmt":"2025-02-04T13:12:45","guid":{"rendered":"https:\/\/staging.tilde.ai\/?p=18159"},"modified":"2026-07-14T14:07:49","modified_gmt":"2026-07-14T12:07:49","slug":"the-race-of-ai-models","status":"publish","type":"post","link":"https:\/\/staging.tilde.ai\/lt\/blog\/the-race-of-ai-models\/","title":{"rendered":"DI modeliai ma\u017eesn\u0117ms kalboms: kaip jie laikosi?"},"content":{"rendered":"<div data-elementor-type=\"wp-post\" data-elementor-id=\"18159\" class=\"elementor elementor-18159\" data-elementor-post-type=\"post\">\n\t\t\t\t<div class=\"elementor-element elementor-element-7ebcd9b e-flex e-con-boxed e-con e-parent\" data-id=\"7ebcd9b\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-11105e3 elementor-widget elementor-widget-shortcode\" data-id=\"11105e3\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"shortcode.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<div class=\"elementor-shortcode\">\t\t<div data-elementor-type=\"page\" data-elementor-id=\"21097\" class=\"elementor elementor-21097\" data-elementor-post-type=\"elementor_library\">\n\t\t\t\t<div class=\"elementor-element elementor-element-58a3430b e-flex e-con-boxed e-con e-parent\" data-id=\"58a3430b\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-6d2f31d9 elementor-widget elementor-widget-shortcode\" data-id=\"6d2f31d9\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"shortcode.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<div class=\"elementor-shortcode\"><nav aria-label=\"duonos trupiniai\" class=\"rank-math-breadcrumb\"><p><span class=\"last\">Pagrindinis<\/span><\/p><\/nav><\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-3856ae94 elementor-widget elementor-widget-theme-post-title elementor-page-title elementor-widget-heading\" data-id=\"3856ae94\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"theme-post-title.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t<h1 class=\"elementor-heading-title elementor-size-default\">DI modeliai ma\u017eesn\u0117ms kalboms: kaip jie laikosi?<\/h1>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-23eb1b5c elementor-align-center elementor-mobile-align-left elementor-widget elementor-widget-post-info\" data-id=\"23eb1b5c\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"post-info.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<ul class=\"elementor-inline-items elementor-icon-list-items elementor-post-info\">\n\t\t\t\t\t\t\t\t<li class=\"elementor-icon-list-item elementor-repeater-item-1935125 elementor-inline-item\" itemprop=\"datePublished\">\n\t\t\t\t\t\t<a href=\"https:\/\/staging.tilde.ai\/lt\/2025\/02\/04\/\">\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t<span class=\"elementor-icon-list-text elementor-post-info__item elementor-post-info__item--type-date\">\n\t\t\t\t\t\t\t\t\t\t<time>4 vasario, 2025 m.<\/time>\t\t\t\t\t<\/span>\n\t\t\t\t\t\t\t\t\t<\/a>\n\t\t\t\t<\/li>\n\t\t\t\t<\/ul>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-343074d3 e-flex e-con-boxed e-con e-parent\" data-id=\"343074d3\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-1aede4a9 elementor-widget elementor-widget-theme-post-featured-image elementor-widget-image\" data-id=\"1aede4a9\" data-element_type=\"widget\" data-e-type=\"widget\" data-settings=\"{&quot;motion_fx_motion_fx_scrolling&quot;:&quot;yes&quot;,&quot;motion_fx_translateY_effect&quot;:&quot;yes&quot;,&quot;motion_fx_translateY_speed&quot;:{&quot;unit&quot;:&quot;px&quot;,&quot;size&quot;:2,&quot;sizes&quot;:[]},&quot;motion_fx_translateY_affectedRange&quot;:{&quot;unit&quot;:&quot;%&quot;,&quot;size&quot;:&quot;&quot;,&quot;sizes&quot;:{&quot;start&quot;:0,&quot;end&quot;:80}},&quot;motion_fx_devices&quot;:[&quot;desktop&quot;,&quot;tablet&quot;,&quot;mobile&quot;]}\" data-widget_type=\"theme-post-featured-image.