{"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\/cs\/blog\/news\/rejting-otkrytosti-generativnyh-ai-modelej","title":{"rendered":"\u017deb\u0159\u00ed\u010dek otev\u0159enosti generativn\u00edch AI model\u016f","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

V\u00fdzkumn\u00edci z Radboudsk\u00e9 univerzity v Nijmegen (Nizozemsko) p\u0159ipravili \u017eeb\u0159\u00ed\u010dek otev\u0159enosti 40 velk\u00fdch jazykov\u00fdch model\u016f a 7 model\u016f pro generov\u00e1n\u00ed obr\u00e1zk\u016f na z\u00e1klad\u011b textov\u00e9ho popisu, kter\u00e9 v\u00fdrobci ozna\u010duj\u00ed jako otev\u0159en\u00e9. Vzhledem k tomu, \u017ee krit\u00e9ria otev\u0159enosti model\u016f strojov\u00e9ho u\u010den\u00ed se teprve formuj\u00ed, vznikla situace, kdy se pod z\u00e1minkou otev\u0159enosti \u0161\u00ed\u0159\u00ed modely, kter\u00e9 maj\u00ed licenci omezuj\u00edc\u00ed rozsah pou\u017eit\u00ed (nap\u0159\u00edklad mnoh\u00e9 modely zakazuj\u00ed u\u017eit\u00ed v komer\u010dn\u00edch projektech). \u010casto tak\u00e9 v\u00fdrobci neposkytuj\u00ed p\u0159\u00edstup k dat\u016fm pou\u017eit\u00fdm p\u0159i tr\u00e9nov\u00e1n\u00ed, neodhaluj\u00ed detaily implementace nebo pln\u011b neotev\u0159ou souvisej\u00edc\u00ed k\u00f3d.<\/p>\n

V\u011bt\u0161ina model\u016f, kter\u00e9 se prezentuj\u00ed jako \u201eotev\u0159en\u00e9\u201c, by m\u011bla b\u00fdt ve skute\u010dnosti vn\u00edm\u00e1na jako \u201eotev\u0159en\u00e9 v\u00e1hov\u00e9 koeficienty\u201c nebo p\u0159esn\u011bji \u201edostupn\u00e9 v\u00e1hov\u00e9 koeficienty\u201c, proto\u017ee jsou distribuov\u00e1ny pod omezen\u00fdmi licencemi, je\u017e zakazuj\u00ed pou\u017eit\u00ed v komer\u010dn\u00edch produktech. Zadn\u00ed v\u00fdzkumn\u00ed pracovn\u00edci mohou experimentovat s t\u011bmito modely, ale nemaj\u00ed mo\u017enost p\u0159izp\u016fsobit model sv\u00fdm pot\u0159eb\u00e1m nebo prozkoumat implementaci. V\u00edce ne\u017e polovina model\u016f neposkytuje podrobn\u00e9 informace o datech pou\u017eit\u00fdch pro tr\u00e9nink a tak\u00e9 nezve\u0159ej\u0148uje informace o vnit\u0159n\u00ed struktu\u0159e a architektu\u0159e.<\/p>\n

Nejotev\u0159en\u011bj\u0161\u00edmi modely jsou pova\u017eov\u00e1ny BloomZ, AmberChat, OLMo, Open Assistant a Stable Diffusion, kter\u00e9 byly publikov\u00e1ny pod otev\u0159en\u00fdmi licencemi spolu se vstupn\u00edmi daty, k\u00f3dem a implementac\u00ed API. Modely od Googlu (Gemma 7B), Microsoftu (Orca 2) a Meta (Llama 3), kter\u00e9 jsou v\u00fdrobci ozna\u010dov\u00e1ny jako otev\u0159en\u00e9, skon\u010dily bl\u00ed\u017ee konci \u017eeb\u0159\u00ed\u010dku, proto\u017ee neposkytuj\u00ed p\u0159\u00edstup k vstupn\u00edm dat\u016fm, neodhaluj\u00ed technick\u00e9 detaily implementace a v\u00e1hov\u00e9 koeficienty modelu \u0161\u00ed\u0159\u00ed pod licencemi, kter\u00e9 omezuj\u00ed rozsah pou\u017eit\u00ed. Popul\u00e1rn\u00ed model Mistral 7B se um\u00edstil p\u0159ibli\u017en\u011b v polovin\u011b \u017eeb\u0159\u00ed\u010dku, proto\u017ee je dod\u00e1v\u00e1n pod otev\u0159enou licenc\u00ed, ale je pouze \u010d\u00e1ste\u010dn\u011b zdokumentov\u00e1n, nezve\u0159ej\u0148uje data pou\u017eit\u00e1 p\u0159i tr\u00e9nov\u00e1n\u00ed a m\u00e1 ne\u00fapln\u011b otev\u0159en\u00fd souvisej\u00edc\u00ed k\u00f3d.<\/p>\n

V\u00fdzkumn\u00edci navrhli 14 krit\u00e9ri\u00ed otev\u0159enosti AI model\u016f, kter\u00e1 zahrnuj\u00ed podm\u00ednky pro zve\u0159ejn\u011bn\u00ed k\u00f3du, dat pro u\u010den\u00ed, vah, variant dat a parametr\u016f optimalizovan\u00fdch pomoc\u00ed u\u010den\u00ed s posilov\u00e1n\u00edm (RL), jako\u017e i dostupnost bal\u00ed\u010dk\u016f, API, dokumentace a podrobn\u00e9ho popisu implementace.<\/p>\n

\"\u017deb\u0159\u00ed\u010dek<\/center><\/p>\n

\"\u017deb\u0159\u00ed\u010dek<\/center><\/p>\n

Podle navr\u017een\u00e9ho konceptu definice otev\u0159en\u00e9ho AI organizac\u00ed OSI (Open Source Initiative) jsou hlavn\u00edmi krit\u00e9rii otev\u0159enosti AI syst\u00e9mu poskytnut\u00ed mo\u017enost\u00ed vyu\u017eit\u00ed k jak\u00fdmkoli \u00fa\u010del\u016fm bez nutnosti z\u00edsk\u00e1n\u00ed samostatn\u00e9ho povolen\u00ed; studium \u010dinnosti syst\u00e9mu a inspekce jeho komponent; prov\u00e1d\u011bn\u00ed zm\u011bn pro jak\u00e9koli \u00fa\u010dely; p\u0159ed\u00e1n\u00ed jin\u00fdm osob\u00e1m jak p\u016fvodn\u00ed verze, tak i edice po proveden\u00ed zm\u011bn.<\/p>\n

Aby bylo mo\u017en\u00e9 prov\u00e1d\u011bt zm\u011bny, AI syst\u00e9m mus\u00ed zahrnovat:<\/p>\n