- Amamodeli ezilimi aqondene nesizinda ahweba ngolwazi olubanzi ukuze athole ulwazi olujulile, athuthukisa ukunemba kanye nokwethenjwa emikhakheni elawulwayo nenezingqinamba ezinkulu.
- Ama-DSLM kanye namamodeli amancane olimi anciphisa izindleko, avumela ukuthunyelwa endaweni noma kudivayisi futhi anikeza ukuvikelwa kwedatha okuqinile kanye nokuhambisana nomthetho.
- Ukuhlanganisa amamodeli akhethekile ne-Retrieval-Augmented Generation kudala izakhiwo eziqinile ezinciphisa ukubona izinto ezingekho futhi zihlale zisesikhathini.
- Amamodeli akhethekile asevele enza kahle kakhulu kune-LLM enkulu kwezezimali, umthetho, ezokwelapha kanye nokubhala amakhodi, okushintsha indlela isofthiwe ehlanganisa ngayo i-AI.
Amamodeli ezilimi aqondene nesizinda (ama-DSLM) asheshe abe umgogodla wangempela we-AI ekhiqizayo esebenzayo, ikakhulukazi ezimbonini lapho ukunemba, ukulawulwa kanye nokwethenjwa kungaxoxiswana ngakho. Esikhundleni sokuzama ukuba muhle kukho konke, lawa mamodeli asebenza kabili endaweni eyodwa - njengokunakekelwa kwempilo, ezezimali, umthetho noma izinhlelo - futhi afunde ngokujulile. Abahlaziyi abafana noDanielle Casey waseGartner sebevele bexwayisa ngokuthi izinkampani ezinamathela kuphela kumamodeli amakhulu olimi (ama-LLM) azoqala ukuzwa ubuhlungu ngendlela yezindleko zokusebenza eziphakeme kanye nengozi ekhulayo.
Ukushintshela ku-GenAI esetshenziswa kabanzi kuye kuma-DSLM akhethekile akuyona nje into edlulayo, kodwa kuyisidingo sezomnotho nokuncintisana.UMcKinsey ulinganisela ukuthi i-AI ekhiqizayo ingafaka phakathi kwamaRandi ayizigidigidi ezingu-2.6 no-4.4 ngonyaka emnothweni womhlaba, okunomthelela omkhulu kakhulu emikhakheni elawulwa kakhulu. Kulezo zimo, imodeli "ezwakala ihlakaniphile" ayanele; izinhlangano zidinga izinhlelo eziqonda ngempela izici zobuchwepheshe zesizinda sazo futhi zingasetshenziswa ngokulawula okuqinile idatha, ukuthobela imithetho kanye nezindleko.
Iyini ngempela imodeli yolimi oluqondene nesizinda?
Imodeli yolimi oluqondene nesizinda iwuhlelo lwe-AI oluqeqeshwe ngokuyinhloko kudatha evela ensimini eyodwa, njengezokwelapha, umthetho, amabhange noma ukuthuthukiswa kwesofthiweNgenkathi ama-LLM ajwayelekile edla ingxube enkulu yombhalo we-inthanethi nolwazi olubanzi, ama-DSLM agxila ezinkampanini ezikhethekile: iziqondiso zezokwelapha, imibono yezomthetho, imibhalo yokulawula, amaphepha ezimali, izincwadi zokubhala eziphathelene nabanikazi kanye nemithombo efanayo.
Umgomo oyinhloko walobu buchwepheshe ukufeza ukunemba okuphezulu kwamaqiniso, ukungaboni kahle okuncane kanye nokucabanga okuthembekile emisebenzini yangempela yomhlabaNgamanye amazwi, la mamodeli ahwebelana ububanzi ngokujula: awazami “ukwazi konke ngakho konke”, kodwa aba nekhono futhi athembeke kakhulu ngaphakathi kwendawo aqeqeshwe kuyo. Yilokhu kanye okudingayo uma iphutha lingasho ukuxilongwa okungalungile, umbiko wezezimali ongahambisani nomthetho noma impikiswano yezomthetho enephutha.
Uma kuqhathaniswa nama-LLM ajwayelekile, ama-DSLM aklanyelwe ukubamba amagama aqondile, imithetho engacacile kanye nomongo ocashile womkhakha othile.Imodeli evamile ingase ihlupheke ngencazelo eqondile yemiqondo efana nethi “habeas corpus” emthethweni noma “PRN” emibhalweni yezokwelapha, noma ichaze kabi amagama asetshenziswayo okulawula. I-DSLM eqeqeshwe ngedatha yesizinda esinegunya inamathuba amaningi okuhumusha imishwana enjalo ngendlela efanele futhi iqonde ukuthi isebenzisana kanjani nemingcele ebanzi, iziqondiso noma izinhlaka zomthetho.
