ZDTaichu5.0-9B Uyasebenza

I-ZDTaichu5.0-9B iyimodeli entsha evulekile, ejwayelekile futhi enemodi eziningi ye-Zidong Taichu, eqondiswe emhlabeni obonakalayo nasezimweni zobuhlakani obugxile emzimbeni. Kwamabhentshimakhi omhlaba ayisishiyagalolunye omhlaba, ithole indawo yokuqala emhlabeni ezivivinyweni eziyisi-8 phakathi kwamamodeli anenani elifanayo lamapharamitha; i-MindCube-tiny ithole u-78.27, idlula kakhulu amamodeli esigaba sayo. Amakhono ajwayelekile okuqonda izithombe nombhalo, i-OCR, izibalo nekhodi nawo ahlala esigabeni esiphambili. Imodeli ingakwazi ukuzivumelanisa nokucabanga okuyindilinga; isebenzisa indlela yokulawula i-entropy ukuze yengeze ukubala okuyindilinga kuma-Token angaqinisekile. Imodeli isiqinisekisiwe emisebenzini yangempela yocwaningo kanye neyokukhiqiza ezimbonini, futhi izisindo kanye nohlelo lokuqeqesha kuvulwe ngokuphelele.

Ukuqonda nokubikezela indawo: Ithole indawo yokuqala emhlabeni ezivivinyweni eziyisi-8 phakathi kwamamodeli anenani elifanayo lamapharamitha kumabhentshimakhi omhlaba ayisishiyagalolunye, futhi inekhono ekucabangeni okuyinkimbinkimbi kwendawo njengokushintsha kwezinhlelo zokubhekisela nokunquma izikhombisi-ndlela.
Amakhono ajwayelekile emodi eziningi: Ukuqonda izithombe nombhalo, i-OCR, izibalo nekhodi kuyalingana; imiphumela ye-AI2D, OCRBench, AIME2026 nezinye ihlala esigabeni esiphambili.
Ukuhlelwa kwemisebenzi egxile emzimbeni: Njengobuchopho obusezingeni eliphezulu berobhothi, iqonda indawo ezungezile, ikhumbule ubunikazi bezinto futhi ihlukanise izinyathelo zomsebenzi.
Usizo ezivivinyweni zocwaningo: Ihlanganisa ukubala kokulingisa nezivivinyo zaselabhorethri, ibize ngokuzenzakalelayo i-Python/SciPy ukuze iqedele ukuhlaziya kwesayensi futhi ikhiqize imibiko.
Ukwenza imisebenzi yokukhiqiza ezimbonini: Ihlanganisa ukusatshalaliswa kwezinto ezilungiselelwe, ukulayisha nokwehlisa imishini, nokuhlelwa kweziteshi zokusebenza, ibhekane nezinkinga zangempela njengezindawo zokugcina eziminyene, ukuvaleka kokubuka nezimfuneko zangaphambili.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI avulekile
Ukucabanga okuyindilinga okuzivumelanisayo: Lapho ikhiqiza i-Token ngayinye, imodeli isebenzisa indlela yokulawula i-entropy ukuhlola ukungaqiniseki kwesibikezelo. Uma ukusatshalaliswa kwamathuba kusabalele kakhulu, iphinda ibale izingxenye ezithile zezendlalelo kuleyo ndawo, ilungise izethulo zangaphakathi emijikelezweni ilandelana ngaphambi kokukhipha umphumela. Ezindaweni ezinokuqiniseka okuphezulu ikhipha umphumela ngokuqondile, igxilise amandla okubala lapho “ingaqiniseki khona”, futhi ayincikile kumathuluzi angaphandle.
Izindlela ezintathu zokuzinza: Ukuvimbela ukuzulazula kwezethulo okubangelwa ukubala okuyindilinga, kufakwa ukulawulwa kokubuyekezwa okunokudambisa ukuze kulawulwe ubukhulu bokulungiswa komjikelezo ngamunye, kusetshenziswa isethulo sokuqala njengesithenjwa. Kusetshenziswa izinqubo ezimbili zokumisa, i-KL divergence kanye nensalela yesimo esifihlekile, ukuze kunqunywe ukuthi ukuphindaphinda sekuguqukile yini. Izimo eziphakathi ziyagcinwa, okuvumela ukukhethwa kwesethulo esinobungozi obuphansi emiphumeleni eminingi kanye nokubuyela emuva lapho kudingeka.
Ikharikhulamu yedatha eqhubekayo: Ukuqeqesha kwenziwa ngezigaba. Ukuqeqeshwa kwangaphambili kusebenzisa idatha yezithombe nombhalo, amawebhusayithi, amavidiyo anemibono eminingi nokulingisa kwe-3D, ngemva kokususa okuphindaphindiwe nokuhlunga ikhwalithi. Esigabeni sokulungisa ngokugadwa kufakwa imibuzo nezimpendulo zebhizinisi langempela, izibonelo zendawo nezemisebenzi egxile emzimbeni, kanye nemizila yokubizwa kwamathuluzi e-Agent; kuqinisekiswa nokuvumelana kwezithombe, ubudlelwano bezinto nemikhawulo yokusebenza kumasampula. Ekugcineni, kusetshenziswa amasampula okucabanga okude ukwenza i-annealing, kuhlanganiswe nokufunda okuqinisa kwe-GRPO, imivuzo isekelwe ezimpawini ezingaqinisekiswa ngokuzenzakalelayo njengokunemba kwempendulo, ukushaya izixhumanisi nokuhambisana nefomethi.
Umjikelezo ophelele wemisebenzi emide: Ukuphindaphinda kwangaphakathi kwemodeli kuxazulula kuphela ukucabanga kwesinyathelo esisodwa; ukuqhubeka komsebenzi kuncike empendulweni yangaphandle. Ngemva kwesinyathelo ngasinye, isimo sendawo siqinisekiswa ngesithombe esisha, kwakheke umjikelezo oqhubekayo “wokubona—ukuqonda—ukwenza—is feedback”.

