OmniEdu Uyasebenza

I-OmniEdu iwuchungechunge lwamamodeli ezemfundo e-multimodal avulelwe umphakathi yiqembu le-DCAI lasePeking University ngokubambisana ne-University of Chinese Academy of Sciences kanye ne-Zhongguancun Academy, futhi inezilinganiso ezintathu: 27B, 9B ne-4B. I-OmniEdu igxile emandleni okufundisa: ingaphendula imibuzo yezifundo, inqume indawo yolwazi kukharikhulamu, ihlonze izimbangela zamaphutha abafundi, futhi ikhethe isenzo sokufundisa esifanele. Iqembu likhethe cishe amasampula angu-69,000 ekuqeqesheni lisuka kudatha yabangenele ukhetho eyizigidi ezingu-1.34, okwenza imodeli ifinyelele ezingeni le-SOTA emisebenzini yezifundo eziningi, yokuxilonga neyokuqondisa. Imodeli encane engu-4B ingasetshenziswa kuseva evamile, ngaleyo ndlela yehlise umkhawulo wokusetshenziswa ezikoleni.

Ukuxazulula imibuzo yezifundo: Kuthinta izibalo, isayensi, ukuqonda okufundwayo nokubhala, inikeze izimpendulo ezifanele futhi iqinisekise ukuthi incazelo iyahambisana nemibandela yombuzo kanye nenqubo yokucabanga.
Ukuqonda ikharikhulamu: Ihlonza ulwazi oluhambisana nombuzo, ibanga, ubunzima nobudlelwano bezimfuneko zangaphambili, bese inquma okudingeka kufundwe kuqala ngaphambi kokuqhubeka.
Ukucabanga kokuxilonga: Ngokusekelwe enqubweni yokuphendula yomfundi nasezimpendulweni ezingalungile, ithola isinyathelo lapho kwenzeke khona iphutha, ihlaziye ukuphambuka ekuqondeni kanye nolwazi oludingekayo, kunokuthembela kuphela ekuqageleni okuvela embuzweni.
Isenzo sokufundisa: Ngokwesidingo somsebenzi, ikhetha indlela efanele yokuphendula, ehlanganisa ukunikeza izeluleko, imibuzo eqondisayo, ukubuyekeza imiqondo eyisisekelo, ukukhomba iphutha noma ukunikeza incazelo ephelele.

Landela i-WeChat uphendule ngokuthi “开源” ukuze ujoyine iqembu lokuxoxa ngamaphrojekthi e-AI avulekile
Ukwakhiwa kwedatha okuqhutshwa yizinjongo zomsebenzi: Idatha yokuqeqesha ihlanzwe, yahlolwa ngokomqondo, yahlolwa ngekhwalithi futhi yahlungwa ngokwehlukahlukana kusuka kumasampula angaba yizigidi ezingu-1.34, kwasala cishe amasampula angu-69,000. Isampula ngayinye yezemfundo ihlukaniswa ngamakhono amane futhi yabelwe owodwa wemiyalelo yemisebenzi engu-20, kuhlukaniswe phakathi kokuthi “isampula ifundisani” nokuthi “imodeli kufanele iphendule kanjani”, ukuze imodeli ikwazi ukushintsha ukuziphatha ngokomsebenzi.
Ukuhlolwa nokubhalwa kabusha ngokomqondo: Imodeli yolimi ihlola amasampula ngokwemigomo yomsebenzi, ihlole ukunemba kokuxazulula, ubudlelwano bekharikhulamu, isisekelo sokuxilonga kanye nekhwalithi yokweseka. Amasampula angafanele abhalwa kabusha ngezimpendulo zokuqondisa bese ehlolwa futhi. Kuphinde kuqinisekiswe ukuhambisana kombhalo nezithombe kumasampula e-multimodal, kugcinwe imininingwane yezimvume zemithombo futhi kukhishwe imithombo engahambisani nemithetho.
Ukususa ukungcoliswa nokuhlanganisa isakhiwo: Ifomethi yemiyalezo ihlanganiswa, kususwe amasampula aphindaphindiwe, angenazo izinkambu noma anesakhiwo esonakele, kuqinisekiswe izimpendulo zemisebenzi engaqinisekiswa, bese kususwa okuphindaphindiwe phakathi kwedatha yokuqeqesha nesethi yokuhlola ukuze kugwenywe imiphumela ephambukayo.
Ukulinganisa isabelo sokuqeqesha ngebhajethi yama-token: Izingxoxo ezinde zokuqondisa ziqukethe ama-token amaningi okuqondisa kunezimpendulo ezimfushane. Uma kusampulwa ngenani lamasampula, izingxoxo ezinde zingaba nesabelo esikhulu kakhulu. Ibhajethi yama-token yabelwa ngamabhakede emisebenzi, kuphakanyiswe nemingcele yezici ezibalulekile ukuze kukhethwe idatha ehlukahlukene, kulawulwe inani langempela lokuqondisa elitholwa yizinhlobo zemisebenzi yokufundisa.

