I-Meta yethula imodeli ye-LLaMA, ithuluzi lokusesha elinamandla kune-OpenAI's GPT-3

I-Meta isanda kukhipha i-generator entsha yolimi lwe-AI ebizwa ngokuthi i-LLaMA, eqinisekisa indima yenkampani esungula izinto ezintsha.

"Namuhla sikhipha imodeli yolimi ye-AI entsha, esezingeni eliphezulu ebizwa ngokuthi i-LLaMA eklanyelwe ukusiza abacwaningi bathuthukise umsebenzi wabo," kusho i-CEO uMark Zuckerberg eposini lika-Facebook.

Kungani LLaMA

Amamodeli ezilimi ezinkulu athathe umhlaba wezobuchwepheshe ngokuzumayo. Banika amandla amathuluzi obuhlakani bokwenziwa, njenge I-ChatGPT namanye amamodeli ezingxoxo. Kodwa-ke, ukusebenzisa lawa mathuluzi kuza nengcuphe enkulu, izimangalo ezibambekayo kodwa ezingamanga, ezikhiqiza okuqukethwe okunobuthi, nokuchema okulingisa okusekelwe kudatha yokuqeqeshwa kwe-AI. 

Ukusiza abacwaningi ukuxazulula lezi zinkinga, ngoLwesihlanu, February 25, Meta  umemezele ukukhululwa yohlobo olusha olukhulu lolimi olubizwa ngokuthi I-LLAMA (Large Language Model I-Meta AI) . 

Yini i-LLaMA?

I-LLaMA akuyona a i-chatbot, kodwa iyithuluzi lokusesha, ngokusho kwe-Meta ai, elizoxazulula izinkinga ezihlobene namamodeli olimi AI. "Amamodeli amancane, asebenza kangcono njenge-LLaMA avumela abanye emphakathini wabacwaningi abangenawo ukufinyelela inani elikhulu lengqalasizinda ukuze bafunde lezi zibonelo, baqhubekisele phambili ukufinyeleleka kwentando yeningi kule nsimu ebalulekile nethuthuka ngokushesha," kusho uMeta kubhulogi yakhe. isikhulu .

I-LLaMA iqoqo lamamodeli olimi asukela ku-7B kuya kumapharamitha angama-65B. Le nkampani ithe iqeqesha amamodeli ayo ngezigidigidi zamathokheni, ithi ingakwazi ukuqeqesha amamodeli asezingeni eliphezulu isebenzisa ama-dataset omphakathi futhi inganciki kumadathasethi obunikazi, angafinyeleleki.

I-LLaMA ihlukile

Ngokusho kwe-Meta, ukuqeqeshwa okuyimodeli efana ne-LLaMA kudinga amandla amancane kakhulu ekhompiyutha ukuhlola, ukuqinisekisa, nokuhlola amacala amasha okusebenzisa. Amamodeli olimi ayisisekelo aqeqeshelwa amabhulokhi amakhulu edatha engenamalebula, okuwenza alungele ukwenziwa ngokwezifiso imisebenzi ehlukahlukene. 

Ephepheni lakhe lokucwaninga, u-Meta waphawula ukuthi i-LLaMA-13B iphumelele i-OpenAI's GPT-3 (175B) kumabhentshimakhi amaningi futhi i-LLaMA-65B incintisana namamodeli aphezulu, Chinchilla70B by DeepMind e I-PaLM-540B evela kwa-Google. 

I-LLaMA okwamanje ayisebenzi kunoma yimiphi imikhiqizo ye-Meta ai, nokho, inkampani inezinhlelo zokuyenza itholakale kubacwaningi. Ngaphambilini le nkampani yethule i-LLM OPT-175B yayo, kodwa i-LLaMA iwuhlelo lwayo oluthuthuke kakhulu. 

Inkampani ikwenza kutholakale ngaphansi kwelayisensi engeyona eyentengiso egxile emacaleni okusetshenziswa kocwaningo. Izotholakala kubacwaningi bezemfundo; labo abaxhumene nohulumeni, izinhlangano zomphakathi kanye nezinhlangano zemfundo; kanye namalabhorethri ocwaningo lwezimboni emhlabeni jikelele.

Ercole Palmeri