Uyiqeqesha kanjani i-chatbot yakho ngobunjiniyela obusheshayo

Esinye sezizathu ezenza ukuthi ama-chatbot asekelwe kubuhlakani bokwenziwa ahlasele umhlaba ezinyangeni ezisanda kwedlula ingoba angakwazi ukukhiqiza noma acwebezele umbhalo ngezinjongo ezahlukahlukene, kungaba ukudala umkhankaso wesikhangiso noma ukubhala kabusha.

Lawa ma-chatbots anikwa amandla ama-algorithms emodeli yolimi olukhulu (LLM), angalingisa ubuhlakani bomuntu futhi adale okuqukethwe kombhalo kanye nomsindo, ividiyo, izithombe, nekhodi yekhompuyutha. Ama-LLM awuhlobo lobuhlakani bokwenziwa obuqeqeshwe ngenqwaba yama-athikili, izincwadi, noma izinsiza ezisekelwe ku-inthanethi nokunye okokufaka ukuze kukhiqizwe izimpendulo ezifana nezomuntu kokokufaka kolimi lwemvelo.

Inani elikhulayo lezinkampani zobuchwepheshe zembule amathuluzi e-AI akhiqizayo asuselwa kuma-LLM ukuze asetshenziswe ibhizinisi ukwenza imisebenzi yohlelo lokusebenza ngokuzenzakalelayo. Ngokwesibonelo, I-Microsoft ngesonto eledlule ikhishwe inombolo elinganiselwe yabasebenzisi i-chatbot esekelwe ku-OpenAI’s ChatGPT; ishumekwe ku-Microsoft 365 futhi ingenza ngokuzenzakalelayo imisebenzi ye-CRM ne-ERP.

I-Salesforce

Ngokwesibonelo, entsha I-Microsoft 365 Copilot ingasetshenziswa ku-Word ukwakha uhlaka lokuqala lwedokhumenti, okungenzeka konga amahora okubhala, ukuthola, nokuhlela. I-Salesforce futhi yamemezela izinhlelo zokukhulula i-chatbot esekelwe ku-GPT ukuze isetshenziswe nengxenyekazi yayo ye-CRM.

Iningi lama-LLM, njenge I-OpenAI’s GPT-4ziqeqeshwe kusengaphambili njenge igama elilandelayo noma izinjini zokubikezela okuqukethwe – yindlela amabhizinisi amaningi azisebenzisa ngayo, “ngaphandle kwebhokisi,” njengokungathi. Futhi nakuba ama-chatbot asekelwe ku-LLM ekhiqize ingxenye yawo yamaphutha, ama-LLM aqeqeshwe kusengaphambili asebenza kahle kakhulu ekondleni okuqukethwe okunembe kakhulu nokuphoqelelayo, okungenani, okungasetshenziswa njengendawo yokweqa.

Izimboni eziningi, nokho, zidinga ama-algorithms e-LLM enziwe ngezifiso, lawo aqonda i-jargon yawo futhi akhiqize okuqukethwe okuqondene nabasebenzisi bazo. Ama-LLM emboni yezokunakekelwa kwempilo, isibonelo, angase adinge ukucubungula futhi ahumushe amarekhodi ezempilo kagesi (EHRs), aphakamise ukwelashwa, noma enze isifinyezo sokunakekelwa kwezempilo kwesiguli esisekelwe kumanothi kadokotela noma ukurekhodwa kwezwi. I-LLM eshuthelwe embonini yezinsizakalo zezezimali ingafingqa izingcingo zeholo, idale imibhalo yemihlangano, futhi yenze ukuhlaziya ukukhwabanisa ukuze kuvikelwe abathengi.

Kuzo zonke izimboni ezihlukahlukene, ukuqinisekisa izinga eliphezulu lokunemba kwempendulo kungaba okubalulekile.