Ukuthola izisindo: Landa izisindo ezivulekile ku-Huggingface noma ku-GitHub.
Ukuyifaka endaweni: Layisha imodeli ku-GPU yezinga labathengi noma yezobuchwepheshe enekhadi elilodwa; inamapharamitha angu-9B futhi izidingo zememori ye-GPU ziphansi uma kuqhathaniswa.
Ukuxhumana okunemodi eziningi: Faka isithombe kanye nomyalelo wombhalo; imodeli ingakhipha imiphumela yokucabanga kwendawo njengokunquma isikhombisi-ndlela, izixhumanisi zento nezindawo ezingabekwa kuzo izinto.
Ukuyixhuma ohlelweni lwerobhothi: Yisebenzise njengemojuli yokuhlela esezingeni eliphezulu, ixhunywe ezinhlelweni zokulawula irobhothi ezisezingeni eliphansi zokubamba, ukuzulazula nokunye.
Ukuqeqeshwa okwengeziwe: Sebenzisa uhlelo oluvuliwe lokukhiqiza idatha nokuqeqesha oluhambisana nezisindo, bese uqhubeka nokulungisa imodeli ngedatha yezimo zakho.

Ikhono lendawo lihamba phambili esigabeni salo: Iwine kuzo zonke izivivinyo eziyisishiyagalolunye zomhlaba zendawo, yathola indawo yokuqala emhlabeni ezivivinyweni ezi-8 zeqembu elijwayelekile elinamapharamitha afanayo; i-MindCube-tiny ihola amamodeli esigaba sayo ngamaphuzu angaphezu kuka-15.
Amakhono ajwayelekile awanciphi: Ngesikhathi ikhono lendawo nomzimba lithuthuka kakhulu, ukuqonda izithombe nombhalo, i-OCR, izibalo nekhodi kusalokhu kusigaba esiphambili.
Ukucabanga okuyindilinga okuzivumelanisayo: Indlela yokulawula i-entropy yengeza ukubala kuma-Token “angaqinisekile” kuphela, okwenza ukucabanga kunembe futhi izindleko ezimaphakathi zilawuleke.
Iqinisekiswe ngamarobhothi angempela: Isivele isetshenziswe ngempumelelo emisebenzini yangempela yocwaningo, njengokuphatha amashubhu okuhlola nokudlulisa uketshezi nge-dropper, kanye neyokukhiqiza, njengokulayisha, ukwehlisa nokuhlela iziteshi zokusebenza.
Iqhathaniswa namamodeli amakhulu: Emabhentshimakhi amaningi ezindawo, idlula i-Gemini 3 Pro ngaphezu kuka-7%, kanti izindleko zokuyisebenzisa ngenxa yenani elincane lamapharamitha ziphansi kakhulu kunamamodeli avaliwe.
Ivuleke ngokuphelele futhi iyalungiseka: Izisindo, ukukhiqizwa kwedatha kanye nezinhlelo zokuqeqesha kuvulwe ngokuphelele, ngakho abathuthukisi bangasebenzisa idatha yabo ukuqhubeka nokulungisa imodeli ukuze ihambisane nezindawo zabo.

Iwebhusayithi yephrojekthi: https://taichu-ai.github.io/ZDTaichu5.0-9B
Indawo yokugcina ye-GitHub: https://github.com/Taichu-AI/ZDTaichu5.0-9B
Indawo yemodeli ye-HuggingFace: https://huggingface.co/TaichuAI/ZDTaichu5.0-9B

Ukukhiqiza ezimbonini: Ukusatshalaliswa kwezinto ezilungiselelwe, ukuhlola ukulayisha nokwehlisa imishini, nokuhlelwa kweziteshi zokusebenza; ibona izinto ezihlosiwe emashalofini aminyene, inqume indawo yokubamba, bese iqinisekisa ukuthi ukubekwa kwenziwe kahle.
Ukuzenzakalela kwezivivinyo zocwaningo: Isebenza njengobuchopho obuphezulu bokuhlela ezisekelweni zezivivinyo ezizenzakalelayo, yenza imisebenzi ehlukahlukene njengokubeka amashubhu okuhlola ngokulandelana nokudlulisa uketshezi ngama-dropper, kuyilapho ilandelela ubunikazi bamasampula, izindawo zawo kanye nenqubekelaphambili yomsebenzi.
Usizo ekubaleni kwesayensi: Ngemva kokwamukela umbuzo wesayensi ngolimi lwemvelo, ibiza ngokuzenzakalelayo i-Python/SciPy ukuze ibale izixazululo ezihlaziyayo nezinombolo, iqhathanise imiphumela futhi ikhiqize umbiko wokuhlaziya, ihlanganise ukulingisa nezivivinyo zaselabhorethri.
Ukuhlelwa okuphezulu kwamarobhothi: Njengobuchopho bobuhlakani obugxile emzimbeni, iqonda indawo ezungezile, ikhumbule izinto, ihlukanise izinyathelo futhi isebenzisane nezinhlelo zokulawula ukubamba nokuzulazula ezisezingeni eliphansi ukuze iqedele imisebenzi emide.
Ukugcinwa nokuthuthwa okuhlakaniphile: Iphatha ama-oda afana nokuthi “thatha ibhokisi lesibili emgqeni wesithathu”, ixhumanise izici, ihlele indawo futhi ithole okuqondiwe, ibhekane nezinkinga zangempela zokugcina izinto njengokuvaleka kokubuka noma into ephuma endaweni yokubuka.