Esinye isici esibalulekile sokuhlukanisa ukuthi ama-DSLM angena kanjani esitokisini se-AI senhlangano, okuhlanganisa ukwakheka kwamaqembu e-ejenti ye-AIEsikhundleni sokusebenza njengezinhlobo zobuchopho ezifanela wonke umuntu efwini, zivame ukuba amamodeli amancane, agxile kakhulu angalungiswa, ahlolwe futhi alawulwe ngokuxhumana okuqinile nochwepheshe besizinda. Lokho kuzenza zifaneleke kangcono ezimbonini lapho kubalulekile ukwazi ukuthi imodeli yakho ingenzani nokuthi yini engenakuyenza, futhi ibhale phansi ukuziphatha kwayo kubahloli bamabhuku noma abalawuli.
Ngokombono webhizinisi, ama-DSLM avumelana ngqo nokusunduza i-AI okuphephile, okuchazekayo nokungahlolwaAbalawuli kuzo zonke izifunda balola imithetho ephathelene nokuvikelwa kwedatha, ukuzibophezela kwe-algorithm kanye nengozi ethile emkhakheni. Imodeli encane, eboshelwe isizinda - okungenzeka isetshenziswe endaweni futhi iqeqeshwe kuphela emithonjeni ehlolwe - kulula kakhulu ukuyibeka ngaphansi kokubusa kune-LLM enkulu ejwayelekile ebambe ingxenye ye-inthanethi.
Ama-DSLM aba kanjani akhethekile?
Ubuchwepheshe be-DSLM buvela esu layo lokuqeqesha kanye nedatha yayo, hhayi emaqhingeni obunjiniyela obusheshayo noma emigqeni embalwa yokucushwa.Ukutshela nje i-LLM ejwayelekile ukuthi “isebenze njengodokotela” noma “iziphathe njengochwepheshe webhange” ngokushesha akubhali kabusha ulwazi oluyisisekelo lwemodeli. Kushintsha isitayela sayo kanye nokugxila kwayo ngokunganaki.
Kunezindlela ezimbili eziyinhloko zobuchwepheshe zokwakha i-DSLM: ukuqeqeshwa kusukela ekuqaleni kanye nokulungisa imodeli eyisisekeloUkuqeqeshwa kusukela ekuqaleni kusho ukuqala ngamapharamitha aqalwe ngokungahleliwe bese unikeza imodeli umbhalo okhethwe kahle kakhulu, othize wesizinda. Ngokuphambene, ukulungisa kahle, kuthatha imodeli esivele iqeqeshwe, evamile bese kuyivumelanisa kusetshenziswa amasethi edatha akhethekile avela emkhakheni oqondiwe.
Ukuqeqeshwa okugcwele kusukela ekuqaleni kunikeza ukulawula okuphezulu phezu kwesethi yedatha kanye nokubandlulula okukhuthazayo kwemodeliUma uhlanganisa i-corpus eyenziwe ngezincwadi zezokwelapha kuphela, imibiko yesilingo sezokwelapha kanye neziqondiso, ungakha imodeli efana ne-BioBERT ehlanganisa amaphethini olimi lwezokwelapha ngokujulile. Ukushintshana ukuthi ukuqoqa idatha, ukuqeqesha imodeli kanye nokuqinisekisa ukuziphatha kwayo kuyabiza ngokwesikhathi, ukubala kanye nomsebenzi wochwepheshe.
Ukulungisa kahle kuvame ukuba yindlela ewusizo kakhulu ezinkampanini eziningiNgokuqala ku-LLM eqinile ejwayelekile, usebenzisa kabusha ikhono lolimi olubanzi lwemodeli kanye nolwazi lomhlaba, bese uyidonsela endaweni yakho ngezibonelo eziqondiwe. Isibonelo, i-DSLM egxile emthethweni ingadalwa ngokulungisa imodeli eyisisekelo ngezinqumo zenkantolo, izinkontileka, imithetho kanye nezimpendulo zemibuzo ezifana nezivivinyo ze-bar, konke kubuyekezwa ochwepheshe bezomthetho.
Kungakhathaliseki ukuthi iyiphi indlela ekhethiwe, ikhwalithi yedatha yesizinda ibaluleke kakhuluAma-DSLM asebenza ngamadokhumenti ambalwa kodwa athembekile kakhulu uma kuqhathaniswa namamodeli ajwayelekile. Lokhu kungafaka phakathi izincwadi zobuchwepheshe zangaphakathi, izinqubo zokusebenza ezijwayelekile, izinqubomgomo zangaphakathi, imithethonqubo ethile yomkhakha, imibiko yamacala engaziwa, noma izinkampani zezimali nezomthetho ezikhethiwe. Izinga elincane livumela ukuhlolwa nokuhlanza okuqinile, okuhunyushwa ngqo emiphumeleni ezinzile nethembekile.
