{"id":118790,"date":"2024-08-01T15:43:32","date_gmt":"2024-08-01T13:43:32","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/google-predstavila-kompaktnuyu-yazykovuyu-model-gemma-2-2b-kotoraya-prevoshodit-gpt-3-5-turbo"},"modified":"2024-08-01T15:43:32","modified_gmt":"2024-08-01T13:43:32","slug":"google-predstavila-kompaktnuyu-yazykovuyu-model-gemma-2-2b-kotoraya-prevoshodit-gpt-3-5-turbo","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/news\/google-predstavila-kompaktnuyu-yazykovuyu-model-gemma-2-2b-kotoraya-prevoshodit-gpt-3-5-turbo","title":{"rendered":"Google prezantoi modelin kompakt t\u00eb gjuh\u00ebs Gemma 2 2B, i cili tejkalon GPT 3.5 Turbo","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

Kompania Google ka prezantuar<\/a><\/noindex> Gemma 2 2B \u2014 nj\u00eb model i kompakt, por i fuqish\u00ebm i gjuh\u00ebs s\u00eb inteligjenc\u00ebs artificiale (LLM), i cili mund t\u00eb konkurroj\u00eb me lider\u00ebt e industris\u00eb, pavar\u00ebsisht nga madh\u00ebsia e tij duksh\u00ebm m\u00eb e vog\u00ebl. Modeli i ri i gjuh\u00ebs, q\u00eb p\u00ebrmban vet\u00ebm 2.6 miliard parameter, tregon performanc\u00eb t\u00eb ngjashme me homolog\u00ebt shum\u00eb m\u00eb t\u00eb m\u00ebdhenj, duke p\u00ebrfshir\u00eb OpenAI GPT-3.5 dhe Mistral AI Mixtral 8x7B.<\/p>\n

<\/center> \t\t\t\t \"Google <\/p>\n

N\u00eb testin e LMSYS Chatbot Arena, nj\u00eb platform\u00eb online popullore p\u00ebr testimin dhe vler\u00ebsimin e cil\u00ebsis\u00eb s\u00eb modeleve t\u00eb inteligjenc\u00ebs artificiale, Gemma 2 2B arriti 1130 pik\u00eb. Ky rezultat \u00ebsht\u00eb pak m\u00eb i lart\u00eb se ai i GPT-3.5-Turbo-0613 (1117 pik\u00eb) dhe Mixtral-8x7B (1114 pik\u00eb) \u2014 modele q\u00eb kan\u00eb dhjet\u00ebfishin e numrit t\u00eb parametrave.<\/p>\n

\"Google <\/p>\n

Google njofton se Gemma 2 2B gjithashtu arriti 56.1 pik\u00eb n\u00eb testin MMLU (Massive Multitask Language Understanding) dhe 36.6 pik\u00eb n\u00eb testin MBPP (Mostly Basic Python Programming), duke sh\u00ebnuar nj\u00eb p\u00ebrmir\u00ebsim t\u00eb duksh\u00ebm n\u00eb krahasim me versionin paraardh\u00ebs. <\/p>\n

<\/center> <\/p>\n

Gemma 2 2B sfidon mendimin e zakonsh\u00ebm q\u00eb modelet m\u00eb t\u00eb m\u00ebdha punojn\u00eb m\u00eb mir\u00eb n\u00eb fillim se ato kompakte. Performanca e Gemma 2 2B tregon se metodat komplekse t\u00eb m\u00ebsimit, efikasiteti i arkitektur\u00ebs dhe setet e dh\u00ebnash t\u00eb cil\u00ebsis\u00eb s\u00eb lart\u00eb mund t\u00eb kompensojn\u00eb munges\u00ebn e numrit t\u00eb parametrave. Zhvillimi i Gemma 2 2B gjithashtu thekson r\u00ebnd\u00ebsin\u00eb n\u00eb rritje t\u00eb metodave t\u00eb kompresimit dhe distilimit t\u00eb modeleve t\u00eb AI. Mund\u00ebsia p\u00ebr t\u00eb p\u00ebrmbledhur informacionin nga modelet m\u00eb t\u00eb m\u00ebdha n\u00eb ato m\u00eb t\u00eb vogla hap mund\u00ebsi p\u00ebr krijimin e mjeteve t\u00eb inteligjenc\u00ebs artificiale m\u00eb t\u00eb arritshme, pa sakrifikuar performanc\u00ebn e tyre.<\/p>\n

Google e ka trajnuar Gemma 2 2B n\u00eb nj\u00eb set t\u00eb madh t\u00eb dh\u00ebnash prej 2 trilion tokenesh, duke p\u00ebrdorur sisteme t\u00eb bazuara n\u00eb akselerator\u00ebt e saj TPU v5e. Mb\u00ebshtetja p\u00ebr disa gjuh\u00eb zgjeron mund\u00ebsit\u00eb e saj p\u00ebr p\u00ebrdorim n\u00eb aplikacione globale. Modeli Gemma 2 2B ka burim t\u00eb hapur. Hulumtuesit dhe zhvilluesit mund t\u00eb ken\u00eb qasje n\u00eb modelin p\u00ebrmes platform\u00ebs Hugging Face<\/a><\/noindex>. Ai gjithashtu mb\u00ebshtet korniza t\u00eb ndryshme, duke p\u00ebrfshir\u00eb PyTorch<\/a><\/noindex> dhe TensorFlow<\/a><\/noindex>.<\/p>\n

Burimi:<\/strong><\/p>\n