I-StableLM, umthombo ovulekile ohlukile ku-ChatGPT

I-StableLM

I-StableLM futhi yakhelwe ukukhiqiza umbhalo nekhodi ngempumelelo

Kwaphuma izindaba ukuthi Ukuzinza kwe-AI, inkampani ngemuva kwemodeli ye-AI ye-Stable Diffusion imaging, isimemezele iqoqo layo lokuqala lezinhlobo zolimi ze-StableLM.

Ngayo Ukuzinza kwethemba ukuphindaphinda imiphumela yemodeli yokuhlanganiswa kwesithombe somthombo ovulekile Ukusabalalisa Okuzinzile, ekhishwe ngo-2022. Ngokulungiswa, i-StableLM ingase isetshenziselwe ukwakha umthombo ovulekile ohlukile ku-ChatGPT.

Kulabo abangazi kahle nge-Stability AI, kufanele wazi ukuthi lena yinkampani ezinze e-London ezibeka njengembangi yomthombo ovulekile we-OpenAI, inkampani ethuthukisa amamodeli ezilimi zokwenziwa ezinamandla kodwa eziphathelene nokuphathelene njenge-ChatGPT.

Mayelana ne-StableLM

I-StableLM igama lomndeni wamamodeli olimi okwenziwa yi-Stability AI, ezitholakala njengomthombo ovulekile ku-GitHub ngaphansi kwelayisensi ye-Creative Commons BY-SA-4.0I-StableLM imodeli yokukhiqiza umbhalo engaqamba umbhalo womuntu futhi ibhale izinhlelo ngokubikezela igama elilandelayo ngokulandelana. Isebenzisa indlela ebizwa ngokuthi “chip prediction” okubandakanya ukuqagela ucezu lwegama elilandelayo kumongo onikezwe ngumuntu ngendlela “yesu”.

Njengamanye ama-LLM "amancane". I-StableLM ithi ithola ukusebenza okufanayo kumodeli yesithenjwa ye-GPT-3 ye-OpenAI kuyilapho kusetshenziswa amapharamitha avamile ambalwa kakhulu (amabhiliyoni angu-7 e-StableLM vs. 175 billion we-GPT-3).

Ukukhishwa kwe-StableLM kwakhela phezu kokuhlangenwe nakho kwethu ngamamodeli olimi omthombo ovulekile wangaphambilini nge-EleutherAI, isikhungo socwaningo esingenzi nzuzo. Lawa mamodeli olimi ahlanganisa i-GPT-J, GPT-NeoX, kanye ne-Pythia suite, eziqeqeshwe kudathasethi yomthombo ovulekile ethi The Pile.

I-StableLM ithi inokusebenza okufanayo ne-GPT-3, imodeli yolimi enika amandla i-ChatGPT, kuyilapho isebenzisa amapharamitha ambalwa kakhulu (amabhiliyoni ayi-7 vs. 175 billion). Amapharamitha ayizinto eziguquguqukayo imodeli ezisebenzisayo ukuze ifunde kudatha yokuqeqeshwa. Ukuba namapharamitha ambalwa kwenza imodeli ibe yincane futhi isebenze kahle, okungenza kube lula ukusebenza kumadivayisi asendaweni njengama-smartphone namakhompyutha aphathekayo.

I-StableLM uqeqeshwe kudathasethi entsha esekelwe ku-The Pile, equkethe 1,5 trillion amathokheni, okuyinto cishe izikhathi 3 ubukhulu Inqwaba. I-Pile iyidathasethi yekhwalithi ephezulu nehlukile yokuqeqesha amamodeli olimi.

Ukuzinza kwe-AI isho ukuthi izifanekiso sezivele zitholakala endaweni yokugcina ye-GitHub nokuthi iphepha elimhlophe eligcwele liyeza maduze, futhi libheke phambili ekuqhubekeni nokusebenzisana nabathuthukisi nabacwaningi njengoba likhipha i-StableLM suite.

Ngaphezu kwalokho, basho ukwethula uhlelo lokusebenzisana oluvulekile lwe-RLHF nokusebenza nemizamo yomphakathi efana ne-Open Assistant ukuze bakhe isethi yedatha yomthombo ovulekile yabasizi be-AI.

Okokugcina, Uma sikhuluma ngokukhishwa kwe-Stability AI, singagqamisa nokuthi imemezele ukukhishwa kwe-beta kwe I-SDXL (okumelela ukuthi I-Stable Diffusion Extra Large), imodeli entsha yobuhlakani bokwenziwa ekwazi ukukhiqiza izithombe ngezincazelo zombhalo. I-SDXL iyisengezo sakamuva ku-Stable Diffusion suite, efaka phakathi amamodeli e-SD, SDT, kanye ne-SDC.

I-SDXL ihlukile kwamanye amamodeli ngosayizi namandla ayo. Ngamapharamitha ayizigidi eziyizinkulungwane ezingu-2300, i-SDXL inkulu izikhathi ezingaphezu kuka-2,5 kunemodeli yasekuqaleni ye-SD, eyayinezigidi ezingu-890 kuphela. Le mingcele eyengeziwe ivumela i-SDXL ukuthi ikhiqize izithombe ezinamathela kangcono kumaphethini ayinkimbinkimbi. Isibonelo, i-SDXL ingakhiqiza umbhalo ofundekayo ezithombeni noma idale izithombe ezingokoqobo ezimangalisayo zezinhlamvu eziqanjiwe.

I-SDXL okwamanje iku-beta ku-DreamStudio nezinye izinhlelo zokusebenza zokuthwebula ezidumile ezifana ne-NightCafe Creator. Njengawo wonke amamodeli we-Stability AI, i-SDXL izokhishwa maduze njengomthombo ovulekile wokufinyeleleka okuphelele. I-Stability AI imemezela ukuthi i-SDXL inikezwe ilayisense ngokuvumelekile ukusetshenziselwa ezohwebo nokungezona ezohwebo, inqobo nje uma ulandela imihlahlandlela yokuziphatha nezomthetho.

Ekugcineni, uma unentshisekelo yokwazi kabanzi ngakho, ungaxhumana nemininingwane Kulesi sixhumanisi esilandelayo.