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<img fetchpriority=\"high\" decoding=\"async\" width=\"1596\" height=\"950\" src=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/Jurgita_blogam.png\" class=\"attachment-full size-full wp-image-18164\" alt=\"Dr. Jurgita Kapo\u010di\u016bt\u0117-Dzikien\u0117 apie DI modelius ma\u017eesn\u0117ms kalboms\" srcset=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/Jurgita_blogam.png 1596w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/Jurgita_blogam-300x179.png 300w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/Jurgita_blogam-1024x610.png 1024w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/Jurgita_blogam-768x457.png 768w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/Jurgita_blogam-1536x914.png 1536w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/Jurgita_blogam-18x12.png 18w\" sizes=\"(max-width: 1596px) 100vw, 1596px\" \/>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-736a2f25 e-flex e-con-boxed e-con e-parent\" data-id=\"736a2f25\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-2fc903b3 elementor-widget elementor-widget-text-editor\" data-id=\"2fc903b3\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t<p><strong>Nuo statistini\u0173 kalb\u0173 modeli\u0173 atsiradimo 1980-aisiais, \u0161i sritis patyr\u0117 nepaprast\u0105 transformacij\u0105. \u0160iandien tikrasis klausimas yra, kaip dirbtinio intelekto modeliai ma\u017eesn\u0117ms kalboms sp\u0117ja su \u0161ia transformacija. Per de\u0161imtme\u010dius \u0161i\u0173 dirbtinio intelekto modeli\u0173 k\u016brimo metodikos evoliucionavo, apdorojam\u0173 duomen\u0173 kiekis i\u0161augo, o patys modeliai tapo sud\u0117tingesni. Parametr\u0173 \u2013 matematini\u0173 koeficient\u0173 formul\u0117se \u2013 skai\u010dius nuolat pl\u0117t\u0117si. Iki 2020-\u0173j\u0173, su <em>\u201eChatGPT\u201c,<\/em>, terminas \u201edidysis kalbos modelis\u201c buvo pla\u010diai pripa\u017eintas. \u0160iandien sunku rasti \u017emog\u0173, kuris apie tai neb\u016bt\u0173 gird\u0117j\u0119s. <\/strong><\/p>\n<p>\u00a0<\/p>\n<p><strong>Bet k\u0105 rei\u0161kia \u201edidelis\u201c kalbant apie kalbos modelius? Vienu metu mil\u017eini\u0161ku buvo laikomas modelis su 3 mlrd. parametr\u0173, ta\u010diau \u0161iandien tokie skai\u010diai jau nebestebina. <em>\u201eOpenAI\u201c<\/em> neatskleid\u017eia tiksli\u0173 duomen\u0173 apie <em>\u201eGPT-4\u201c<\/em>model\u012f, ta\u010diau manoma, kad jo parametr\u0173 skai\u010dius gali siekti nuo 1 iki 10 trilijon\u0173.<\/strong><\/p>\n<p>\u00a0<\/p>\n<h2>Kam skirti DI modeliai ma\u017eesn\u0117ms kalboms?<\/h2>\n<p>Ma\u017eiau paplitusioms kalboms skirti DI modeliai yra kalbos technologijos, skirtos suprasti, generuoti ir versti kalbas, kurioms yra ma\u017eiau skaitmenini\u0173 i\u0161tekli\u0173. Skirtingai nuo modeli\u0173, apmokyt\u0173 daugiausia naudojant didelius angl\u0173 kalbos duomen\u0173 rinkinius, \u0161iems modeliams reikia kruop\u0161\u010diai atrinkt\u0173 mokymo duomen\u0173, vietin\u0117s kalbos ekspert\u0173 ir patikim\u0173 vertinimo metod\u0173, kad jie gerai veikt\u0173 realiose naudojimo situacijose.<\/p>\n<p>\u00a0<\/p>\n<h2>K\u0105 dirbtinio intelekto kalbiniuose modeliuose rei\u0161kia \u201edidelis\u201c?<\/h2>\n<p><!-- \/wp:paragraph --><\/p>\n<p><!-- wp:paragraph --><\/p>\n<p>Mokslo pasaulyje vyksta tikras prover\u017eis \u2013 publikacij\u0173 apie did\u017eiuosius kalbos modelius gausu kaip niekada anks\u010diau, o kiekviena \u0161alis ir tyrim\u0173 grup\u0117 stengiasi sukurti savo model\u012f \u2013 kuo galingesn\u012f, tikslesn\u012f ir universalesn\u012f. Tokie modeliai turi geb\u0117ti atsakyti \u012f klausimus, generuoti santraukas, versti tekstus ir atlikti daugyb\u0119 kalbini\u0173 u\u017eduo\u010di\u0173. Ta\u010diau vien tik didelis parametr\u0173 skai\u010dius dar nieko negarantuoja. Kiekvienas parametras turi b\u016bti suderintas (t. y. jam turi b\u016bti priskirta tinkama matematin\u0117 reik\u0161m\u0117), o tai \u012fmanoma tik turint pakankamai mokymo duomen\u0173.