Ama-LLM amaningi angafinyelelwa ngesixhumi esibonakalayo sohlelo lokusebenza (i-API) esivumela umsebenzisi ukuthi akhe imingcele noma ukulungisa indlela i-LLM ephendula ngayo. Umbuzo noma isicelo esithunyelwe ku-chatbot siyabizwa ukwaziswa, ngokuthi umsebenzisi ucela impendulo. Ukwaziswa kungaba imibuzo yolimi lwemvelo, amazwibela ekhodi, noma imiyalo, kodwa ukuze i-LMM yenze umsebenzi wayo ngokunembile, ukwaziswa kufanele kube iphuzu.

Futhi leso sidingo sinikeze ikhono elisha: ubunjiniyela obusheshayo.

Unjiniyela osheshayo wachaza

Ubunjiniyela obusheshayo yinqubo yokubhala kanye nokwenza kahle ukwaziswa kombhalo kumamodeli amakhulu olimi ukuze kuzuzwe imiphumela efiselekayo. “[It] isiza ama-LLM ukuthi aphindaphindeke ngokushesha ekufanekiseni umkhiqizo nasekuhloleni, njengoba ihlobanisa i-LLM ukuthi ihambisane kangcono nencazelo yomsebenzi ngokushesha futhi kalula,” kusho uMarshall Choy, iphini likamongameli omkhulu wemikhiqizo kwa-. SambaNova Systemsisiqalo seSilicon Valley esenza ama-semiconductors obuhlakani bokwenziwa (AI).

Mhlawumbe njengokubalulekile kubasebenzisi, ubunjiniyela obusheshayo sebulungele ukuba yikhono elibalulekile le-IT kanye nezingcweti zebhizinisi, ngokusho kuka-Eno Reyes, unjiniyela wokufunda ngomshini oneHugging Face, inkundla eqhutshwa umphakathi edala futhi isingathe ama-LLM.

UNeil Lockhart/Shutterstock

“Abantu abaningi engibaziyo ku-software, IT, kanye nokubonisana basebenzisa ubunjiniyela obusheshayo ngaso sonke isikhathi emsebenzini wabo siqu,” kusho uReyes ephendula nge-imeyili. I-Computerworld. “Njengoba ama-LLM ehlanganiswa nezimboni ezihlukahlukene, amandla awo okuthuthukisa umkhiqizo makhulu kakhulu.”

Ngokusebenzisa ngempumelelo ubunjiniyela obusheshayo, abasebenzisi bebhizinisi bangakwazi ukuthuthukisa ama-LLM ukuze benze imisebenzi yabo ethile ngendlela ephumelelayo nenembile, kusukela ekusekelweni kwamakhasimende kuya ekukhiqizeni okuqukethwe nokuhlaziywa kwedatha, kusho uReyes.

I-LLM eyaziwa kakhulu njengamanje – i-OpenAI’s GPT-3 – iyisisekelo sengxoxo edume kakhulu ye-ChatGPT. I-GPT-3 LLM isebenza kumodeli yepharamitha yebhiliyoni engu-175 engakwazi ukukhiqiza umbhalo kanye nekhodi yekhompuyutha ngemiyalo emifushane ebhaliwe. Inguqulo yakamuva ye-OpenAI, i-GPT-4, ilinganiselwa ukuthi inepharamitha efika ku-280 bhiliyoni, okwenza kube maningi amathuba okuthi ikhiqize izimpendulo ezinembile.

Kanye ne-OpenAI’s GPT LLM, amapulatifomu e-AI akhiqizayo adumile afaka amamodeli avulekile afana nala UBUSO OBUGQUMAYO futhi I-XLM-RoBERTa, I-Nvidia NeMO LLM, XLNet, Co: lapha futhi I-GLM-130B.

Ngenxa yokuthi ubunjiniyela obusheshayo buyisiyalo esisafufusa futhi esisafufusa, amabhizinisi athembele kumabhukwana kanye nemihlahlandlela esheshayo njengendlela yokuqinisekisa izimpendulo ezifanele ezivela kwizicelo zabo ze-AI. Kukhona nezimakethe ezisafufusa, njenge 100 ukwaziswa okungcono kakhulu kwe-ChatGPT.