Olunye ungqimba lobuchwepheshe luvela ezindleleni zokuhlola ezinolwazi lwesizinda kanye nezilinganisoEsikhundleni sokuhlola ukusebenza emisebenzini ejwayelekile njengokubhala okuvulekile noma izibalo ezilula, ama-DSLM aqinisekiswa kusetshenziswa izivivinyo ezithile zomkhakha: izilinganiso ze-QA yezokwelapha, izilinganiso zokubona izinto ezingekho emthethweni, imisebenzi yokuhlaziya imizwa yezezimali kanye nemibhalo, noma izinselele zekhodi yokuhlela. Ochwepheshe abavela emacaleni okuhlola umkhakha, balungisa amalebula futhi basize ekuchazeni ukuthi "okulungile ngokwanele" kubukeka kanjani ekusebenzeni.
Kungani ama-LLM ajwayelekile efika ophahleni ezindaweni ezikhethekile
Ama-LLM ayisisekelo njenge-GPT, i-Gemini, i-Claude noma i-LLaMA abangele uguquko lwangempela endleleni isofthiwe esebenzelana ngayo nolimi lwemveloBangafingqa imibhalo emide, babhale okuqukethwe, bahumushe phakathi kwezilimi, bakhiqize ikhodi futhi baphendule imibuzo yolwazi olubanzi ngokushelela okumangalisayo. Emisebenzini eminingi yansuku zonke, kakade sekwanele.
Kodwa-ke, lawa mamodeli afanayo ahlala ehlushwa imininingwane emincane ebaluleke kakhulu emikhakheni ekhethekile nelawulwayo, ukuboniswa kwe imikhawulo kanye nezingozi zama-LLMUma umbuzo udinga ukuchazwa okucashile kwemithetho, ukufundwa okuseduze kwesiqondiso sezokwelapha noma ukuhambisana okunembile nendinganiso yobuchwepheshe ekhethekile, ama-LLM ajwayelekile anamathuba amaningi okuphutha noma enze izimpendulo ezizwakala zinegunya kodwa zinganembile.
Lo mkhawulo awugcini nje ngamaphutha ezikhathi ezithile; uphazamisa inani lokusebenza kohlelo.Uma uhlaka lwakho lokuphathwa kwezingozi luphoqa uchwepheshe ongumuntu ukuthi aqinisekise yonke impendulo ye-AI ngaphambi kokuyisebenzisa, inzuzo yokukhiqiza elindelekile iyanyamalala. Udokotela, ummeli noma isikhulu sezingozi abakwazi ukuthembela kumodeli oziphatha njengomfundi oqeqeshwayo onekhono kodwa ongathembekile.
Ukuze balungise lobu buthakathaka, amaqembu amaningi aphendukele ku-Retrieval‑Augmented Generation (RAG). Ekusethweni kwe-RAG, imodeli ayiphenduli nje kuphela kusuka kumapharamitha ayo angaphakathi; kunalokho, iqala ngokusesha isisekelo solwazi noma isitolo samadokhumenti, ithole izindima ezifanele bese izisebenzisa njengomongo lapho ikhiqiza impendulo. Lokhu kugcina okuqukethwe kusha futhi kukuvumela ukuthi uqinise izimpendulo emithonjeni oyilawulayo.
I-RAG iwusizo kakhulu, kodwa ayishintshi indlela imodeli eyisisekelo ebangela ngayo. I-LLM eyisisekelo ingase isaqonda imiqondo yesizinda, ifunde kabi izingcezu ezitholiwe noma intule ukuqonda okujulile kwesakhiwo semithetho ensimini yakho. I-RAG isiza ukuvimbela imibono engaqondakali ngokubeka izimpendulo kumadokhumenti, kodwa ayikwazi ukulungisa ngokuphelele ukuntuleka kobuchwepheshe ngaphakathi kwemodeli ngokwayo, ikakhulukazi lapho imibuzo ixubile noma lapho imibhalo eminingi ingqubuzana.
Ngenxa yalokhu, ukuthembela kuphela ku-LLM ejwayelekile kanye ne-RAG ngokuvamile akwanele ekusetshenzisweni okuyingozi kakhuluUngagcina unesistimu ethola idokhumenti efanele kodwa ichaze kabi imiphumela yayo, noma ehluleka ukuvumelanisa imithethonqubo ehlukene ngendlela efanele. Yilesi kanye igebe ama-DSLM aklanyelwe ukuligcwalisa: ukuqonda kwangaphakathi, kweqiniso kwesizinda okuhlanganiswe nokutholwa kwangaphandle lapho kudingeka khona.
Ushintsho lobuchwepheshe ngaphakathi kwe-DSLM
Ngaphansi kwe-hood, ama-DSLM ahlukile kuma-LLM abanzi ikakhulukazi ngobubanzi bedatha, ukuhlolwa kanye namaphethini okusetshenziswaNgokuvamile basebenzisa isethi yedatha encane kodwa eqinile futhi bahlolwa ngokubheka amaphrofayili amaphutha athile kakhulu: imibono engekho emthethweni, izincomo ezingaphephile ngokwezokwelapha, ukuchazwa kabi kwemithetho yezezimali, noma ukuphathwa ngokunganaki kwezihlonzi ezibucayi.