<\/p>\n<p>\u00a0<\/p>\n<h2>Kod\u0117l DI modeliai ma\u017eesn\u0117ms kalboms susiduria su duomen\u0173 problema<\/h2>\n<p><!-- \/wp:paragraph --><\/p>\n<p><!-- wp:paragraph --><\/p>\n<p>Kaip mokomi \u0161iuolaikiniai didieji kalbos modeliai? Pirmiausia sukonstruojama j\u0173 architekt\u016bra \u2013 matematin\u0117s formul\u0117s ir j\u0173 koeficientai, kurie i\u0161 prad\u017ei\u0173 u\u017epildomi atsitiktin\u0117mis reik\u0161m\u0117mis. V\u0117liau, analizuojant mokymo duomenis, \u0161ie koeficientai tikslinami. Jei architekt\u016bra per didel\u0117, o duomen\u0173 per ma\u017eai \u2013 modelis veiks netiksliai ir \u201ehaliucinuos\u201c. Jei modelis per ma\u017eas, bet duomen\u0173 labai daug \u2013 jis netur\u0117s pakankamai paj\u0117gum\u0173 \u012fsisavinti vis\u0105 informacij\u0105.  \nAugant vartotoj\u0173 l\u016bkes\u010diams, vis aktualesne problema tampa duomen\u0173 tr\u016bkumas. Tekst\u0173 lietuvi\u0173 kalba gerokai ma\u017eiau nei, pavyzd\u017eiui, angl\u0173, vokie\u010di\u0173 ar lenk\u0173 kalbomis. Tod\u0117l tyr\u0117jai vis da\u017eniau imasi spr\u0119sti \u0161i\u0105 problem\u0105. Bendrov\u0117 <em>\u201eTilde\u201c<\/em> kartu su <em>VDU prof. Jurgita Kapo\u010di\u016bt\u0117-Dzikien\u0117<\/em> ir <em>Vilniaus universitetais<\/em> \u0161iuo metu vykdo projektus, siekdama sukaupti daugiau duomen\u0173, kurie ateityje taps pagrindu kurti dar tikslesnius modelius. Did\u017eiosios kompanijos, tokios kaip <em>\u201eOpenAI\u201c<\/em>, <em>\u201eMeta\u201c<\/em>, ir <em>\u201eGoogle DeepMind\u201c,<\/em> taip pat susiduria su ma\u017eesni\u0173 kalb\u0173 duomen\u0173 tr\u016bkumo problema, ta\u010diau j\u0173 modeliai yra daugiakalbiai. Daugiakalbiai modeliai pasi\u017eymi lankstumu \u2013 jie lengviau palaiko ma\u017eesnes kalbas, nes perima \u017einias i\u0161 did\u017ei\u0173j\u0173 kalb\u0173 ir i\u0161naudoja tarpkalbinius ry\u0161ius. B\u016btent d\u0117l \u0161ios prie\u017easties <em>\u201eTilde\u201c<\/em> kuria savo daugiakalb\u012f model\u012f <strong><em><a href=\"https:\/\/staging.tilde.ai\/lt\/tildeopen-llm\/\">\u201eTildeLM\u201c<\/a><\/em><\/strong>, su ai\u0161kiu d\u0117mesiu ma\u017eesn\u0117ms kalboms, tokioms kaip lietuvi\u0173, latvi\u0173 ir est\u0173. Tai vienas i\u0161 nedaugelio dirbtinio intelekto modeli\u0173 ma\u017eesn\u0117ms kalboms, sukurt\u0173 su \u0161iomis rinkomis kaip prioritetu, o ne kaip papildoma mintimi.<\/p>\n<p>\u00a0<\/p>\n<p><!-- \/wp:paragraph --><\/p>\n<p><!-- wp:qi-blocks\/spacer {\"uniqueClass\":\"qodef-block-6a64be29\",\"space\":20} --><\/p>\n<div class=\"wp-block-qi-blocks-spacer qodef-block-container qodef-block-6a64be29\">\n<div class=\"qi-block-spacer qodef-block qodef-m\">\u00a0<\/div>\n<\/div>\n<p><!-- \/wp:qi-blocks\/spacer --><\/p>\n<p><!-- wp:image {\"id\":18162,\"sizeSlug\":\"full\",\"linkDestination\":\"none\"} --><\/p>\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" class=\"alignnone wp-image-18162 size-full\" src=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/17690.jpg\" alt=\"DI modeliai ma\u017eesn\u0117ms kalboms \u2013 pasaulin\u0117s kalbos technologij\u0173 lenktyn\u0117s\" width=\"1000\" height=\"667\" srcset=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/17690.jpg 1000w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/17690-300x200.jpg 300w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/17690-768x512.jpg 768w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2025\/02\/17690-18x12.jpg 18w\" sizes=\"(max-width: 1000px) 100vw, 1000px\" \/><\/figure>\n<p><!-- \/wp:image --><\/p>\n<p><!-- wp:qi-blocks\/spacer {\"uniqueClass\":\"qodef-block-f4885452\",\"space\":20} --><\/p>\n<div class=\"wp-block-qi-blocks-spacer qodef-block-container qodef-block-f4885452\">\n<div class=\"qi-block-spacer qodef-block qodef-m\">\u00a0<\/div>\n<\/div>\n<p><!-- \/wp:qi-blocks\/spacer --><\/p>\n<p><!