Buyela Ezindabeni
UkuqambaAI Understanding ukwaziswa

Iphepha liphakamisa indlela esebenza kahle kakhulu yokuklama ukuhlolwa kwedatha yokuqeqeshwa kwe-LLM

Ukuhlanganiswa kwedatha yemodeli yolimi olusha oluphrintiwe ngaphambili njengokuhlolwa kwezibalo, kubika ukuthi ukugijima kommeleli okukhethwe ngokucophelela kungase kubuyise izinga lengxube ngemva kokugijima okumbalwa okungaba ngu-25% esilinganisweni esilinganiselwe.

5 min readRead the primary source
Source-page capture accompanying Paper proposes a more efficient way to design LLM training-data experiments
Idokhumenti yomthombo oyinhlokoUmthombo urekhodiwe
Umshicileli
arxiv.org
Isixhumanisi somthombo
arxiv.orghttps://arxiv.org/abs/2608.23922
Uhlobo lomthombo
Idokhumenti eyisisekelo — isimemezelo esisemthethweni, iphepha, ukugcwalisa, noma ikhasi lomuntu wokuqala esilifunda ngokuqondile.
UmongoQonda lokhu ngemizuzwana engama-60

Qala lapha

Imigomo ebalulekile

Imodeli Yolimi Olukhulu (LLM)
Imodeli yolimi eqeqeshwe ku-massive text corpora ukuze ikhiqize futhi ihlaziye umbhalo.
Ukuziqeqesha
Ukuqeqeshwa kokuqala kwemodeli yezinga elikhulu kudatha ebanzi ngaphambi kokujwayela umfula.
Ipayipi
Ukugeleza komsebenzi oku-oda kokucutshungulwa kwangaphambili, izinyathelo zemodeli, nezigaba zokucubungula ngemuva.
ZihloleImibuzo Ecacisiwe yamamodeli e-AI

Kwenzekeni

Abacwaningi u-Yicheng Mao no-Hongru Du bahlongoza ukuphatha ukwabiwa kwedatha yokuqeqeshwa kuzo zonke izizinda njengokuhlolwa kwengxube yakudala. Uhlaka lwabo lumodela ukuthi izilinganiso zesizinda zithinta kanjani ukulahlekelwa kokuqinisekisa futhi zisebenzisa izindlela zokuklama zezibalo ukuze ukhethe ukuthi yikuphi ukuqeqeshwa kommeleli okusebenza ukuze kwenziwe.

I-preprint, ethunyelwe ku-arXiv ngo-Aug. 24, ichaza ukuxuba idatha njengenkinga yokuklama: lapho inani eliphelele lamathokheni okuqeqesha lilungisiwe, ochwepheshe kufanele banqume ukuthi yisiphi isabelo okufanele sivele esizindeni ngasinye. Ababhali baphikisa ngokuthi ukuhamba komsebenzi okususelwe kummeleli okukhona kakade kunesakhiwo sokuhlolwa kwengxube. Kuleyo ncazelo, izizinda ziyizingxenye, amasheya amathokheni alinganisa, amamodeli amancane agijima amaphuzu okuhlola, futhi ukulahlekelwa kokuqinisekisa kuyimpendulo elinganisiwe. Isiphakamiso esimaphakathi sephepha siwukukhetha lawo maphuzu okuhlola ngamabomu esikhundleni sokuphatha izingxube zommeleli njengeqoqo lamaresiphi ekhandidethi elikhethelwe ikakhulukazi ukubikezela. Uhlaka lugcina kugxilwe ekutheni ungafunda kanjani ezivivinyweni ezitholakalayo zommeleli kuyilapho inani eliphelele lethokheni lihlala lingashintshi. Ngakho-ke iphathelene nokukhethwa kwamaphuzu okuhlola kanye nokuchazwa kwezimpendulo zabo zokuqinisekisa-ukulahlekelwa, hhayi ushintsho enanini ledatha etholakalayo yokuqeqeshwa.