Isethi yedatha eyinhloko ye-DSLM ivame ukugxila emithonjeni yolwazi lwesizinda olunenani eliphezuluEzindaweni zezimboni, lokho kungaba imibhalo yobuchwepheshe enemininingwane, izincazelo zenqubo, amazinga obunjiniyela kanye nezisekelo zolwazi lwangaphakathi. Emkhakheni wezomthetho, kungafaka umthetho, umthetho, isiqondiso sokulawula kanye nokuphawula kwezimfundiso. Kwezokwelapha, izincwadi zezokwelapha, iziqondiso zezokwelapha, amarekhodi ezempilo kagesi angachazwanga kanye nezincwadi ezibuyekezwa ontanga zidlala indima ebalulekile.
Ngaphezu kwedatha eluhlaza, ama-DSLM abhekana nokulungiswa okuhle nokuqondiswa okuqondiswayo okuholwa ochwepheshe besizindaAbameli bangase babhale izingcaphuno ezifanele kanye nezindlela zokucabanga, odokotela bangase baveze izincomo ezingaphephile noma ezidukisayo, futhi izikhulu zokuthobela imithetho zingasiza ekubhaleni ukuziphatha okungenangozi okungalungile. Lokhu kuqapha kuqondisa imodeli kude nezimpendulo ezinengqondo kodwa eziyingozi.
Ukuhlola kulandela ifilosofi efanayo egxile endaweniEsikhundleni sokusebenzisa amabhentshi ajwayelekile kuphela emisebenzini yokucabanga evamile noma yolimi, ama-DSLM ahlolwa kusetshenziswa ama-metric namasethi edatha akhethekile: amabhentshi okubona izinto ezingekho emthethweni njenge-Stanford Legal Hallucination Benchmark, izinselele zokuqashelwa kwezinhlangano zezokwelapha, imisebenzi yokukhipha ulwazi lwezezimali, ukuhlolwa kokuqedwa kwekhodi kanye nokulungisa amaphutha, noma amasethi e-Q&A athile embonini. Ukusebenza kulezi zivivinyo kubonisa ngqo inani lemodeli ekusetshenzisweni kwangempela.
Amamodeli amancane, aqaphela isizinda nawo enza kube lula ukuhlanganisa izakhiwo ezithuthukisiwe njenge-RAG ngendlela elawulwa kakhuluEsikhundleni sokuthembela kumodeli omkhulu ojwayelekile futhi unethemba lokuthi ukubuyisa kuzovala izikhala zolwazi, izinhlangano zingasebenzisa i-DSLM encane njengenjini yokucabanga eyinhloko bese zinamathisela ungqimba lwe-RAG ukuze zilunike amadokhumenti amasha noma aqondene kakhulu nomongo, zinciphise kokubili ukuphelelwa yisikhathi kanye nokubona izinto ezingekho.
Umphumela uba ukwakheka lapho i-DSLM isebenza njenge-nucleus yokuqonda, kuyilapho i-RAG inikeza ibhuloho elinamandla lokuthola ulwazi oluphilayo.Lokhu kuhlanganiswa kunamandla kakhulu ezindaweni lapho imithetho nolwazi kushintsha khona njalo – isibonelo, imithethonqubo eshintshashintshayo, iziqondiso zokwelashwa noma izimo zezimali ezishintsha ngokushesha – ngoba ukuqonda komqondo wemodeli kuzinzile, kodwa usengashintsha idatha ebuyekeziwe ngaphandle kokuqeqeshwa kabusha kusukela ekuqaleni.
Izinzuzo zebhizinisi ze-DSLM zamabhizinisi
Ngokombono wesu, ukwamukela ama-DSLM kune-LLM ezijwayelekile kunikeza izinhlangano izinzuzo ezicacile nezilinganisekayoLezi zinzuzo zisukela ekunembeni okungcono kanye nokuqondaniswa kwemithetho kuya ekongeni izindleko kanye nokwethenjwa kwabasebenzisi okuthuthukisiwe, konke okuhambisana ngqo nembuyiselo yokutshalwa kwezimali.
Okokuqala, ama-DSLM avame ukuletha ukunemba okuphezulu kakhulu kobuchwepheshe kanye nokuqonda isizindaNgenxa yokuthi baqeqeshwe futhi bahlelwe kahle yizinkampani ezikhethekile, cishe abaqondi kahle amagama athile esizinda, bahlanganise imiqondo efanayo noma bangazinaki izinkomba ezicashile zomongo. Emthethweni, lokho kusho ukubhekisela okuthembekile kakhulu emithethweni nasemthethweni wamacala; kwezempilo, ukunamathela kangcono eziqondisweni zemitholampilo; kwezezimali, ukuhluzwa okunembile kwemibiko kanye nezinkomba zezingozi.