-- wp:paragraph --><\/p>\n<h2>LLM lenktyn\u0117s: GPT, Llama, Gemma ir kiti<\/h2>\n<p>Kalbos modeli\u0173 pasaulyje vyksta tikros lenktyn\u0117s. <em>GPT<\/em>, <em>\u201eMistral\u201c<\/em>, <em>\u201eLlama3.1\u201c<\/em>, <em>\u201eGemma\u201c<\/em>, <em>\u201eBloom\u201c<\/em>, <em>\u201eClaude\u201c<\/em>, <em>\u201eSolar\u201c<\/em> ir daugelis kit\u0173 did\u017ei\u0173j\u0173 kalbos modeli\u0173 nuolat konkuruoja tarpusavyje. Ta\u010diau i\u0161lieka viena problema \u2013 apie \u0161i\u0173 modeli\u0173 veikim\u0105 naudojant ma\u017eesni\u0173 kalb\u0173 duomenis vis dar tr\u016bksta informacijos. Paprastam vartotojui toki\u0173 modeli\u0173 auk\u0161tas tikslumas atrodo \u012fsp\u016bdingas, ta\u010diau yra sri\u010di\u0173, pavyzd\u017eiui, medicina ar teis\u0117, kur reikalavimai yra gerokai grie\u017etesni, nes klaidos gali tur\u0117ti rimt\u0173 pasekmi\u0173. Be to, daugelis populiariausi\u0173 modeli\u0173 (pvz., <em>GPT<\/em>)  yra u\u017edari, priklauso i\u0161orin\u0117ms kompanijoms, kurios valdo pa\u010dius modelius ir kontroliuoja duomen\u0173 srautus. D\u0117l to toki\u0173 modeli\u0173 taikymas tampa sud\u0117tingesnis, ypa\u010d kai svarbu u\u017etikrinti jautri\u0173 duomen\u0173 saugum\u0105.<\/p>\n<p>\u00a0<\/p>\n<p><!-- \/wp:paragraph --><\/p>\n<p><!-- wp:paragraph --><\/p>\n<p>Alternatyva? <strong>Atvir\u0173 svori\u0173 modeliai!<\/strong> Juos galima tiesiog parsisi\u0173sti \u012f savo kompiuter\u012f ir naudoti dviem b\u016bdais: 1)\u202fpakoreguojant su savo duomenimis arba 2) be pakeitim\u0173, tiesiog teisingai suformuluojant u\u017eklausas. Pirmasis variantas yra priimtinas ne visiems vartotojams, nes reikia papildom\u0173 mokymo duomen\u0173 ir pakankamai dideli\u0173 skai\u010diavimo paj\u0117gum\u0173. Tod\u0117l buvo svarbu i\u0161testuoti antrojo varianto galimybes: atliktas tyrimas detaliai apra\u0161ytas <a href=\"https:\/\/arxiv.org\/abs\/2501.03952\" target=\"_blank\" rel=\"noopener\"><strong>mokslin\u0117je publikacijoje,<\/strong><\/a> kuri \u012ftraukta \u012f \u201eNoDaLiDa &amp; Baltic HLT\u201c konferencijos program\u0105.<\/p>\n<p>\u00a0<\/p>\n<h2>M\u016bs\u0173 tyrimas: 12 LLM i\u0161bandyta lietuvi\u0173, latvi\u0173 ir est\u0173 kalbomis<\/h2>\n<p><!-- \/wp:paragraph --><\/p>\n<p><!-- wp:paragraph --><\/p>\n<p>\u0160io tyrimo metu testavome <strong>12 skirting\u0173 kalbos modeli\u0173,<\/strong> siekdami \u012fvertinti j\u0173 geb\u0117jim\u0105 suprasti ir generuoti tekstus lietuvi\u0173, latvi\u0173 bei est\u0173 kalbomis. I\u0161band\u0117me tiek mokamas u\u017edarojo kodo modeli\u0173 versijas \u2013<em>\u201eGPT-3.5<\/em> Turbo\u201c, <em>\u201eGPT-4\u201c<\/em>, ir <em>\u201eGPT-4o\u201c<\/em>tiek atvir\u0173 svori\u0173 modelius:<em>\u201eLlama\u201c<\/em>(3, 3.1, 3.2 versijas su 3, 8 ir 70 mlrd. parametr\u0173), <em>\u201eMistral\u201c<\/em> (12 mlrd.), <em>\u201eGemma2\u201c<\/em> (9 ir 27 mlrd.) ir <em>\u201ePhi\u201c<\/em> su 3 ir 14 mlrd.). Tai vienas i\u0161 nedaugelio vie\u0161\u0173 dirbtinio intelekto modeli\u0173, skirt\u0173 ma\u017eesn\u0117ms kalboms, etalon\u0173.<\/p>\n<p>\u00a0<\/p>\n<p><!-- \/wp:paragraph --><\/p>\n<p><!-- wp:paragraph --><\/p>\n<p>Pirmasis eksperimentas buvo skirtas <a href=\"https:\/\/staging.tilde.ai\/lt\/translate\/tilde-machine-translation\/\"><strong>ma\u0161ininio vertimo<\/strong><\/a>tikslumui \u012fvertinti, lyginant vertimus i\u0161 angl\u0173 \u012f lietuvi\u0173\/latvi\u0173\/est\u0173 kalbas ir atvirk\u0161\u010diai. GPT modeli\u0173 rezultatai buvo auk\u0161\u010diausi, ta\u010diau nustebino ir 27 mlrd. parametr\u0173 <em>\u201eGemma2\u201c<\/em> bei 70 mlrd. parametr\u0173 <em>\u201eLlama3.1\u201c<\/em> , kuri\u0173 vertimo kokyb\u0117 buvo gana artima <em>GPT<\/em> modeliams. Pras\u010diausiai veik\u0117 <em>\u201ePhi\u201c<\/em> modeliai.  \nDid\u017ei\u0173j\u0173 kalbos modeli\u0173 rezultatus palyginome su <em>\u201eDeepL\u201c<\/em>\u2013 viena pa\u017eangiausi\u0173 ma\u0161ininio vertimo sistem\u0173 \u2013 ir pasteb\u0117jome, kad jos vertimo kokyb\u0117 prilygo <em>\u201eGPT-4o\u201c<\/em>modeliui. Tai rodo, kad didieji kalbos modeliai jau geba generuoti vertimus, konkuruojan\u010dius su specializuotomis vertimo sistemomis.