“Abantu baze bathengise iziphakamiso ngokushesha,” kusho u-Arun Chandrasekaran, umhlaziyi wephini likamongameli ohlonishwayo e-Gartner Research, wengeza ngokuthi ukunakwa kwakamuva kwe-AI ekhiqizayo kuveze isidingo sobunjiniyela obusheshayo.

“Kuyisizinda esisha kakhulu,” esho. “Izicelo ze-Generative AI zivame ukuthembela kumamodeli amakhulu we-AI azigadayo ngakho ukuthola izimpendulo ezilungile kuzo kudinga ukwazi okwengeziwe, izivivinyo kanye nomzamo owengeziwe. Ngiqinisekile ngokukhula okukhulayo singabona isiqondiso esingcono kanye nemikhuba engcono kakhulu evela kubadali bemodeli ye-AI ngezindlela ezisebenzayo zokuthola okungcono kakhulu kumamodeli we-AI nezinhlelo zokusebenza.”

Okokufaka okuhle kufana nokukhiphayo okuhle

Ingxenye yokufunda ngomshini ye-LLMs ifunda ngokuzenzakalelayo ekufakweni kwedatha. Ngaphezu kwedatha eyasetshenziswa ekuqaleni i-LLM, njenge-GPT-4, i-OpenAI idale into ebizwa ngokuthi I-Reinforcement Learning Feedback yabantulapho umuntu eqeqesha imodeli yendlela yokunikeza izimpendulo ezinjengomuntu.

Isibonelo, umsebenzisi uzofaka umbuzo ku-LLM bese ebhala impendulo efanelekile. Ngemuva kwalokho umsebenzisi uzophinda abuze imodeli umbuzo ofanayo, futhi imodeli izonikeza ezinye izimpendulo eziningi ezahlukahlukene. Uma kungumbuzo osekelwe eqinisweni, ithemba liwukuthi impendulo izohlala injalo; uma kungumbuzo ovulekile, umgomo uwukukhiqiza izimpendulo zokudala eziningi ezifana nezomuntu.

Isibonelo, uma umsebenzisi ecela i-ChatGPT ukuthi yenze inkondlo ngomuntu ohlezi ogwini lolwandle e-Hawaii, okulindelekile ukuthi izokhiqiza inkondlo ehlukile isikhathi ngasinye. “Ngakho-ke, okwenziwa abaqeqeshi abangabantu ukulinganisa izimpendulo ukusuka kokuhle kakhulu kuye kokubi kakhulu,” kusho uChandrasekaran. “Lokho kuwumbono wemodeli yokuqinisekisa ukuthi inikeza impendulo efana neyomuntu noma engcono kakhulu, kuyilapho izama ukunciphisa izimpendulo ezimbi kakhulu. Kepha uyiphendula kanjani imibuzo [has] umthelela omkhulu ekuphumeni okuthola kumamodeli. “

Izinhlangano zingaqeqesha imodeli ye-GPT ngokungenisa amasethi edatha yangokwezifiso angaphakathi kuleyo nkampani. Isibonelo, bangathatha idatha yebhizinisi bayilebule futhi bayichasise ukuze bakhulise ikhwalithi yayo bese beyifaka kumodeli ye-GPT-4. Lokho kushuna kahle imodeli ukuze ikwazi ukuphendula imibuzo eqondene naleyo nhlangano.

Fine Tuning cna futhi imboni ethize. Sekuvele kunemboni ye-cottage ephumayo yokuqala ethatha i-GPT-4 futhi ingenisa ulwazi oluningi oluqondene nezimboni eziqondile, ezifana nezinsizakalo zezezimali.

