akuyona inkinga yekhwalithi yedatha. Akuyona inkinga yokuqeqeshwa. Akuyona inkinga ongayixazulula nge-RLHF eyengeziwe, ukuhlunga okungcono, noma iwindi lokuqukethwe elikhudlwana. Kuyimpahla yesakhiwo yalokho lezi zinhlelo ezithuthukisiwe ukuze zikwenze.
Ngibambe lesi sikhundla izinyanga, futhi ukusabela kuyabikezelwa: abacwaningi abasebenza ekukhuliseni ukubuyisa, amapayipi okulungisa kahle, nezindlela zokuqondanisa bangakhetha uhlaka olunethemba kakhudlwana. Ngiyaqonda ukuthi kungani.
Obekulahlekile kule mpikiswano i-geometry. Ukuqonda mayelana nezinjongo kanye nezakhiwo kuyadingeka kodwa akwanele. Kudingeka sivule imodeli futhi sibheke ukuthi empeleni kwenzekani ngaphakathi lapho isistimu ikhiqiza impendulo engalungile ngokuzethemba. Hhayi kumalogi. Hhayi kumaphethini wokunaka. Emgudwini wangaphakathi wokumelwa ngokwawo, isendlalelo ngokwesendlalelo, kusukela kokokufakayo kuye kokuphumayo. Yilokho okwenzile umsebenzi engiwethula lapha.
Lokho Okwaziwa Ukusakaza Okusele Ngaphambi kokuthi Imodeli Iqambe Amanga
Ukusetha kulula kakhulu. Sithatha ukwaziswa okuyiqiniso – uhlobo lapho i-transformer kufanele ibuyise khona inhlangano egciniwe – futhi siyisebenzisa ezimeni ezimbili: eyodwa lapho imodeli ikhiqiza impendulo efanele, enye lapho ikhiqiza khona impendulo engalungile (i-hallucination). Bese, silandelela umkhondo wokusakaza okusele – i-vector emele yangaphakathi – isendlalelo ngesendlalelo ngenethiwekhi. Umbuzo uwukuthi: ingabe lezi zindlela ezimbili ziyahlukana ngenxa yokuthi imodeli imane nje ingenayo inhlangano efanele? Noma kukhona okwenzekayo okuqondile?
Ukuze uqonde ukuthi lokho kusho ukuthini, cabanga ngesimo sangaphakathi semodeli kusendlalelo ngasinye njengephuzu esikhaleni – indawo enobukhulu obuphezulu. Njengoba imodeli isebenza ngokushesha, lelo phuzu liyahamba. Ilandela indlela. Yiziphi izinyathelo zokuhlola ukuthi indlela ethathwe ngesikhathi sempendulo efanele futhi indlela ethathwe ngesikhathi sokuphupha ziyahlukana ngoba indlela eyodwa imfishane – imodeli iphelelwa ulwazi – noma ngenxa yokuthi ziya ngezindlela ezihlukene ngenkathi zithatha ibanga elifanayo.
Impendulo ingeyesibili. Izindlela zinobude obufanayo. Bakhomba izindawo ezahlukene. Yilokho okubonisa uMfanekiso 1: ama-trajectories amabili ashiya imvelaphi efanayo, ehamba ibanga elifanayo, efika ezindaweni ezihlukene zesikhala. Eyodwa ibheke empendulweni efanele. Omunye kude nayo.
Isilinganiso Sokuzibophezela: Lapho Ukucindezelwa Kuvela Khona
Iphepha lethula imethrikhi ebizwa ngokuthi isilinganiso sokuzibophezela κ — empeleni, mangakanani amathuba esisindo semodeli aqondiswa ngenkuthalo noma kude nethokheni elungile kusendlalelo ngasinye.
Ekucubunguleni okulungile κ iphakama i-monotonically ngokusebenzisa inethiwekhi (Umfanekiso 2 — amajika abomvu, aluhlaza okwesibhakabhaka nampunga amnyama). Imodeli yakha ukuzibophezela empendulweni efanele ngokuqhubekayo. Yilokhu ongakulindela ohlelweni lokubuyisa inhlangano efundiwe.