<\/p>\n<p>\u00a0<\/p>\n<p><!-- \/wp:paragraph --><\/p>\n<p><!-- wp:paragraph --><\/p>\n<p>Kitoje u\u017eduotyje modeliai tur\u0117jo <strong>atsakyti \u012f klausimus<\/strong> lietuvi\u0173, latvi\u0173 ir est\u0173 kalbomis, pasirinkdami tinkamo atsakymo varianto ID. \u0160i u\u017eduotis reikalavo ne tik suprasti klausim\u0105 ir \u017einoti atsakym\u0105, bet ir geb\u0117ti j\u012f pateikti tinkama forma. Geriausiai pasirod\u0117 tie patys trys modeliai: <em>\u201eGPT-4o\u201c<\/em>, <em>\u201eLlama3.1\u201c<\/em> (70 mlrd.) ir <em>\u201eGemma2\u201c<\/em> (27 mlrd.). Rezultatus palyginome su angl\u0173 kalba \u2013 skirtumas akivaizdus. Tikslumas ma\u017eosiomis kalbomis vis dar gerokai atsilieka nuo did\u017ei\u0173j\u0173 kalb\u0173.<\/p>\n<p>\u00a0<\/p>\n<p><!-- \/wp:paragraph --><\/p>\n<p><!-- wp:paragraph --><\/p>\n<p>Tre\u010diojo eksperimento metu tikrinome, <strong>kaip didieji kalbos modeliai geba atsakyti \u012f atvirus \u012fvairi\u0173 sri\u010di\u0173 klausimus<\/strong> lietuvi\u0173 ir latvi\u0173 kalbomis. Geriausi modeliai pasiek\u0117 80\u201390 % tikslum\u0105. Taip pat vertinome j\u0173 generuojamo teksto sklandum\u0105, bet ir \u010dia lyderi\u0173 s\u0105ra\u0161as nesikeit\u0117. Be to, vertindami rezultatus lietuvi\u0173 kalba, i\u0161band\u0117me ir <em>\u201eNeurotechnology\u201c<\/em> sukurtus vienakalbius <em>\u201eLlama2\u201c<\/em>(7 ir 13 mlrd. parametr\u0173) modelius, pritaikytus b\u016btent lietuvi\u0173 kalbai. Jie geb\u0117jo generuoti itin skland\u0173 lietuvi\u0161k\u0105 tekst\u0105, ta\u010diau atsakym\u0173 tikslumu vis dar gerokai nusileido didiesiems daugiakalbiams modeliams.<\/p>\n<p>\u00a0<\/p>\n<h2>K\u0105 LLM rezultatai rei\u0161kia ma\u017eesn\u0117ms kalboms<\/h2>\n<p><!-- \/wp:paragraph --><\/p>\n<p><!-- wp:paragraph --><\/p>\n<p>\u0160is tyrimas tik dar kart\u0105 patvirtino, koks <strong>svarbus yra kokybi\u0161k\u0173 duomen\u0173 kiekis ir \u012fvairov\u0117. <\/strong>Jei norime, kad kalb\u0173 modeliai pasiekt\u0173 tok\u012f pat\u012f tikslumo lyg\u012f ma\u017eesn\u0117se kalbose, kaip ir didesn\u0117se, b\u016btinas aktyvus bendradarbiavimas mokslo bendruomen\u0117je. Tai apima nuolatin\u012f duomen\u0173 rinkim\u0105 ir specializuot\u0173 modeli\u0173 k\u016brim\u0105. Taip dirbtinio intelekto modeliai ma\u017eesn\u0117ms kalboms suma\u017eins atotr\u016bk\u012f. <strong>Tod\u0117l ne konkuruokime, o bendradarbiaukime!<\/strong><\/p>\n<div class=\"wp-block-qi-blocks-call-to-action qodef-block-container qodef-block-7fff17c8\">\n<div class=\"qi-block-call-to-action qodef-block qodef-m qodef-layout--standard\">\n<div class=\"qodef-m-inner\">\n<div class=\"qodef-m-button\">\n<p><!-- \/wp:qi-blocks\/button --><\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p><!-- \/wp:qi-blocks\/call-to-action --><\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-55ea5ed e-flex e-con-boxed e-con e-parent\" data-id=\"55ea5ed\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t<div class=\"elementor-element elementor-element-b0a0897 e-con-full e-flex e-con e-child\" data-id=\"b0a0897\" data-element_type=\"container\" data-e-type=\"container\" data-settings=\"{&quot;background_background&quot;:&quot;classic&quot;}\">\n\t\t\t\t<div class=\"elementor-element elementor-element-a4a1bc8 elementor-widget elementor-widget-heading\" data-id=\"a4a1bc8\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"heading.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t<h2 class=\"elementor-heading-title elementor-size-default\">I\u0161tirkite DI modelius Europos kalboms<\/h2>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-64a9f0e elementor-widget elementor-widget-text-editor\" data-id=\"64a9f0e\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t<div class=\"qodef-m-content\">\n<div class=\"qodef-m-text\">Ma\u017eesn\u0117ms kalboms reikia kokybi\u0161k\u0173 duomen\u0173, vietin\u0117s kalbos ekspert\u0173 ir saugios AI infrastrukt\u016bros. Su\u017einokite daugiau apie <a href=\"https:\/\/staging.tilde.ai\/lt\/tildeopen-llm\/\">\u201eTildeOpen LLM\u201c<\/a> ir <a href=\"https:\/\/staging.tilde.ai\/lt\/ai-data-services\/\">Tild\u0117s AI duomen\u0173 paslaugos<\/a> kalbos technologij\u0173 k\u016brimui, vertinimui ir pritaikymui.