I-OpenAssistant, umthombo ovulekileyo we-AI bot

OpenAssistant

Vula uMncedisi yiprojekthi ejolise ekunikeni wonke umntu ukufikelela kwincoko enkulu esekwe kwimodeli yolwimi olukhulu.

Kutshanje LION uluntu (Isixa esikhulu soBukrelekrele obuVulile iNethiwekhi) yatyhilwa ngesibhengezo ukukhutshwa kokuqala kweprojekthi ye "OpenAssistant"., ephuhlisa i-chatbot yengqondo eyenziweyo ekwaziyo ukuqonda kunye nokuphendula imibuzo ngolwimi lwendalo, ukusebenzisana neenkqubo zomntu wesithathu kunye nokukhupha ulwazi oluyimfuneko.

Kulabo abangaqhelananga ne-LAION, kufuneka ukwazi ukuba iphuhlisa izixhobo, iimodeli kunye nokuqokelela idatha ukudala iinkqubo zokufunda zomatshini wamahhala (umzekelo, iqoqo le-LAION lisetyenziselwa ukuqeqesha imodeli ye-Stable Diffusion image synthesis system).

Ngaphandle kwe ikhowudi yokuqeqesha kunye nokulungelelanisa umsebenzi yebot kwikhompyuter yakho, kucetywa ukuba kusetyenziswe ingqokelela yeemodeli esele zenziwe ukusebenzisa esele iqeqeshiwe kunye nemodeli yolwimi, eqeqeshwe ngokusekelwe kwimizekelo engama-600 amawaka eengxoxo ngendlela yesicelo-impendulo (ukwenziwa komyalelo), elungiselelwe kwaye yahlaziywa ngokuthatha inxaxheba koluntu olunomdla.

Inkonzo ye-intanethi yokuvavanya umgangatho we-chatbot nayo yasungulwa, kusetyenziswa imodeli yolwazi ye-OA_SFT_Llama_30B_6, equka i-30 yeebhiliyoni zeeparamitha.

Iqela lethu lisebenze ngokungadinwayo kwezi nyanga zidlulileyo liqokelela ulwazi oluninzi kunye nengxelo esekwe kwiteksti ukwenza uluhlu lwedatha olwahluke kakhulu nolukhethekileyo olulungiselelwe ngokukodwa ukuqeqeshwa kolwimi okanye ezinye iinkqubo zeAI.

Ngaphezulu kwe-600 yamanqaku edatha eyenziwe ngumntu aquka uluhlu olubanzi lwezihloko kunye nezimbo zokubhala, isethi yethu yedatha iya kuba sisixhobo esixabisekileyo kuye nawuphi na umphuhlisi ojonge ukudala imodeli yokufundisa kwisizukulwana esilandelayo.

Ukwandisa ukusebenza kakuhle yenkqubo kwaye uphephe imfuneko gcina izixa ezikhulu zeeparamitha ezicwangcisiweyo, iprojekthi ibona kwangaphambili ukuba kunokwenzeka ukusebenzisa isiseko solwazi esihlaziyiweyo esinokuthi sifumane ulwazi olufunekayo ngeenjini zokukhangela okanye iinkonzo zangaphandle.

Ngokomzekelo, xa uvelisa iimpendulo, i-bot inokufikelela kwii-API zangaphandle ukuze ufumane idatha eyongezelelweyo. Kwimiba ephambili, inkxaso yobuntu ikwaqaqambile, oko kukuthi, ukukwazi ukuziqhelanisa nomsebenzisi othile ngokusekwe kumabinzana abo angaphambili.

Kwabo banomdla wokufaka i-OpenAssistant, kuya kufuneka uyazi ukuba ungayifaka endaweni, kwaye iimodeli zePythia SFT zomgqatswa ziyafumaneka kwiHuggingFace kwaye zinokulayishwa ngethala leencwadi leHuggingFace Transformers. Ngaloo ndlela, kunokwenzeka ukuba zingasetyenziswa kunye ne-hardware eyaneleyo. Kukho nezithuba kwi-HF ezinokusetyenziswa ukuncokola nomgqatswa we-OA ngaphandle kwehardware yakho. Nangona kunjalo, ezi modeli aziqinisekanga kwaye zinokuvelisa iziphumo ezibi okanye ezingafunwayo.

Iimodeli ze-LLaMa SFT azinakukhutshwa ngokuthe ngqo ngenxa yelayisensi ye-Meta, kodwa ii-weights ze-XOR ziya kukhutshwa kungekudala.

Kubalulekile ukukhankanya ukuba imodeli encinci yangoku (i-Pythia) ineeparamitha ze-12B kwaye kunzima ukuqhuba kwi-hardware yabathengi, kodwa inokuqhuba kwi-GPU eyodwa yobuchwephesha. Kusenokubakho imifuziselo emincinci kwixesha elizayo, kwaye sinethemba lokuqhubela phambili iindlela ezifana nenani elipheleleyo elinokunceda ukuqhuba imodeli kwihardware encinci.

Iprojekthi ayicwangcisi ukuyeka ukuphinda izakhono zeChatGPT. I-Open-Assistant ilindeleke ukuba ikhuthaze uphuhliso lophuhliso oluvulekileyo kwinkalo yokuveliswa komxholo kunye nokuphendula imibuzo ngeelwimi zendalo, kanye njengokuba iprojekthi yomthombo ovulekileyo we-Stable Diffusion ivuselela ukuphuhliswa kwezixhobo zokuvelisa umfanekiso.

Ikhowudi yeprojekthi ibhalwe kwiPython kwaye ihanjiswa phantsi kwelayisensi ye-Apache 2.0. Uphuhliso lwe-OpenAssistant lunokusetyenziswa ukwenza abancedisi bakho abakrelekrele kunye neenkqubo zengxoxo ezingabotshelelwanga kwii-APIs zangaphandle kunye neenkonzo. I-hardware yabathengi eqhelekileyo yanele ukusebenza, umzekelo, kunokwenzeka ukuba usebenze kwi-smartphone. Idatha yoMncedisi ovulekileyo ikhutshwe phantsi kwelayisensi ye-Creative Commons evumela ukuba kusetyenziswe uluhlu olubanzi, kuquka ukusetyenziswa kwezorhwebo.

Okokugqibela, ukuba unomdla wokwazi ukufunda ngakumbi malunga nayo kunye nokukwazi ukudibana nekhowudi yomthombo, unokujonga iinkcukacha. Kule khonkco ilandelayo.


Shiya uluvo lwakho

Idilesi yakho ye email aziyi kupapashwa. ezidingekayo ziphawulwe *

*

*

  1. Inoxanduva lwedatha: I-AB Internet Networks 2008 SL
  2. Injongo yedatha: Ulawulo lwe-SPAM, ulawulo lwezimvo.
  3. Umthetho: Imvume yakho
  4. Unxibelelwano lwedatha: Idatha ayizukuhanjiswa kubantu besithathu ngaphandle koxanduva lomthetho.
  5. Ukugcinwa kweenkcukacha
  6. Amalungelo: Ngalo naliphi na ixesha unganciphisa, uphinde uphinde ucime ulwazi lwakho.