Ekuboneni izinto ezingekho, kwenzeka okuhlukile. κ ayihlali nje isicaba, okungase kubonise ukwehluleka kokubuyisa — ukungabikho kwephethini yezibalo efanelekile. Kunalokho, u-κ uyagoqa (amajika anedeshi kuMfanekiso 2). Kuwo wonke amamodeli ahloliwe, u-κ ufinyelela ubuncane ngokuphawulekayo ngaphansi kwenani lawo lokuqala ngaphambi kokululama kancane ezingqimbeni zokugcina. Ku-LLaMA-2 13B naku-Mistral 7B, yehlela ku-κ_min = 0.08. Amanani we-p angaphansi kuka-10⁻¹⁰⁰. Lokhu akuwona umphumela “ocashile”.
Kwenzekani? Imodeli ayehluleka ukuthola impendulo efanele. Ihambisa ngenkuthalo isisindo samathuba kude nethokheni elungile kuzindlalelo ezifanayo lapho izobe ihambisa isisindo samathuba ibheke kuyo esimweni esifanele. Ukwehluleka kuwukweqa.
Imodeli ibhale ngekhodi impendulo efanele. Yilokho okwenza ukugoqa kuka-κ kuphawuleke. Uma imodeli ivele yantula ukuhlobana okufanelekile – uma i-“Paris” ingakaze ixhumeke “enhlokodolobha yase-France” ezisindweni – besizobona umkhondo oyisicaba noma onomsindo. Akukho ukucindezelwa. Ijiyomethri ibingeke ibe nolwazi.
Esikubonayo esikhundleni salokho i-trajectory eqala ngendlela efanele (wonke amajika ku-Figure 2 aqala ngokuyisisekelo endaweni efanayo) kodwa abese ephenduka. Ithokheni elungile iqongelela amathuba ezingqimbeni zakuqala, njengoba kugijima okulungile, bese ilahlekelwa yizingqimba ezimaphakathi, ekujuleni lapho kufanele ikhuphuke khona esimweni esifanele (amajika abomvu, aluhlaza okwesibhakabhaka kanye namnyama ampunga kuMfanekiso 1). Kungani? Impendulo eqotho ukuthi iphepha lisungula lokho ngokunemba futhi lishiya ukuthi kungani livulekile. Kodwa incazelo ezwakalayo kakhulu ukuncintisana. Lawa mamodeli awabuyisi amaqiniso angawodwa. Babikezela ithokheni elandelayo kumongo, futhi umongo udala ingcindezi yawo. Umusho obulokhu uya ohlangothini oluthile – ngokwesitayela, ngokwesihloko, ngokwe-syntactically – udala okwangaphambili okuqinile kokuthi kufanele kuqhubeke kanjani. Uma impendulo eyiqiniso ishayisana nalowo mkhangi onomongo, imodeli ayiphenyi uhlamvu lwemali. Isignali yomongo, eminyene futhi eqhubekayo kulo lonke ukulandelana, ingadlula isignali yeqiniso, okungenzeka ibe yingcosana kudatha yokuqeqeshwa.
Isignali yokuqeqesha ayizange itshele imodeli ngokusobala ukuthi ikhethe ukuhambisana kunokunemba. Itshele imodeli ukuthi ibikezele ithokheni elandelayo. Ukuhambisana nokunemba kuvame ukuqondanisa. Uma bengakwenzi, esikutholayo umugqa ompunga odayishiwe kuMfanekiso 2.
Imodeli ayiqambi amanga. Lenza lokho kanye ebelilungiselelwe ukukwenza. Lokho ingxenye engakhululekile.