Indlela yokuqala: Ukusebenzisa endaweni (Transformers) Faka okuncikile: sebenzisa pip install -U torch torchvision transformers accelerate safetensors pillow (izisindo ze-FP8 zidinga nokufakwa kwe-compressed-tensors).
Layisha imodeli: Sebenzisa i-AutoProcessor kanye ne-Qwen3_5ForConditionalGeneration.from_pretrained ukuze ulayishe i-ID yemodeli efanele (njenge lhpku20010120/Omni-Edu-4B; ku-9B/27B shintsha i-ID kuphela).
Yakha okokufaka: Hlela imibuzo yomfundi, izimpendulo noma izithombe zibe ama-messages usebenzisa i-chat template; emibuzweni enesithombe faka into ethi {"type": "image"}.
Yenza impendulo: Shayela i-generate (qinisekisa ukuthi enable_thinking=False), bese usebenzisa i-batch_decode ukukhipha imiphumela.
Indlela yesibili: Ukusebenzisa isevisi (vLLM, kuyanconywa) Qalisa isevisi: sebenzisa vllm serve lhpku20010120/Omni-Edu-4B --max-model-len 32768 --reasoning-parser qwen3 --default-chat-template-kwargs '{"enable_thinking": false}' (enguqulweni ye-FP8 sebenzisa i-Omni-Edu-4B-FP8; kumakhadi amaningi lungisa --tensor-parallel-size).
Thumela isicelo: Sebenzisa i-curl ukuze ufinyelele ku-http://localhost:8000/v1/chat/completions, ubuze usebenzisa ifomethi ehambisana ne-OpenAI.
Indlela yesithathu: Sebenzisa iphakheji esemthethweni Faka i-SDK: sebenzisa pip install omniedu (ekuhloleni faka omniedu[eval]).
Shayela ingxoxo: Ngemva kwe-OmniEdu.from_pretrained("4B", base_url=ikheli_lesevisi), sebenzisa model.chat(text=...) ukubuza.
Ukuqhathanisa ukuhlola: Sebenzisa omniedu eval --list ukuze ubuke izisekelo zokuhlola ezingu-12 zezemfundo futhi ukhiqize umbiko.

Ukushintsha ukuziphatha okuqhutshwa yimiyalelo yomsebenzi: Imiyalelo yesistimu engu-20 yenza imodeli efanayo ikwazi ukuhlukanisa phakathi kokuxazulula imibuzo ngesinyathelo ngesinyathelo, ukuxilonga amaphutha, imibuzo ye-Socratic kanye nezingxoxo zokuqondisa ngokomsebenzi.
Idatha yokuqeqesha yekhwalithi ephezulu eklanywe ngokucophelela: Kusampula ezingaba yizigidi ezingu-1.34, kukhethwe cishe ezingu-70,000 ngemva kokuhlolwa ngokomqondo, ukususa ukungcoliswa kanye nokulinganisa isabelo sama-token, kwacaciswa “okufundwayo” kanye “nendlela yokufundisa”.
Ukusebenza kwezinga le-SOTA: I-27B inqobe izikhathi ezingu-51 kwezingama-60 zokuqhathanisa namamodeli angu-5 aphambili, ahlanganisa i-GPT-5.4 ne-Claude-Opus-5, yahola ezinkomba eziningi zemfundo.
Ukuthuthuka okuphawulekayo ekufundiseni okusekelayo: Izinga lokunqoba lesikafula ku-MathTutorBench lenyuke lisuka ku-14% laya ku-78.74%, okufakazela ukuthi ukuziphatha kokufundisa kungaqeqeshwa ngempumelelo.
Ukuzivumelanisa nezimo ezahlukene ngamazinga amaningi: Ihlanganisa i-4B, 9B ne-27B, kanti inguqulo ye-FP8 inciphisa umthamo wokulanda ngo-37%–44%; imodeli encane ingasetshenziswa kuseva evamile yesikole.

Iwebhusayithi esemthethweni yephrojekthi: https://haolpku.github.io/Omni-Edu/
Ikhosombe le-GitHub: https://github.com/haolpku/Omni-Edu
Ikhosombe lamamodeli le-HuggingFace: https://huggingface.co/datasets/lhpku20010120/Omni-Edu
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2609.23088

Ukuqondisa ukufunda okuqondene nomfundi: Ngesikhathi umfundi exazulula umbuzo, imodeli iqala ngokuthola lapho ebambeke khona, bese isebenzisa izeluleko nemibuzo elandelayo ukuze imqondise ekucabangeni ngokuzimela, okulungele ukuzifundela.
Ukuhlaziya nokuxilonga imibuzo engalungile: Uthisha ufaka inqubo yokuphendula yomfundi kumodeli, yona ihlaziye ngokuzenzakalelayo imbangela yephutha nezikhala zolwazi ezihambisanayo, ikhiqize iziphakamiso zokulungisa eziqondene nesidingo futhi inciphise umthwalo wokumaka.
Ukukhiqiza okuqukethwe okuhambisana nekharikhulamu: Uthisha ukhiqiza imibuzo yokuzilolonga, umsebenzi ohlukaniswe ngamazinga noma izinto zokuchaza ngokwezindinganiso zekharikhulamu nebanga elithile, bese imodeli imaka ngokuzenzakalelayo ubulungu bolwazi kanye nobudlelwano bezimfuneko zangaphambili.
Umsizi wokuxhumana ekilasini (ukufundisa kwe-Socratic): Imodeli idlala indima yomsizi wokufundisa, iqhubekisele phambili ukucabanga kwabafundi ngemibuzo eqondisayo, ilungise amasu okufundisa ngokusekelwe ebufakazini futhi isekele ukufundisa okuhlukaniswe ngamazinga.