<\/div>\n<\/div>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-4861db1 e-con-full e-flex e-con e-child\" data-id=\"4861db1\" data-element_type=\"container\" data-e-type=\"container\" data-settings=\"{&quot;background_background&quot;:&quot;classic&quot;}\">\n\t\t\t\t<div class=\"elementor-element elementor-element-839a1f5 elementor-widget elementor-widget-button\" data-id=\"839a1f5\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"button.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t<div class=\"elementor-button-wrapper\">\n\t\t\t\t\t<a class=\"elementor-button elementor-button-link elementor-size-sm\" href=\"https:\/\/staging.tilde.ai\/lt\/tildeopen-llm\/\">\n\t\t\t\t\t\t<span class=\"elementor-button-content-wrapper\">\n\t\t\t\t\t\t\t\t\t<span class=\"elementor-button-text\">I\u0161bandykite TildeOpen LLM<\/span>\n\t\t\t\t\t<\/span>\n\t\t\t\t\t<\/a>\n\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-028bf2b e-flex e-con-boxed e-con e-parent\" data-id=\"028bf2b\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-619bd10 elementor-widget elementor-widget-shortcode\" data-id=\"619bd10\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"shortcode.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<div class=\"elementor-shortcode\">\t\t<div data-elementor-type=\"section\" data-elementor-id=\"20957\" class=\"elementor elementor-20957\" data-elementor-post-type=\"elementor_library\">\n\t\t\t<div class=\"elementor-element elementor-element-acd72f3 e-flex e-con-boxed e-con e-parent\" data-id=\"acd72f3\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-2a7fada elementor-widget elementor-widget-heading\" data-id=\"2a7fada\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"heading.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t<h4 class=\"elementor-heading-title elementor-size-default\">Kas naujo \u201eTilde\u201c\n<\/h4>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-cafb956 e-flex e-con-boxed e-con e-parent\" data-id=\"cafb956\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-84bcf49 elementor-grid-4 elementor-posts--align-left e-transform elementor-grid-tablet-2 elementor-grid-mobile-1 elementor-posts--thumbnail-top elementor-widget elementor-widget-posts\" data-id=\"84bcf49\" data-element_type=\"widget\" data-e-type=\"widget\" data-settings=\"{&quot;classic_columns&quot;:&quot;4&quot;,&quot;classic_row_gap&quot;:{&quot;unit&quot;:&quot;em&quot;,&quot;size&quot;:1.5,&quot;sizes&quot;:[]},&quot;classic_row_gap_tablet&quot;:{&quot;unit&quot;:&quot;em&quot;,&quot;size&quot;:&quot;&quot;,&quot;sizes&quot;:[]},&quot;classic_row_gap_mobile&quot;:{&quot;unit&quot;:&quot;em&quot;,&quot;size&quot;:&quot;&quot;,&quot;sizes&quot;:[]},&quot;classic_columns_tablet&quot;:&quot;2&quot;,&quot;classic_columns_mobile&quot;:&quot;1&quot;,&quot;_transform_translateX_effect_hover&quot;:{&quot;unit&quot;:&quot;px&quot;,&quot;size&quot;:&quot;&quot;,&quot;sizes&quot;:[]},&quot;_transform_translateX_effect_hover_tablet&quot;:{&quot;unit&quot;:&quot;px&quot;,&quot;size&quot;:&quot;&quot;,&quot;sizes&quot;:[]},&quot;_transform_translateX_effect_hover_mobile&quot;:{&quot;unit&quot;:&quot;px&quot;,&quot;size&quot;:&quot;&quot;,&quot;sizes&quot;:[]},&quot;_transform_translateY_effect_hover&quot;:{&quot;unit&quot;:&quot;px&quot;,&quot;size&quot;:&quot;&quot;,&quot;sizes&quot;:[]},&quot;_transform_translateY_effect_hover_tablet&quot;:{&quot;unit&quot;:&quot;px&quot;,&quot;size&quot;:&quot;&quot;,&quot;sizes&quot;:[]},&quot;_transform_translateY_effect_hover_mobile&quot;:{&quot;unit&quot;:&quot;px&quot;,&quot;size&quot;:&quot;&quot;,&quot;sizes&quot;:[]}}\" data-widget_type=\"posts.classic\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<div