Ababhali bakha imodeli ye-Scheffé yesibili elandelanayo yokuphendula. Ngamagama asobala, imodeli ilinganisela kokubili umnikelo ngamunye wesizinda sedatha kanye nomthelela wokuyihlanganisa nesinye isizinda. I-abstract ithi ukuhlaziya kuthola ukuthi inani lesizinda lihlobene kakhulu: ezinye izizinda ezibonakala zibuthakathaka uma zicatshangelwa ngemiphumela yokwengeza ziba zinhle kuzinhlanganisela ezithile, ikakhulukazi uma zibhangqwe nombhalo osuselwe kuwebhu. Lesi isimangalo esimayelana nokusebenzisana ekuhlaziyweni kocwaningo, hhayi ukutholwa okuvamile kokuthi yonke imithombo yedatha izothuthukisa i-LLM uma ixutshaniswa nombhalo wewebhu.

I-RegMix isetshenziswa njengesifundo sephepha elisebenzayo. Ababhali bathi imodeli yezibalo ezimbalwa igcina izilinganiso zengxube kuzo zonke izikali zemodeli futhi ihlala inokuncintisana nesibikezelo sokufunda ngomshini esivumelana nezimo, kuyilapho ihlinzeka nokuhlukaniswa okucacile kwemiphumela yokungeza nokusebenzisana. Ekulingiseni okulinganiselwe kuzimpendulo zokuqeqeshwa kommeleli, imiklamo yabo ye-I-optimal eqinile ibuyisela ukuhleleka okuhambisanayo kwezingxube ngemva kokukhishwa okungaba ngu-25% kokwenziwa kommeleli kwasekuqaleni. Umthombo awunikezi inani lokugijima, osayizi bemodeli, amasethi edatha, amanani okulahlekelwa okuqinisekisa noma ukuqhathaniswa kwezindleko zokubala zangempela ku-abstract.

Imininingwane yomthombo: arxiv.org ↗

Kungani kubalulekile

Le ndlela ingasiza abacwaningi ukuthi bafunde ukwakheka kwedatha yokuqeqeshwa ngokuhlolwa kwamamodeli amancane kuyilapho benza ukusebenzisana kwesizinda kubonakale kakhudlwana. Umphumela obikiwe osebenza kahle usuka ekulingiseni okulinganiselwe ukuze kube nezimpendulo zokuqeqeshwa kommeleli, ngakho inani layo elisebenzayo lincike ekutheni idlulisela kwamanye amasethi edatha, amamodeli nezilungiselelo zokuqeqesha.

Ukwakhiwa kwedatha yokuqeqesha kuwukukhetha okuyisisekelo ekwakheni amamodeli amakhulu olimi, kodwa ukuhlola izingxube eziningi ezingaba khona kungadinga ukuqeqeshwa okuphindaphindiwe kommeleli. Uhlaka oluhlongozwayo lubhekana nenqubo yokuhlola ngokwalo: luzama ukuhlonza ukuthi yiziphi izingxube ezifundisa kakhulu ngaphambi kokuba kwenziwe konke ukugijima kommeleli wekhandidethi. Uma umphumela wokulingisa okubikiwe usasebenza, abacwaningi bangachitha izivivinyo ezimbalwa befunda ukuthi yiziphi izilinganiso ezisebenza kangcono, noma basebenzise isabelomali sokuhlola esifanayo ukuze bahlole ezinye izindlela.

Ukugcizelela kwephepha ekusebenzelaneni ngababili nakho kuyahambisana. Isizinda esibukeka singakhangi sodwa singanikela inani lapho sihlanganiswa nesinye isizinda, futhi ingxube ebonakala ithembisa kusuka kuzikolo ezihlukene zesizinda ingase yenze ngendlela ehlukile uma izingxenye sezihlanganisiwe. Indlela edalula lobo budlelwano ingenza izinqumo zokuhlanganisa idatha kube lula ukuzihlola nokuzichaza kunesibikezelo esikhiqiza kuphela izinga lokugcina. Umthombo wethula lokhu kuhunyushwa njengenzuzo yemodeli ye-Scheffé encane.

