I-Meta isungula imodeli ye-LLaMA, isixhobo sokukhangela esinamandla ngakumbi kune-OpenAI's GPT-3

I-Meta isandula ukukhupha ijenereyitha yolwimi lwe-AI entsha ebizwa ngokuba yi-LLaMA, eqinisekisa indima yenkampani eyenza izinto ezintsha.

"Namhlanje sikhupha imodeli entsha yolwimi lwe-AI ebizwa ngokuba yi-LLaMA eyenzelwe ukunceda abaphandi ukuba baqhubele phambili umsebenzi wabo," utshilo u-CEO uMark Zuckerberg kwiposti ye-Facebook.

Kutheni LLaMA

Iimodeli zolwimi olukhulu zithathe ilizwe letekhnoloji ngoqhwithela. Banika amandla izixhobo zobuntlola ezenziweyo, ezifana NcokolaGPT kunye neminye imizekelo yokuncokola. Nangona kunjalo, ukusebenzisa ezi zixhobo kuza nomngcipheko omkhulu, amabango abambekayo kodwa abubuxoki, avelisa umxholo onetyhefu, kunye nokulinganisa umkhethe osekelwe kwidatha yoqeqesho ye-AI. 

Ukunceda abaphandi bacombulule ezi ngxaki, ngoLwesihlanu, ngoFebruwari 25, iMeta  ibhengeze ukukhululwa yemodeli entsha yolwimi olukhulu ebizwa LLaMA (Large Language Model Meta AI) . 

Yintoni iLLAMA?

I-LLaMA ayiyo ingxoxo, kodwa sisixhobo sokukhangela ukuba, ngokutsho kweMeta ai, iya kusombulula iingxaki ezinxulumene nemizekelo yolwimi AI. "Iimodeli ezincinci, ezisebenza ngcono ezifana ne-LLaMA zivumela abanye kuluntu lophando abangenalo ukufikelela kwizixa ezikhulu zeziseko zophuhliso ukuba bafunde le mizekelo, ukufikelela kwidemokhrasi ngakumbi kule ntsimi ibalulekileyo kwaye ikhula ngokukhawuleza," utshilo uMeta kwibhlog yakhe. igosa .

I-LLaMA yingqokelela yeemodeli zolwimi ukusuka kwi-7B ukuya kwi-65B iiparamitha. Inkampani ithe iqeqesha iimodeli zayo kwiibhiliyoni zeethokheni, isithi inokuqeqesha iimodeli ezibukhali zisebenzisa iiseti zedatha yoluntu kwaye ingathembeli kubunikazi, iiseti zedatha ezingafikelelekiyo.

I-LLaMA yahlukile

Ngokutsho kwe-Meta, uqeqesho lwemodeli efana ne-LLaMA lufuna amandla amancinci kakhulu okuvavanya, ukuqinisekiswa, kunye nokuhlola iimeko ezintsha zokusetyenziswa. Imifuziselo yolwimi olusisiseko iqeqeshelwa kwiibhloko ezinkulu zedatha engabhalwanga, izenza zilungele ukwenziwa kwimisebenzi eyahlukeneyo. 

Kwiphepha lakhe lophando, i-Meta yaqaphela ukuba i-LLaMA-13B iphumelele i-OpenAI's GPT-3 (175B) kwiibenchmarks ezininzi kwaye i-LLaMA-65B inokhuphiswano kunye neemodeli eziphezulu, Chinchilla70B yi-DeepMind e I-PaLM-540B evela kuGoogle. 

I-LLaMA ayisebenzi ngoku kuyo nayiphi na imveliso ye-Meta ai, nangona kunjalo, inkampani inezicwangciso zokuyenza ifumaneke kubaphandi. Inkampani ibikhe yaphehlelela i-LLM OPT-175B yayo ngaphambili, kodwa i-LLaMA yeyona nkqubo iphambili. 

Inkampani iyenza ifumaneke phantsi kwelayisenisi engeyiyo eyentengiso egxile kumatyala okusetyenziswa kophando. Iya kufumaneka kubaphandi bezemfundo; abo banxulumene norhulumente, imibutho yoluntu kunye nemibutho yezemfundo; kunye neelebhu zophando lwamashishini kwihlabathi jikelele.

Ercole Palmeri