“Bangase badle ulwazi lwe-Lexus-Nexus ne-Bloomberg, bangase badle ulwazi lwe-SEC njengemibiko ye-8K ne-10K. Kodwa iphuzu ukuthi imodeli ifunda ulimi oluningi noma ulwazi oluqondene ngqo naleso sizinda,” kusho uChandrasekaran. “Ngakho-ke, ukuhlela okuhle kungenzeka ezingeni lemboni noma lenhlangano.”

Ngokwesibonelo, Harvey iyisiqalo esisebenzisana ne-OpenAI ukudala lokho ekubiza ngokuthi “ikhophi yabameli” noma inguqulo ye-ChatGPT yochwepheshe bezomthetho. Abameli bangasebenzisa i-chatbot ye-ChatGPT eyenziwe ngokwezifiso ukuze bathole noma ikuphi okusemthethweni ukuze amajaji athile alungiselele icala lawo elilandelayo, kusho u-Chandrasekaran.

“Ngibona ukubaluleka kokuthengisa ukwaziswa hhayi kakhulu ngolimi kodwa izithombe,” kusho Chandrasekaran. “Zikhona zonke izinhlobo zamamodeli esikhaleni se-AI esikhiqizayo, okubandakanya amamodeli wombhalo kuya kwesithombe.”

Isibonelo, umsebenzisi angacela imodeli ekhiqizayo ye-AI ukuze akhiqize isithombe somdlali wesiginci egida enyangeni. “Ngicabanga ukuthi isizinda sombhalo kuya esithombeni sinokugcizelelwa kakhulu ezimakethe ezisheshayo,” kusho u-Chandrasekaran.

Ubuso Obugonayo njengehabhu ye-LLM yesitobhi esisodwa

Nakuba i-Hugging Face idala amanye ama-LLM ayo, okuhlanganisa i-BLOOM, indima eyinhloko yenhlangano iwukuba ihabhu lamamodeli okufunda emishini evela eceleni, njengoba i-GitHub yenza ngekhodi; I-Hugging Face okwamanje isingathe amamodeli okufunda ngomshini angaphezu kuka-100,000, okuhlanganisa nama-LLM ahlukahlukene asuka ekuqaleni kanye nobuchwepheshe obukhulu.

Njengoba amamodeli amasha anemithombo evulekile, ngokuvamile enziwa atholakale kuhabhu, okwenza indawo eyodwa yokumisa eyodwa yama-LLM avelayo avuleke.

Ukushuna kahle i-LLM yebhizinisi elithile noma imboni esebenzisa i-Hugging Face, abasebenzisi bangasebenzisa “inhlangano “Ama-Transformers” Ama-API kanye nemitapo “yedathasethi”. Isibonelo, ezinsizeni zezezimali, umsebenzisi angangenisa i-LLM eqeqeshwe ngaphambilini njenge I-Flan-UL2, layisha idathasethi yama-athikili ezindaba zezimali, futhi usebenzise umqeqeshi “we-transformers” ukushuna kahle imodeli ukuze ukhiqize izifinyezo zalezo zihloko. Ukuhlanganiswa ne AWS, I-DeepSpeedfuthi Sheshisa thuthukisa futhi uthuthukise ukuqeqeshwa.

Yonke inqubo ingenziwa ngemigqa yekhodi engaphansi kwe-100, ngokusho kukaReyes.

Enye indlela yokuqalisa ngobunjiniyela obusheshayo ifaka i-Hugging Face’s Inference API; iwukuphela kwesicelo se-HTTP esilula esisekela amamodeli we-transformer angaphezu kuka-80,000, ngokusho kukaReyes. “Le API ivumela abasebenzisi ukuthi bathumele iziyalezo zombhalo futhi bathole izimpendulo kumamodeli wemithombo evulekile endaweni yethu, okuhlanganisa nama-LLM,” kusho uReyes. “Uma ufuna ukwenza lula nakakhulu, ungakwazi ukuthumela umbhalo ngaphandle kwekhodi ngokusebenzisa iwijethi ye-inference kumamodeli we-LLM ku- Ihabhu yobuso obungangana.”