Okwesibili, ama-DSLM anikeza iziqinisekiso eziqinile mayelana nokuphepha kwedatha, ubumfihlo kanye nokuthobela imithetho. Eziningi zalezi zinhlobo zenzelwe ukusebenza endaweni noma ngaphakathi kwendawo yamafu elawulwa ngokuqinile, kusetshenziswa kuphela amasethi edatha ahlangabezana nokuphathwa kwangaphakathi kanye nezimfuneko zomthetho zangaphandle. Lokhu kufaneleka ngokwemvelo emikhakheni enemithetho eqinile yedatha yomuntu siqu (PII), izimfihlo zokuhweba noma ubumfihlo bamakhasimende.
Okwesithathu, amamodeli akhethekile angaba ngcono futhi ashibhile ukuwasebenzisa kunalawo amakhulu, anezinjongo ezijwayelekileNgenxa yokuthi ama-DSLM avame ukuba namapharamitha ambalwa futhi alungiselelwe imisebenzi emincane, ukuqagela kungaba okusheshayo futhi kungadingi kakhulu izinsiza. Lokho kuholela ezindleleni eziphansi zokukhonza, okuhlangenwe nakho komsebenzisi okubushelelezi kanye nokwenzeka kokusebenzisa amamodeli kumadivayisi asemaphethelweni noma kumaseva aphansi esikhundleni samaqoqo amakhulu e-GPU.
Okwesine, ama-DSLM ayithuluzi elinamandla lokunciphisa ukubona izinto ezingekho emthethweni ezisetshenziswayo. Uma zihlanganiswe ne-RAG, azivamile ukusungula imiqondo noma izingcaphuno ezingekho, ngoba ulwazi lwazo lwangaphakathi kanye nokuhlola kuye kwaklanywa ukuze kubekwe phambili ukunemba kwesizinda. Lokhu kunciphisa umzamo wesandla odingekayo ukuqinisekisa imiphumela ye-AI futhi kusiza ekwakheni ukwethembana phakathi kwabasebenzisi abangochwepheshe.
Idatha yemboni isivele ibonisa lolu shintshoUcwaningo lwasekuqaleni lubonisa ukuthi ingxenye enkulu yezinkampani ezisebenzise ama-DSLM zibika ukunemba okuphezulu kanye ne-ROI enamandla kunalezo ezithembele kuphela kumamodeli enhloso ejwayelekile. Abahlaziyi bacabanga ukuthi ngo-2027, amamodeli e-GenAI angaphezu kwengxenye asetshenziswa kakhulu emabhizinisini azobe eqondene nesizinda, kunokuba kube ama-LLM ajwayelekile atholakala ngama-API ajwayelekile.
Izindaba zempumelelo ze-DSLM zomhlaba wangempela
Umqondo wokuthi "okukhulu kuhlale kungcono" ku-AI uphikiswe ngokusobala uhlu olukhulayo lwamamodeli akhethekile asebenza kangcono kunezinhlelo ezinkulu ezijwayelekile endaweni yazo.Lezi zimo zomhlaba wangempela zibonisa ukuthi ukugxila kwesizinda okuqinile kanye nedatha ekhethiwe kunganqoba kanjani ukubalwa kwamapharamitha angahluziwe.
I-BioBERT iyisibonelo esivelele esivela emkhakheni wezokwelapha. Yakhelwe phezu kokwakhiwa kwe-BERT kodwa iqeqeshwe ngqo kuma-corpora njengezifinyezo ze-PubMed kanye nezihloko ze-biomedical ezigcwele umbhalo, i-BioBERT ikhombisa ukusebenza okungcono kakhulu emisebenzini efana nokuqashelwa kwezinhlangano eziqanjwe nge-biomedical, ukukhipha ubudlelwano kanye nokuphendula imibuzo uma kuqhathaniswa namamodeli ajwayelekile esitayela se-BERT. Ubuhle bayo buvela ekwazini okujulile amagama esizinda, ama-acronym kanye nemigomo yocwaningo.
Kwezezimali, iBloombergGPT ikhombisa ukuthi imodeli eqeqeshwe yisizinda ingawubumba kanjani kabusha umsebenzi ohamba phambili onenani eliphezuluNjengoba inamapharamitha angaba yizigidigidi ezingama-50, akuyona imodeli enkulu kunazo zonke ekhona, kodwa iqeqeshwe ngamanani amakhulu edatha yezezimali kanye nezindaba. Ngokwezilinganiso zangaphakathi, iBloombergGPT kubikwa ukuthi isebenza kangcono kunezinhlobo ezijwayelekile ezifanayo ngamaphesenti angaphezu kwama-60 emisebenzini efana nokuhlukaniswa kwamadokhumenti, ukukhipha ulwazi kanye nokuhlaziywa kwemizwa yemibhalo efanelekile emakethe.
