Isoftware inokunceda abaphuhlisi bakhokele izicelo ze-AI ezivelisayo ukwenza iimpendulo zeteksti ezichukumisayo ezihlala zisendleleni.
Kutshanje kutyhilwe ukuba UNvidia ukhuphe isoftware entshakubizwa NeMo Guardrails, ukunceda abaphuhlisi kunye neenkampani zikhokele kwaye zilawule iimpendulo zemveliso yee-AIs zabo.
NeMo Guardrails ijolise ekwenzeni i-AI chatbots kunye nezinye izicelo Yenziwe kwiiModeli zoLwimi olude (LLM) zichanekile, zifanelekile, zifanelekile kwaye zikhuselekile. Ngamanye amagama, iya kunceda abasebenzisi ukuthintela iimodeli ze-AI ekuboniseni iinyani ezingachanekanga, phuma ngaphandle kwesihloko, thetha ngezihloko eziyingozi okanye uvule imingxuma yokhuseleko. Isibhengezo sicebisa ukuba i-NeMo Guardrails inokunceda ukusombulula ingxaki ye-AI ye-chatbot hallucinations, kodwa abaninzi bayathandabuza ngayo.
Nangona i-hype, iimodeli ezinkulu ze-AI ezivelisa isicatshulwa, ezifana ne-OpenAI's GPT-4, zenza iimpazamo ezininzi, ezinye zazo ziyingozi. Kwaye njengoko sisazi, ii-AIs ziqeqeshwe ukusuka kwi-terabytes yedatha ukwenza iinkqubo ezikwaziyo ukudala iibhloko zetekisi ezifundwayo ngokungathi zibhalwe ngumntu.
Kodwa bavame ukwenza izinto, apho oogqirha bahlala bebiza ngokuthi "i-hallucination." Ukusetyenziswa kwangethuba kobuchwephesha, okufana nokushwankathela amaxwebhu okanye ukuphendula imibuzo esisiseko, kufuneka kuncitshiswe "imibono" ukuze ibe luncedo.
Abadali bale mifuziselo ye-AI bathi bathatha inyathelo ukujongana nemiba, efana nokuseka izihluzi kunye namaqela abantu abamodareyitha ukujongana nemiba xa ixelwa. Kodwa asikho isisombululo esinye.
Nezona modeli zigqwesileyo zanamhlanje zisesichengeni sokucalucalula, ubuthi, kunye nohlaselo olunolunya. Kwiphulo layo lee-LLM “ezikhuselekileyo”, uNvidia uphehlelele iNeMo Guardrails, isethi yezixhobo ezivulelekileyo ezijoliswe ekwenzeni izicelo ezisekelwe kwi-AI zichaneke ngakumbi, zifanelekile, zifanelekile kwaye zikhuselekile. UNvidia uthe isixhobo sakhe sisiphumo seminyaka yophando.
Ngokutsho kukaNvidia, NeMo Guardrails ngumaleko wesoftware oza phakathi komsebenzisi kunye neLLM okanye ezinye izixhobo ze-AI ukususela ngoko kuphelisa iziphumo ezimbi phambi kokuba imodeli ivelise kwaye ithintele imodeli ekuqhubeni izibonakaliso ezingachanekanga.
Kwingxelo, uJonathan Cohen, usekela mongameli wophando olusetyenzisiweyo eNvidia, wachaza:
Ukuba unenkonzo yabathengi ye-chatbot, eyenzelwe ukuthetha ngeemveliso zakho, awufuni ukuba iphendule imibuzo malunga nabo sikhuphisana nabo. Ufuna ukubeka iliso kwincoko. Yaye ukuba oko kuyenzeka, incoko yalathisela kwimibandela oyithandayo.”
I-NeMo Guardrails iquka ikhowudi, iisampuli, kunye namaxwebhu okongeza ukhuseleko kwizicelo ze-AI ezivelisa isicatshulwa kunye nentetho. U-Nvidia uthi i-toolkit yenzelwe ukusebenza kunye nee-LLM ezininzi, evumela abaphuhlisi ukuba benze imithetho ngokusebenzisa imigca embalwa yekhowudi.
“Ekugqibeleni, abaphuhlisi balawula oko kungaphandle kwemida yesicelo sabo ngeNeMo Guardrails. Banokuphuhlisa iintsimbi ezibanzi kakhulu okanye ezimxinwa kakhulu ukuba bangazisebenzisa, ”utshilo uCohen.
NeMo Guardrails ivumela abaphuhlisi ukuba bachaze iindidi ezintathu zemida:
- Amanyathelo okhuseleko Imixholo inqanda usetyenziso ekubhaduleni ukuya kwimimandla engafunwayo. Ngokomzekelo, bathintela abancedisi benkonzo yabathengi ukuba baphendule imibuzo malunga nemozulu.
- Amanyathelo okhuseleko, Baqinisekisa ukuba izicelo ziphendula ngolwazi oluchanekileyo nolufanelekileyo. Banokucoca ulwimi olungafunwayo kwaye baqinisekise ukuba iimbekiselo zenziwe kuphela kwimithombo ethembekileyo;
- Izicelo zithintelwe ekusekweni koqhakamshelwano nezicelo zangaphandle zomntu wesithathu ezaziwa ngokuba zikhuselekile.
Ngokutsho kukaNvidia, i-NeMo Guardrails inokusebenzisa i-LLM enye ukubona i-hallucinations ngokucela enye i-LLM ukuba iqinisekise impendulo yokuqala ye-LLM.
Emva koko ibuyisela "Andazi" ukuba imodeli ayifumani naziphi na iimpendulo ezihambelanayo. Nangona le ndlela ibonakala iluncedo, abahlalutyi abaninzi baye bavakalisa ukuthandabuza ngayo. Ngaloo ndlela, uNvidia uye waqaphela ukuba i-NeMo Guardrails ayisona isisombululo esisodwa kwiintsilelo zemodeli yolwimi. Isigebenga saseSanta Clara, saseCalifornia sathi isixhobo saso esitsha asinaziphene. Ngamanye amagama, i-NeMo Guardrails ayizukugubungela yonke into.
Cohen nayen iphawula ukuba i-NeMo Guardrails isebenza ngcono kakhulu ngeemodeli ezilungileyo ngokwaneleyo ukulandela imiyalelo, efana ne-ChatGPT, kwaye esebenzisa isakhelo esidumileyo seLangChain ukwenza usetyenziso olusekwe kwi-AI.
Oku kukhupha ezinye zeenketho ezikhoyo zemithombo evulekileyo. Ukongeza kwimibuzo malunga nokusebenza kwesixhobo, kufuneka kuqatshelwe ukuba INvidia ayinyanzelekanga ukuba ikhulule i-NeMo Guardrails ngaphandle kokulunga okumsulwa. Iyinxalenye yesakhelo seNeMo yenkampani, efumaneka ngesoftware ye-AI yeshishini leNvidia kunye nenkonzo yayo yelifu yeNeMo elawulwa ngokupheleleyo.
Okokugqibela, ukuba unomdla wokwazi ngakumbi ngayo, ungajongana neenkcukacha kwi ukulandela ikhonkco.