Ukufunda okumbalwa futhi okungasho lutho

Ubunjiniyela bokushesha be-LLM ngokuvamile buthatha uhlobo olulodwa kwezimbili: ukufunda noma ukuqeqeshwa okungasho lutho futhi ungadumi.

Ukufunda ungasho lutho kubandakanya ukuphakela imiyalelo elula njengomyalezo oveza impendulo elindelekile evela ku-LLM. Iklanyelwe ukufundisa i-LLM ukwenza imisebenzi emisha ngaphandle kokusebenzisa idatha enelebula kuleyo misebenzi ethile. Cabanga nge-zero-shot njengokufunda kokuqinisa.

Ngokuphambene, ukufunda okumbalwa kusebenzisa inani elincane lolwazi lwesampula noma idatha ukuqeqesha i-LLM ngezimpendulo ezifiswayo. Ukufunda okumbalwa iqukethe izingxenye ezintathu ezibalulekile:

• Incazelo Yomsebenzi: Incazelo emfushane yalokho okumele kwenziwe yimodeli, isb “Humusha isiNgisi kuya kusiFulentshi”
• Izibonelo: Izibonelo ezimbalwa ezibonisa imodeli lokho okulindeleke ukuba ikwenze, isibonelo, “sea otter => loutre de mer”
• Ngokushesha: Ukuqala kwesibonelo esisha, okumele imodeli iqedele ngokukhiqiza umbhalo ongekho, njengokuthi “shizi => “

Eqinisweni, zimbalwa izinhlangano namuhla ezinamamodeli okuqeqesha ngokwezifiso azohambisana nezidingo zazo ngoba amamodeli amaningi asesesigabeni sokuqala sokuthuthuka, ngokusho kwe-Gartner’s Chandrasekaran. Futhi nakuba ukufunda okumbalwa kanye nokushutha kancane kungasiza, ukufunda ubunjiniyela bamashesha njengekhono kubalulekile, kokubili kubasebenzisi be-IT nabamabhizinisi ngokufanayo.

“Ubunjiniyela obusheshayo buyikhono elibalulekile ongalithola namuhla njengoba amamodeli ezisekelo esebenza kahle ekudubuleni kanye nasekufundeni iqanda, kodwa ukusebenza kwawo kuthonywa ngezindlela eziningi indlela esenza ngayo izinto ngendlela,” kusho uChandrasekaran. “Ngokuya ngecala lokusebenzisa kanye nesizinda, lawa makhono azobalulekile kubo bobabili abasebenzisi be-IT nabamabhizinisi.”

Ama-API amaningi avumela abasebenzisi ukuthi basebenzise amasu abo wobunjiniyela ngokushesha. Noma nini lapho umsebenzisi ethumela umbhalo ku-LLM, kunethuba lokucwengisiswa kokwaziswa ukuze kuzuzwe imiphumela ethile, ngokusho kukaReyes.

“Kodwa-ke, lokhu kuvumelana nezimo kubuye kuvule umnyango wamacala okusebenzisa kabi, njengokujova ngokushesha,” kusho uReyes. “Izimo ezinjengalezi [Microsoft’s] I-Bing’s Sydney wabonisa ukuthi abantu bangasebenzisa kanjani ubunjiniyela obusheshayo ngezinjongo ezingahlosiwe. Njengomkhakha okhulayo wocwaningo, ukubhekana nomjovo osheshayo kuzo zombili izimo zokusetshenziswa ezinonya kanye ‘nokuhlanganisa iqembu elibomvu’ ukuze kuhlolwe ipeni kuzobaluleka esikhathini esizayo, ukuqinisekisa ukusetshenziswa okunesibopho nokuvikelekile kwama-LLM kuzo zonke izinhlelo zokusebenza ezihlukahlukene.”