class=\"elementor-posts-container elementor-posts elementor-posts--skin-classic elementor-grid\" role=\"list\">\n\t\t\t\t<article class=\"elementor-post elementor-grid-item post-26289 post type-post status-publish format-standard has-post-thumbnail hentry category-blog tag-ai-data-services tag-llm\" role=\"listitem\">\n\t\t\t\t<a class=\"elementor-post__thumbnail__link\" href=\"https:\/\/staging.tilde.ai\/lt\/blog\/ai-adoption-barriers\/\" tabindex=\"-1\">\n\t\t\t<div class=\"elementor-post__thumbnail\"><img decoding=\"async\" width=\"750\" height=\"317\" src=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/3_vz_lt_blogam-3.png\" class=\"attachment-full size-full wp-image-26283\" alt=\"Verslo profesionalas vertina AI diegimo kli\u016btis ir duomen\u0173 saugumo rizik\u0105\" srcset=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/3_vz_lt_blogam-3.png 750w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/3_vz_lt_blogam-3-300x127.png 300w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/3_vz_lt_blogam-3-18x8.png 18w\" sizes=\"(max-width: 750px) 100vw, 750px\" \/><\/div>\n\t\t<\/a>\n\t\t\t\t<div class=\"elementor-post__text\">\n\t\t\t\t<h4 class=\"elementor-post__title\">\n\t\t\t<a href=\"https:\/\/staging.tilde.ai\/lt\/blog\/ai-adoption-barriers\/\">\n\t\t\t\tDirbtinio intelekto diegimo kli\u016btys, d\u0117l kuri\u0173 \u012fmon\u0117s dvejoja diegti DI\t\t\t<\/a>\n\t\t<\/h4>\n\t\t\t\t<\/div>\n\t\t\t\t<\/article>\n\t\t\t\t<article class=\"elementor-post elementor-grid-item post-26172 post type-post status-publish format-standard has-post-thumbnail hentry category-blog tag-machine-translation\" role=\"listitem\">\n\t\t\t\t<a class=\"elementor-post__thumbnail__link\" href=\"https:\/\/staging.tilde.ai\/lt\/blog\/will-the-ai-era-swallow-up-languages-with-a-small-number-of-speakers\/\" tabindex=\"-1\">\n\t\t\t<div class=\"elementor-post__thumbnail\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1668\" src=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/Tilde2092025-4-scaled.jpg\" class=\"attachment-full size-full wp-image-26173\" alt=\"\" srcset=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/Tilde2092025-4-scaled.jpg 2560w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/Tilde2092025-4-300x196.jpg 300w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/Tilde2092025-4-1024x667.jpg 1024w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/Tilde2092025-4-768x500.jpg 768w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/Tilde2092025-4-1536x1001.jpg 1536w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/Tilde2092025-4-2048x1335.jpg 2048w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/05\/Tilde2092025-4-18x12.jpg 18w\" sizes=\"(max-width: 2560px) 100vw, 2560px\" \/><\/div>\n\t\t<\/a>\n\t\t\t\t<div class=\"elementor-post__text\">\n\t\t\t\t<h4 class=\"elementor-post__title\">\n\t\t\t<a href=\"https:\/\/staging.tilde.ai\/lt\/blog\/will-the-ai-era-swallow-up-languages-with-a-small-number-of-speakers\/\">\n\t\t\t\tAr dirbtinio intelekto era \u201epraris\u201c kalbas, kuriomis kalba nedaug \u017emoni\u0173?\u00a0\t\t\t<\/a>\n\t\t<\/h4>\n\t\t\t\t<\/div>\n\t\t\t\t<\/article>\n\t\t\t\t<article class=\"elementor-post elementor-grid-item post-26124 post type-post status-publish format-standard has-post-thumbnail hentry category-news\" role=\"listitem\">\n\t\t\t\t<a class=\"elementor-post__thumbnail__link\" href=\"https:\/\/staging.tilde.ai\/lt\/news\/tildeopen-will-henceforth-process-8-times-larger-text-volumes\/\" tabindex=\"-1\">\n\t\t\t<div class=\"elementor-post__thumbnail\"><img loading=\"lazy\" decoding=\"async\" width=\"800\" height=\"510\" src=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/TildeOpen_8x_web.jpg\" class=\"attachment-full size-full