Emkhakheni wezomthetho, amathuluzi anjenge-Paxton AI aqokomisa indlela ama-DSLM ahlelwe ngokucophelela anganciphisa ngayo amazinga okubona izinto ezingekho emthethweni. Kuhlolwe ku-Stanford Legal Hallucination Benchmark, lolu hlobo lwemodeli lufinyelela amazinga aphezulu kakhulu okunemba kwemibuzo nezimpendulo zomthetho, ukuhlaziywa kwamacala kanye nokuchazwa komthetho, okwenza kube umsizi othembekile kakhulu kubameli uma kuqhathaniswa nama-LLM ajwayelekile angase aqambe izingcaphuno zamacala noma afunde kabi imithetho yenqubo.
Ukuhlela kungenye indawo lapho amamodeli akhethekile ekhanya khona. I-StarCoder, isibonelo, yakhelwe ekuqondeni nasekukhiqizeni ikhodi. Ukuphindaphinda kwayo kuka-2024 kubonise ukuthi imodeli enamapharamitha angaba yizigidigidi eziyi-15, uma iqeqeshwe ezindaweni zokugcina amakhodi ezikhethwe ngokucophelela, ingadlula amamodeli amakhulu okubhala amakhodi ajwayelekile njenge-CodeLlama yepharamitha eyizigidigidi ezingama-34 kumabhentshimakhi amaningi afanele onjiniyela. Futhi, ukuqeqeshwa okugxile kanye nekhwalithi yedatha kudlula usayizi omkhulu.
Ngaphandle kwalezi zimo eziyinhloko, abadlali abaningi bezimboni basebenzisa buthule ama-DSLM aboIzinkampani ezifana ne-Siemens ne-Bosch ziye zazama amamodeli avunyelaniswe namadokhumenti azo obunjiniyela bangaphakathi nolwazi lwenqubo, kuyilapho i-Med-PaLM ye-Google DeepMind ihlose imibuzo nezimpendulo zezokwelapha kanye nokucabanga kwesitayela sezokwelapha. UHarvey ukhonza imakethe yezomthetho ngokugxila ocwaningweni, ekubhaleni nasekuhlaziyeni okulungiselelwe umkhuba wezomthetho.
Ukuvela Kwamamodeli Olimi Oluncane (ama-SLM)
Okuhlobene kakhulu nama-DSLM ukuthambekela okusha kwamaModeli Olimi Oluncane (ama-SLM)Lawa amamodeli amancane ngamabomu, avame ukuqeqeshwa kusukela ekuqaleni noma anqunyiwe futhi alungiswe kakhulu, agxila ezizindeni ezithile noma imindeni yemisebenzi ngenkathi egcina ukusetshenziswa kwezinsiza kuphansi. Ahambisana kahle nezidingo zebhizinisi zokulawula, ukusebenza kahle kwezindleko kanye nokusetshenziswa endaweni.
Ukuqeqesha i-SLM ethile yesizinda kusukela ekuqaleni kunikeza izinhlangano ithuba lokuklama imodeli ezungeze idatha yazo kanye nemikhawulo yazo.Esikhundleni sokushintsha imodeli enkulu evamile, bangakha uhlelo oluncane oluhambisana nesilulumagama sabo, isakhiwo sedokhumenti kanye namaphethini okusebenza. Lokhu kuyakhanga kakhulu lapho idatha yobunikazi ingakwazi ukuphuma engqalasizinda yenhlangano ngenxa yezizathu zomthetho noma zokuncintisana.
Enye yezinzuzo ezibaluleke kakhulu zama-SLM ukuphetha okushibhile nokusheshayo. Njengoba zinamapharamitha ambalwa kanye nenhloso ebanzi, zingasebenza kahle kuma-CPU noma kuma-GPU aphansi, noma ngisho ngqo kumadivayisi asemaphethelweni. Lokhu kwenza kube ngokoqobo ukufaka amakhono e-AI ngqo kumikhiqizo yesofthiwe, imishini yezimboni noma kumadivayisi abasebenzisi ngaphandle kokuthembela njalo kumasevisi efu.
Ama-SLM aphinde avule ukuthunyelwa okusebenzayo endaweni ethile emikhakheni enezidingo eziqinile zobumfihlo kanye nobumfihloIzinhlelo zezempilo, amabhange, izinkampani zomshuwalense kanye nabaqhubi bezingqalasizinda ezibalulekile bavame ukungathandi ukusakaza idatha ebucayi kubahlinzeki bezinkampani zangaphandle. Ukusingatha i-SLM encane neqondakala kahle ngaphakathi kwendawo yabo kubavumela ukuthi bagcine idatha yendawo ngenkathi besavuna izinzuzo ze-GenAI.
Izakhiwo ezibheke phambili manje sezihlanganisa ama-SLM noma ama-DSLM njengenjini yokucabanga eyinhloko nesendlalelo se-RAG njengomhlinzeki womongo oguqukayoImodeli ihlanganisa ukuqonda okuzinzile kwesizinda kanye nokuziphatha okuzenzakalelayo, kuyilapho i-RAG iyivumela ukuthi ilande izinqubomgomo, iziqondiso, izinkontileka noma imininingwane yobuchwepheshe esesikhathini. Le ndlela inciphisa isidingo sokuqeqeshwa kabusha njalo, ngoba ulwazi lwangaphandle kuphela oludinga ukubuyekezwa njengoba amadokhumenti eshintsha.