Umphumela ubaluleke kakhulu njengomnikelo wendlela yokusebenza, hhayi njengobufakazi bokuthi ingxube ethile yokuqeqeshwa manje isisungulwe njengengcono kakhulu. Iphepha alimemezeli imodeli yolimi olusha, isethi yedatha noma umkhiqizo. Inikeza indlela yokuhlela izivivinyo mayelana nesabelomali sethokheni engashintshi kanye nempendulo yokuqinisekisa-ukulahlekelwa. Ngakho-ke ukubaluleka kwayo okungokoqobo kuncike kukhwalithi yamamodeli ommeleli, ukumelela kokulahlekelwa kokuqinisekisa okulinganiselwe kanye nezinga lapho amazinga ahlala ezinzile lapho ukuqeqeshwa kwenyuswa.

Interactive Mechanism

I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela

Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.

Model Parameter Size:8B Parameters
VRAM Required5.5 GBGPU memory footprint
Target HardwareMacBook / Single GPUDeployment tier
Privacy100% Air-GappedLocal device capability
Core takeaway: Small, quantized models (3B–8B) now run directly inside smartphones and laptops with complete data privacy, while mammoth 400B+ models remain the domain of datacenter clusters.
I-Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

Ongakubuka ngokulandelayo

Imibuzo ebalulekile ukuthi ingabe indlela ithuthukisa izinqumo ekuqeqeshweni kwangaphambilini okukhulu, ukuthi izilinganiso zayo zithembeke kangakanani ngaphandle kocwaningo lwendaba lwe-RegMix, nokuthi ingabe ukonga kuyaqhubeka yini lapho ukusebenza kommeleli kuhluka kakhulu kunemodeli yokugcina. Umthombo awubiki ukusetshenziswa kwenani elikhulu, ukuphindaphinda okuzimele noma ukonga okuphelele kwezindleko zokuqeqesha.

Inkinga yokuqala okufanele ibukwe ukuqinisekiswa kwangaphandle. I-abstract ibika ucwaningo lwecala lwe-RegMix olunamandla kanye nokulingisa okulinganiselwe ukuze kuqashwe izimpendulo zokuqeqeshwa kommeleli, kodwa ayisho ukuthi imiklamo ehlongozwayo yahlolwa ekugijimeni okusha kwesilinganiso esikhulu sokuqeqesha. Umsebenzi wesikhathi esizayo uzodinga ukubonisa ukuthi ukukhetha izingxube ezimbalwa zommeleli kuholela ezinqumweni ezifanayo lapho imodeli yokugcina, isabelomali sethokheni, ipayipi lokucubungula idatha noma izinguquko zokuhlola.

Udaba lwesibili wububanzi bokwehliswa okufunwayo ngo-25%. Amagama abonisa ukuthi umphumela uvela ekususeni cishe ingxenye eyodwa kwezine yommeleli wangempela osebenza ngokulingisa ngenkathi ithola uku-oda kwengxube efanele. Ayisunguli ukwehla okungama-25% kwezindleko zokuqeqesha, isikhathi sewashi lasodongeni noma amandla. Ukonga kungahluka ngenani lezizinda, umumo wendawo yokuphendula kanye nenani lomsindo ezilinganisweni zokuqinisekisa.

Umthombo uphinde ushiye imininingwane ebalulekile yokusebenza ingaziwa. I-abstract ayibiki idizayini yokuhlola ephelele, izizinda ezifakwe ku-RegMix, osayizi bamamodeli abamba iqhaza, umehluko ophelele wokuqinisekisa-ukulahlekelwa noma ukuthi indlela ikhetha kaningi kangakanani ingxube ephansi. Ayichazi ukuphindaphinda okuzimele noma izikhawu zokungaqiniseki. Leyo mininingwane izonquma ukuthi uhlaka luqine ngokwanele yini ezinqumweni zokuqeqesha ezibizayo noma ngokuyinhloko luyilensi yokuhlaziya ewusizo yokuhlolwa kocwaningo.

Imihlahlandlela ehlobene nemibuzo

Amamodeli e-AI AchaziweUkuqeqeshwa kwe-AIAma-TransformersYini i-AI?Hlola okwaziyo — zama imibuzo ye-AI yamahhalaBheka igama le-AI kuhlu lwethu lwamagamaLandela i-tracker yokukhishwa kwemodeli ye-AI
Uthole lokhu kuwusizo?