Imibuso Emithathu
Okunye kokutholwe okuhlanzekile kwe-empirical ukuthi amamodeli ayisikhombisa awasabalalisi ngokuqhubekayo kunoma iyiphi i-axis yokuziphatha kokuphupha. Ziwela emaqenjini amathathu ahlukene:
| Amamodeli at 1B amapharamitha akhombisa ukuqala kokwabiwa kabusha kokunaka – ukuhlukaniswa okuthile kwejometri – kodwa ukucindezela okungaphelele. | Amamodeli at 1.6B–3B bonisa ukucindezelwa okuphakathi. Ukugoqa u-κ kukhona kodwa akujulile. I-StableLM-2 1.6B ifinyelela ku-κ_min = 0.32 kune-0.08. | Bese kuba ne-Gemma 2 2B, ehambisana nokujula kokucindezelwa kwe-LLaMA-2 13B kanye neMistral 7B naphezu kokuba nengxenye yamapharamitha azo (κ_min = 0.08, p <10⁻⁹¹). |
Okuthile okungokoqobo okwenzekayo ngokwezakhiwo, hhayi nje njengomsebenzi wesikali. Izinketho zezakhiwo – izindlela zokunaka, ukujwayela, ukwakheka kongqimba – kunquma uphahla ekujuleni kokucindezela ngaphandle kokubala kwepharamitha. Lesi yisakhiwo sesigaba.
Ukuthola Ama-Hallucinations
Senze imephu, ngokunemba kwejometri, ukuthi isigaba esithile sesistimu sihluleka kanjani. Umbuzo oyimbangela – yiziphi izifunda ezithile ezisebenzisa ukucindezelwa, futhi kungani – zihlala zivulekile. Leyo inkinga elandelayo. Lokho i-geometry esungulayo ukuthi ukucindezelwa akukona ngengozi. Akulona iphutha lokulinganisa ongalilungisa ngokutshelwa okungcono noma izinga lokufunda elihlukile. Kuyimpahla evelayo yamasistimu alungiselelwe ukuqagela kwethokheni elandelayo. Ukuhambisana kokuqukethwe kanye nokunemba kweqiniso yizinhloso ezehlukene. Lapho bengqubuzana, isignali yokuqeqesha ayihluleli phakathi kwabo. Ukweqa ukuthi lokho kungqubuzana kubukeka kanjani ngaphakathi.
Okushiwo okungokoqobo kuqondile. Ungasebenzisa lesi siginesha yejiyomethri ukuze wakhe izitholi ze-hallucination – ama-probe ahlonza izehlakalo zokucindezelwa ngaphambi kokuthi zifinyelele okukhiphayo. Basebenza kahle. Kodwa ngabendawo. Uphenyo oluqeqeshelwe ukubuyisa amaqiniso aludluliseli ngokuhlanzekile emisebenzini yokucabanga noma ezizindeni zolwazi ezihlukene. I-geometry iyashintsha ngokwanele ukuthi ukutholwa kwehlise isithunzi. Lokhu akulona iphutha endleleni. Kuwulwazi. Ikutshela ukuthi ukuqapha kudinga ukuthi kucaciswe isizinda, kulinganiswe ngomongo ngamunye wokuthunyelwa, hhayi ukufakwa kanye futhi kukhohlakale.
Kunoma ubani owakha amasistimu okukhiqiza ngezinga, leso isiphetho sokusebenza: imonitha eyodwa ngesizinda ngasinye, oqeqeshwe ngedatha emele evela kuleso sizinda. Okunye – umtshina owodwa wendawo yonke – awusekelwa ubufakazi.
Lokho Ijiyomethri Engakwazi Ukukulungisa
Indlela yokukhipha la madokhumenti omsebenzi akusona “isiphazamisi esilindele ukupeshishwa”. Kungumphumela oqondile womsebenzi wenhloso osetshenziselwa ukuqeqesha ama-LLM. Ukubikezela kwethokheni elandelayo phezu kokulandelana okuhlukahlukene akuniki imodeli noma iyiphi indlela yokunikeza ilungelo lokunemba kweqiniso ngaphezu kokuhambisana komongo. Isignali yokuqeqesha ayikwazi ukuhlukanisa phakathi kwabo. Imodeli ifunda ukukhuluma kahle, okuyinto ephawulekayo. Inkinga iwukuthi ukuqephuza nokunemba ngokuvamile kuyahambisana. Uma bengakwenzi, ukushelela kuyaphumelela. Kuyi-a ukuxazulula ukungqubuzana indlela ekhiqiza umphumela ongalungile. Ijiyomethri ikubonisa ngesikhathi leso sinqumo senzeka ngaso.