Abahlaziyi bezimboni sebevele bakhetha ama-SLM kanye nama-DSLM njengobuchwepheshe obubalulekile okufanele buqashwe eminyakeni embalwa ezayo.Esikhundleni sekusasa elibuswa imodeli eyodwa enkulu, yendawo yonke, sibheke endaweni ehlukahlukene lapho kuhlangana khona amamodeli amaningi amancane, akhethekile, ngalinye lenzelwe ingxenye ethile yangempela futhi lihlanganiswe emikhiqizweni, emisebenzini nasemadivayisini.
Ukusebenzisa ama-LLM nama-DSLM endaweni: imiphumela kudivayisi
Uma ucabanga ngendlela yokuletha amakhono e-DSLM kubasebenzisi, izinketho zokusebenzisa zibaluleke kakhulu njengokwakhiwa kwemodeli.Ungasebenzisa amamodeli ngama-API efu, uwaphathe ngokwakho engqalasizinda yakho noma uwafake ngqo kumadivayisi abasebenzisi kusiphequluli, kudeskithophu noma kuselula.
Izinsizakalo ze-LLM ezisekelwe efwini zisanikeza izinzuzo ezinamandla. Banikeza ukufinyelela kumamodeli amakhulu kakhulu futhi anekhono, ngokubona okuphendulayo kanye nentengo yokukhokha ngethokheni engaba eyongayo ngezinga. Amanye amamodeli akhethekile kubathengisi abathile bamafu, njenge Ukuhlanganiswa kwe-Gemini ku-OCI, futhi amabhizinisi angazuza emsebenzini wokuthuthukisa nokuthuthukisa oqhubekayo wabahlinzeki ngaphandle kokuphatha ingqalasizinda ngokwabo.
Kodwa-ke, izindlela zasendaweni kanye nezama-app ziye zaba zikhanga kakhulu, ikakhulukazi kuma-DSLM kanye nama-SLMUkusebenzisa amamodeli ngqo kusiphequluli ngokusebenzisa ubuchwepheshe obufana ne-WebLLM, noma ngezixhumi zokuhlola ezifana ne-Chrome's Prompt API, kuvumela ukusebenza okungaxhunyiwe ku-inthanethi, ukubambezeleka okuqhubekayo kanye nokulawula okugcwele idatha yomsebenzisi. Lokhu kulungele izinhlelo zokusebenza ezifana nabaphathi bemisebenzi, amathuluzi okukhiqiza noma amadeshibhodi athile esizinda acebile ngezici ze-chatbot.
Ama-LLM kanye nama-DSLM akudivayisi nawo athuthukisa kakhulu ubumfihlo kanye nokuphephaUma idatha yomsebenzisi ingaphumi kudivayisi, asikho isidingo sokudlulisa ulwazi lomuntu siqu noma okuqukethwe kwebhizinisi okubucayi kumaseva ezinkampani zangaphandle. Kuma-domain alawulwayo, lokhu kungenza kube lula ukuthobela imithetho kakhulu futhi kunciphise indawo yokuhlaselwa kokwephulwa kwedatha.
Vele, kukhona ukuhwebelana ngokusebenzisa amamodeli endaweni. Osayizi bamamodeli banqunywa isitoreji sedivayisi kanye nememori, ukulandwa kwezindawo zokuhlola ze-multi‑gigabyte kungaba kancane, futhi amamodeli amancane endawo angase asale ngemuva kwama-giant aphethwe amafu ngekhono lokucabanga elijwayelekile. Kuma-DSLM, lokhu kugcizelela kakhulu ukukhethekile ngokucophelela, ukunqunywa kanye nokwenza ngcono ukuze imodeli inikeze amakhono aqinile esizinda ngaphakathi kwesabelomali esincane sezinsizakusebenza.
Naphezu kwalezi zithiyo, inhlanganisela yama-SLM, ama-DSLM kanye nezikhathi zokusebenza ezikudivayisi kuvula umnyango wesigaba esisha sesofthiwe enikwe amandla yi-AI.Cabanga ngethuluzi lokucwaninga lezomthetho, umsizi wenothi yezokwelapha noma ideshibhodi yezezimali ene-chatbot ekhethekile eyakhelwe ngaphakathi eqhubeka isebenza ngisho nangaphandle kokuxhumana kwenethiwekhi, ehlonipha izinqubomgomo zedatha yendawo futhi elawulwa ngokugcwele yinhlangano eyisebenzisayo.