wp-image-26149\" alt=\"\" srcset=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/TildeOpen_8x_web.jpg 800w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/TildeOpen_8x_web-300x191.jpg 300w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/TildeOpen_8x_web-768x490.jpg 768w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/TildeOpen_8x_web-18x12.jpg 18w\" sizes=\"(max-width: 800px) 100vw, 800px\" \/><\/div>\n\t\t<\/a>\n\t\t\t\t<div class=\"elementor-post__text\">\n\t\t\t\t<h4 class=\"elementor-post__title\">\n\t\t\t<a href=\"https:\/\/staging.tilde.ai\/lt\/news\/tildeopen-will-henceforth-process-8-times-larger-text-volumes\/\">\n\t\t\t\tNuo \u0161iol \u201eTildeOpen\u201c apdoros 8 kartus didesnius teksto kiekius\t\t\t<\/a>\n\t\t<\/h4>\n\t\t\t\t<\/div>\n\t\t\t\t<\/article>\n\t\t\t\t<article class=\"elementor-post elementor-grid-item post-25645 post type-post status-publish format-standard has-post-thumbnail hentry category-news\" role=\"listitem\">\n\t\t\t\t<a class=\"elementor-post__thumbnail__link\" href=\"https:\/\/staging.tilde.ai\/lt\/news\/secure-ai-solutions-in-government\/\" tabindex=\"-1\">\n\t\t\t<div class=\"elementor-post__thumbnail\"><img loading=\"lazy\" decoding=\"async\" width=\"1200\" height=\"675\" src=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/AI-solutions-in-government-institutions-risk.webp\" class=\"attachment-full size-full wp-image-25830\" alt=\"AI saugos kompiuteris\" srcset=\"https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/AI-solutions-in-government-institutions-risk.webp 1200w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/AI-solutions-in-government-institutions-risk-300x169.webp 300w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/AI-solutions-in-government-institutions-risk-1024x576.webp 1024w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/AI-solutions-in-government-institutions-risk-768x432.webp 768w, https:\/\/staging.tilde.ai\/wp-content\/uploads\/2026\/04\/AI-solutions-in-government-institutions-risk-18x10.webp 18w\" sizes=\"(max-width: 1200px) 100vw, 1200px\" \/><\/div>\n\t\t<\/a>\n\t\t\t\t<div class=\"elementor-post__text\">\n\t\t\t\t<h4 class=\"elementor-post__title\">\n\t\t\t<a href=\"https:\/\/staging.tilde.ai\/lt\/news\/secure-ai-solutions-in-government\/\">\n\t\t\t\tU\u017esienieti\u0161k\u0173 DI sprendim\u0173 naudojimas valstybin\u0117se institucijose: didel\u0117 rizika saugumui, bet sprendimas jau yra\u00a0\t\t\t<\/a>\n\t\t<\/h4>\n\t\t\t\t<\/div>\n\t\t\t\t<\/article>\n\t\t\t\t<\/div>\n\t\t\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>","protected":false},"excerpt":{"rendered":"<p>Nuo statistini\u0173 kalb\u0173 modeli\u0173 atsiradimo 1980-aisiais, \u0161i sritis patyr\u0117 nepaprast\u0105 transformacij\u0105. \u0160iandien tikrasis klausimas yra, kaip AI modeliai<\/p>","protected":false},"author":23,"featured_media":18164,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[108],"tags":[148,149],"class_list":["post-18159","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog","tag-ai-data-services","tag-llm"],"_links":{"self":[{"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/posts\/18159","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/users\/23"}],"replies":[{"embeddable":true,"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/comments?post=18159"}],"version-history":[{"count":21,"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/posts\/18159\/revisions"}],"predecessor-version":[{"id":27340,"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/posts\/18159\/revisions\/27340"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/media\/18164"}],"wp:attachment":[{"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/media?parent=18159"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/categories?post=18159"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/staging.tilde.ai\/lt\/wp-json\/wp\/v2\/tags?post=18159"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}