Ukuze siphendule umbuzo oyimbangela – yiziphi izifunda ezithile ezisebenzisa ukucindezelwa, nokuthi zingashintshwa yini – sidinga ukucushwa kokuvula esikalini, ukuhlaziywa kwezinga lesifunda, kanye nokuhlolwa kokungenela okuyimbangela okudlula ubufakazi bokuhlobana obunikezwa yileli phepha. Leso isinyathelo esilandelayo. Amaqembu amaningana asebenza ngakho.
Ukuthi impendulo yalowo mbuzo oyimbangela ingasivumela yini ukuthi silungise ukubona izinto ezingekho ngaphakathi kwepharadigm yezakhiwo zamanje kuyindaba ehlukile. Umbono wami ukuthi ngeke – hhayi ngokuyisisekelo. Singakwazi ukucindezela ukucindezelwa. Singangeza isendlalelo sokuqapha esibamba ukugoqa kuka-κ ngaphambi kokuthi sifinyelele ekuphumeni. Singakwazi ukushuna kahle ezizindeni lapho ukungqubuzana kushube kakhulu. Lokhu ukuthuthuka kwangempela. Kodwa ukungezwani okuyisisekelo phakathi isibikezelo somongo futhi isisekelo esiyiqiniso ayisuki kuze kube yilapho imodeli inezethulo zomhlaba ezingasuselwe ekuhlanganeni kwethokheni. Lokho kudinga i-architecture ehlukile.
Kungani Lo msebenzi Ubalulekile
Ingqalasizinda ebonisa ngokunembile izindlela zokuhluleka kwama-LLM amanje iyisinyathelo esidingekayo sokudlulela kwangcono. Asikwazi ukuklama i-architecture ezolandela ngaphandle kokuqonda, ngokuningiliziwe, lokho owandulelayo empeleni wenzani ngaphakathi. Lo msebenzi usitshela okuthile okuqondile:
- Kuma-LLM e-autoregressive (i-transformers architecture), i-geometry yokucubungula okuyiqiniso nokungalungile iyahlukana ngokujikeleza, hhayi ngobukhulu;
- ukwehlukana kuyasebenza esikhundleni sokungenzi lutho;
- ukujula kokucindezelwa kufakwe isango lokuklama, hhayi nje umsebenzi wesikali;
- isiginesha yejiyomethri idluliswa kuzo zonke izizinda ngokuwohloka okuhlelekile kodwa okunemingcele.
Ijiyomethri ayiqambi amanga. Esikhetha ukukwenza ngalo ngumbuzo ohlukile.
Code, data, and related papers will be available at cert-framework.com soon.Ukufunda okunconyiwe
- Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, kanye Shan Carter. 2020. Sondeza isithombe: Isingeniso samasekhethi. I-Distill, 5(3):e00024–001.
- Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion Kamamode, Tom Conerly, Jackson Kernion, Dawn Drain UBrown, uJack Clark, uJared Kaplan, uSam McCandlish, noChris Olah. 2021. Uhlaka lwezibalo lwamasekethe ama-transformer. Transformer Circuits Thread.
- Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Mditya Ramesh, Daniel, Rewon Zimesh. Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, noDario Amodei. 2020. Amamodeli olimi angabafundi abathwebulayo abambalwa. Ku-Advances ku-Neural Information Processing Systems 33: Inkomfa Yaminyaka Yonke Yezinhlelo Zokucubungula Ulwazi Lwe-Neural 2020, NeurIPS 2020, Disemba 6-12, 2020, ebonakalayo.
- Bereska, L., & Gavves, E. (2024). Ukutolika kwemishini yokuphepha kwe-AI – isibuyekezo. I-arXiv preprint arXiv:2404.14082.
- U-Guillaume Alain no-Yoshua Bengio. Ukuqonda izendlalelo ezimaphakathi kusetshenziswa ama-linear classifier probes. I-ICLR, 2016.
