{"id":116805,"date":"2024-06-27T18:05:52","date_gmt":"2024-06-27T16:05:52","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/rejting-otkrytosti-generativnyh-ai-modelej"},"modified":"2024-06-27T18:05:52","modified_gmt":"2024-06-27T16:05:52","slug":"rejting-otkrytosti-generativnyh-ai-modelej","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/news\/rejting-otkrytosti-generativnyh-ai-modelej","title":{"rendered":"Vler\u00ebsimi i transparenc\u00ebs s\u00eb modeleve gjeneruese t\u00eb AI","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

K\u00ebrkuesit nga Universiteti i Nijmegenit (Holand\u00eb) p\u00ebrgatit\u00ebn nj\u00eb renditje t\u00eb hapjes s\u00eb 40 modeleve t\u00eb m\u00ebdha gjuh\u00ebsore dhe 7 modeleve p\u00ebr gjenerimin e imazheve sipas p\u00ebrshkrimit tekstual, t\u00eb cilat jan\u00eb deklaruar nga prodhuesit si t\u00eb hapura. P\u00ebr shkak se kriteret e hapjes s\u00eb modeleve t\u00eb m\u00ebsimit t\u00eb makinerive ende po formohen, aktualisht ka nj\u00eb situat\u00eb ku, n\u00ebn emrin e modeleve t\u00eb hapura, shp\u00ebrndahen modele q\u00eb kan\u00eb licenca q\u00eb kufizojn\u00eb fush\u00ebn e p\u00ebrdorimit (p\u00ebr shembull, shum\u00eb modele ndalojn\u00eb p\u00ebrdorimin n\u00eb projekte komerciale). Gjithashtu, shpesh prodhuesit nuk ofrojn\u00eb qasje n\u00eb t\u00eb dh\u00ebnat e p\u00ebrdorura p\u00ebr trajnimin, nuk zbulojn\u00eb detajet e implementimit ose nuk hapin plot\u00ebsisht kodin mb\u00ebsht\u00ebs.<\/p>\n

Shumica e modeleve t\u00eb pozicionuara si \"t\u00eb hapura\" n\u00eb t\u00eb v\u00ebrtet\u00eb duhet t\u00eb percepohen si \"koeficient\u00eb t\u00eb hapur\" ose m\u00eb sakt\u00eb \"koeficient\u00eb t\u00eb disponuesh\u00ebm\", pasi ato shp\u00ebrndahen n\u00ebn licenca kufizuese q\u00eb ndalojn\u00eb p\u00ebrdorimin n\u00eb produkte komerciale. Hulumtuesit e pal\u00ebve t\u00eb treta mund t\u00eb eksperimentohet me k\u00ebto modele, por nuk kan\u00eb mund\u00ebsin\u00eb ta adaptojn\u00eb modelin sipas nevojave t\u00eb tyre ose ta inspektojn\u00eb implementimin. M\u00eb shum\u00eb se gjysma e modeleve nuk ofrojn\u00eb informacion t\u00eb detajuar mbi t\u00eb dh\u00ebnat q\u00eb jan\u00eb p\u00ebrdorur p\u00ebr trajnimin, si dhe nuk publikojn\u00eb informacione mbi struktur\u00ebn dhe arkitektur\u00ebn e brendshme.<\/p>\n

Modelet m\u00eb t\u00eb hapura jan\u00eb njohur si BloomZ, AmberChat, OLMo, Open Assistant dhe Stable Diffusion, t\u00eb cilat jan\u00eb publikuar n\u00ebn licenca t\u00eb hapura s\u00eb bashku me t\u00eb dh\u00ebnat burimore, kodin dhe realizimin e API-s\u00eb. Modelet nga Google (Gemma 7B), Microsoft (Orca 2) dhe Meta (Llama 3), t\u00eb pozicionuara nga prodhuesit si t\u00eb hapura, kan\u00eb p\u00ebrfunduar m\u00eb af\u00ebr fundin e renditjes, pasi ato nuk ofrojn\u00eb qasje n\u00eb t\u00eb dh\u00ebnat burimore, nuk zbulojn\u00eb detajet teknike t\u00eb realizimit, dhe koeficient\u00ebt e modelit shp\u00ebrndahen n\u00ebn licenca q\u00eb kufizojn\u00eb fush\u00ebn e p\u00ebrdorimit. Modeli i njohur Mistral 7B p\u00ebrfundoi af\u00ebrsisht n\u00eb mes t\u00eb renditjes, pasi ofrohet n\u00ebn nj\u00eb licenc\u00eb t\u00eb hapur, por \u00ebsht\u00eb dokumentuar vet\u00ebm pjes\u00ebrisht, nuk zbulon t\u00eb dh\u00ebnat e p\u00ebrdorura p\u00ebr trajnim dhe ka kodin mb\u00ebsht\u00ebs q\u00eb nuk \u00ebsht\u00eb plot\u00ebsisht i hapur.<\/p>\n

K\u00ebrkuesit kan\u00eb propozuar 14 kritere p\u00ebr hap\u00ebsir\u00ebn e modeleve AI, duke p\u00ebrfshir\u00eb kushtet p\u00ebr shp\u00ebrndarjen e kodit, t\u00eb dh\u00ebnave p\u00ebr trajnimin, koeficient\u00ebve, opsioneve t\u00eb t\u00eb dh\u00ebnave dhe koeficient\u00ebve t\u00eb optimizuar p\u00ebrmes m\u00ebsimit t\u00eb forc\u00ebs (RL), si dhe disponueshm\u00ebrin\u00eb e pakove t\u00eb gatshme p\u00ebr p\u00ebrdorim, API-ve, dokumentacionit dhe p\u00ebrshkrimeve t\u00eb detajuara t\u00eb implementimit.<\/p>\n

\"Vler\u00ebsimi<\/center><\/p>\n

\"Vler\u00ebsimi<\/center><\/p>\n

Sipas versionit t\u00eb par\u00eb t\u00eb propozitit nga organizata OSI (Open Source Initiative) p\u00ebr definimin e hap\u00ebsir\u00ebs s\u00eb AI, kriteret kryesore t\u00eb hap\u00ebsir\u00ebs s\u00eb sistemit AI jan\u00eb ofrimi i mund\u00ebsive p\u00ebr p\u00ebrdorim p\u00ebr q\u00ebllime t\u00eb ndryshme pa pasur nevoj\u00eb p\u00ebr nj\u00eb leje t\u00eb ve\u00e7ant\u00eb; studimi i funksionimit t\u00eb sistemit dhe inspektimi i komponent\u00ebve t\u00eb tij; b\u00ebrja e ndryshimeve p\u00ebr q\u00ebllime t\u00eb ndryshme; transferimi te pal\u00ebt e tjera si variant origjinal ashtu edhe modifikimet pas ndryshimeve.<\/p>\n

P\u00ebr t\u00eb ofruar mund\u00ebsin\u00eb e ndryshimeve, sistemi AI duhet t\u00eb p\u00ebrfshij\u00eb:<\/p>\n