Amacala okusetshenziswa okusebenzayo: kusukela ohlwini lwezinto okufanele zenziwe kuya emisebenzini yezimboni
Ubuchwepheshe obufanayo be-LLM obusebenzisa amathuluzi ezimboni athile esizinda bungathuthukisa nezinhlelo zokusebenza ezilula kakhuluCabanga ngohlelo lokusebenza lwewebhu lohlu lwezinto okufanele zenziwe olujwayelekile: abasebenzisi bangangeza imisebenzi, bayimake ngokuthi iqediwe bese beyisusa. Uma uqala ukuyibuka, iyi-interface ye-CRUD elula engenasidingo se-AI ethuthukisiwe - kodwa ama-LLM nama-DSLM angathuthukisa ulwazi ngendlela enenjongo.
Ukuhlanganisa i-chatbot yendawo kulolu hlobo lohlelo lokusebenza kuvumela abasebenzisi ukuthi babuze futhi basebenzise idatha yabo ngolimi lwemvelo.Bangase babuze ukuthi mingaki imisebenzi evulekile esele, bacele uhlu lwezinto ezidlulelwe yisikhathi, noma bathole iziphakamiso zezinyathelo ezilandelayo ngokusekelwe emisebenzini eqediwe ngaphambilini. Imodeli elungisiwe yesizinda semisebenzi yokukhiqiza ingaqagela izigaba, ithole okuphindwe kabili futhi iphakamise ukuhlanganiswa ngokuhlakanipha kakhulu kunemithetho embalwa ebhalwe ngokuqinile.
Ama-Chatbot kuzinhlelo zokusebenza ezinjalo angadlula imibuzo elula futhi enze izinguquko zokuqukethweAbasebenzisi bangase bafune ukuhumusha imisebenzi baye kwezinye izilimi, bathumele uhlu lwabo nge-XML noma ezinye izinhlobo ezihlelekile, noma bakhiqize imisebenzi emisha ngokusekelwe kumaphethini emlandweni wabo. I-LLM efakwe nge-WebLLM noma isikhathi sokusebenza esifanayo ingaphatha lezi zicelo kudivayisi, ilondoloze ubumfihlo ngenkathi inikeza isikhombikubona esicebile sengxoxo.
Izimo zebhizinisi ezifuna kakhulu zilandela iphethini efanayo kodwa nge-DSLM ezikhethekileEsimweni sezokwelapha, i-DSLM ingasiza odokotela ukufingqa amanothi eziguli, izinketho zokwelapha ezihambisanayo zesiqondiso noma ukuhlola ukuthi umbiko owuhlaka uyahambisana yini nezindinganiso zamadokhumenti. Kwezezimali, imodeli ehlelwe ngezinhlaka zengozi zangaphakathi ingahlaziya amaphothifoliyo, iveze izinkinga zomthetho noma ifingqa amaphepha amade ngendlela ehambisana ne-taxonomy yenkampani.
Kuzo zonke izimo, ulimi lwemvelo luba umnyango ongaphambili wezinhlelo eziyinkimbinkimbi kanye namasethi edathaEsikhundleni sokuphoqa abasebenzisi ukuthi bafunde ukugeleza kwe-UI okuqinile noma izilimi zemibuzo, ungabavumela ukuthi bachaze inhloso yabo ngamagama avamile. I-DSLM ihumusha leyo nhloso, ibize amathuluzi noma ithole amadokhumenti nge-RAG lapho kudingeka khona, futhi ibuyisele izimpendulo ezizwakala sengathi ziyaxoxwa kodwa zinamathela emithethweni yesizinda.
Kwabathuthukisi besofthiwe, lokhu kumelela ushintsho olubanzi lwendlela yokucabangaEsikhundleni sokuhlanganisa ama-API amaningi namafomu athile kakhulu, bangaluke imodeli ekhethekile ekwakhiweni kwabo futhi bayisebenzise njengesendlalelo sesikhombimsebenzisi esiguquguqukayo. Ngakho-ke ama-DSLM nama-SLM ahambisana ne-backend logic yendabuko kanye nezizindalwazi, kunokuba azithathele indawo, asebenze njenge-semantic glue phakathi kwabantu nezinhlelo.
Ekugcineni, umfutho ongemuva kwamamodeli ezilimi athile kanye nezilimi ezincane ukhomba endaweni ye-AI eyakhiwe ngezinto eziningi ezigxile futhi ezinokwethenjelwa esikhundleni se-giant eyodwa yenhloso ejwayelekile.Izinhlangano ezitshala imali kusenesikhathi kuma-DSLM - zihlanganisa idatha ekhethiwe, ukuhlolwa okuqinile, ukuthunyelwa okusebenzayo, kanye, lapho kufaneleka khona, ukuqaliswa kwendawo - zizibeka esimweni sokubamba inzuzo yangempela yezomnotho ye-AI ekhiqizayo ngenkathi zigcina izingozi zilawulwa futhi ziqinisekisa ukuthi izinhlelo zazo ziyaziqonda ngempela izindawo ezisebenza kuzo.
