Unlimited OCR 2 Uyasebenza

I-Unlimited OCR 2 iyimodeli yokuhlaziya amadokhumenti esekupheleni kuya ekupheleni, ekhishwe yi-Baidu Wenxin njengomthombo ovulekile. Isekela ukuhlaziywa kwamadokhumenti amade ananoma yiliphi inani lamakhasi, futhi ingacubungula izinkulungwane zamakhasi ezincwadi ngesikhathi esisodwa igcine ukunemba nokusebenza kahle. I-Unlimited OCR 2 isebenzisa indlela ye-sparse reference sliding-window attention kanye nendlela ye-infinite decoding, yehlise izindleko zokubala nezememori kumadokhumenti amade. Ifinyelela ezingeni le-SOTA kuma-benchmark afana ne-MPDocBench ne-MarathonDocBench, iqhubekisele i-OCR enkathini ye-3.0.

Ukuhlaziya noma yiliphi inani lamakhasi: Yeqa umkhawulo wamakhasi wendabuko, isekele ukucutshungulwa kwamadokhumenti amade anezinkulungwane zamakhasi ngesikhathi esisodwa.
Ukuqonda okunokunemba okuphezulu kumadokhumenti amade: Igcina ukunemba okuphezulu kokuhlaziya nokusebenza kahle lapho ubude bedokhumenti bukhula kakhulu.
Ukuhlobanisa ulwazi phakathi kwamakhasi: Ingabuyela emibhalweni yangaphambilini futhi iqinisekise okuqukethwe kwamakhasi ahlukene, yehlise ukunqamuka komongo okubangelwa ukuhlanganiswa kwamakhasi amaningi.
Iyatholakala njengomthombo ovulekile: Ikhodi nezisindo zemodeli kuvulekile, futhi isiyatholakala ku-Baidu Intelligent Cloud naku-Baidu AI Open Platform.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI anomthombo ovulekile
SR-SWA sparse reference sliding-window attention: Isebenzisa umqondo wokuthi umuntu oguqula incwadi ugxila ekhasini lamanje futhi abuyele emibhalweni yangaphambilini lapho kudingeka. I-Page Gate inciphisa indawo yokubhekisela ebonakalayo yombuzo ngamunye isuka kuwo wonke amadokhumenti iye emakhasini ahlobene, iphinde yakhe ubudlelwano bendawo phakathi kwamakhasi. Ama-Visual Token Buckets anezikolo eziphezulu ayafakwa ukuze abambe ukuncika okubukwayo okude, kugwenywe ukubala okuphindaphindwayo okuphezulu kwe-Full Attention kuwo wonke ama-token edokhumenti.
Unlimited-Horizon Decoding: Ilungiselela inqubo yokukhipha impendulo ukuze ixazulule izingqinamba zokubala nememori kumadokhumenti amade kakhulu, ivumele imodeli ukuthi icubungule amakhasi amaningi ngaphandle kokubanjwa kanzima ubude bomongo, ngaleyo ndlela isekele umgomo wokuhlaziya “noma yiliphi inani lamakhasi”.
Ukumodela amadokhumenti amade kusukela ekuqaleni kuya ekupheleni: Ngokungafani ne-OCR yendabuko ehlonza ikhasi ngalinye bese ihlanganisa imiphumela, futhi ngokungafani ne-R-SWA yesizukulwane esedlule eyayinweba kuphela kumashumi ambalwa amakhasi alandelanayo, i-Unlimited OCR 2 ihlanganisa ukukhethwa kwamakhasi ahlukene, ukubuyela kumareferensi nokuqonda kwekhasi lamanje kuhlaka olulodwa lokuphela kuya ekupheleni, ithuthukise ukuvumelana nokuphelela kokuhlaziya amadokhumenti amade.
Ukulawula ukusebenza kahle nezindleko: Ijubane lokukhipha impendulo ngesinyathelo esisodwa le-SR-SWA liphakeme ngokuphindwe ka-7 kune-Full Attention, futhi liphakeme ngokuphindwe ka-5.6 kune-R-SWA yesizukulwane esedlule. Inzuzo esemqoka ivela ekusebenziseni ukubhekisela okungasindi esikhundleni sokunaka okuphelele komhlaba wonke.

Khetha indawo yokufinyelela: Landa ikhodi nezisindo ku-GitHub noma ku-Hugging Face ukuze ukufake endaweni yangakini.
Lungisa amadokhumenti: Hlela ama-PDF, izikena noma amadokhumenti amade, uqinisekise ukuthi amakhasi aphelele, acacile futhi abheke kahle.
Layisha noma layisha ngaphakathi: Endaweni ye-inthanethi layisha idokhumenti ngqo; ekufakweni kwasendaweni lungisa imvelo bese ulayisha izisindo zemodeli ye-Unlimited OCR 2.
Setha umsebenzi: Khetha ifomethi yokuphumayo, ububanzi bokuhlaziya kanye nezinketho zokubona amatafula nesakhiwo ngokwamakhasi noma ngamapharamitha e-API.
Qalisa ukuhlaziya: Thumela umsebenzi, ulinde imodeli iqede ukuqonda amakhasi ahlukene, ukukhipha okuqukethwe nokukhiqiza okuhlelekile.
Qinisekisa imiphumela: Hlola ngokukhethekile izihloko eziphakathi kwamakhasi, amatafula, amafomula, izinombolo zamakhasi, izingcaphuno nokuqhubeka kwezigaba.
Thumela futhi uhlanganise: Thumela umbhalo/Markdown/JSON, noma usebenzise i-API ukuze uyixhume ekucutshungulweni kwamaqoqo nasezinhlelweni zebhizinisi.

Ukuhlaziya noma yiliphi inani lamakhasi: Yeqa umkhawulo wamakhasi we-OCR yendabuko futhi ikwazi ukucubungula amadokhumenti amade anezinkulungwane zamakhasi.
Ukunemba okuzinzile kumadokhumenti amade: Ngisho nalapho inani lamakhasi likhula kakhulu, igcina ukunemba nokusebenza kahle okuphezulu kokuhlaziya.
Ukuqonda okungcono phakathi kwamakhasi: Ingabuyela emibhalweni yangaphambilini futhi ihlobanise ulwazi lwamakhasi ahlukene, yehlise ukunqamuka komongo okubangelwa ukuhlanganisa.
Ukusebenza kahle kokubala okuphezulu: I-SR-SWA isebenzisa ukubhekisela okungasindi esikhundleni sokunaka okuphelele, yehlise kakhulu ukubala okuphindaphindwayo.
Izindleko zememori eziphansi: Ihlanganiswe nendlela ye-infinite decoding, yehlisa ingcindezi yememori lapho kuhlaziywa amadokhumenti amade kakhulu.
Ukusebenza okuvelele kuma-benchmark: Umbhalo uthi ifinyelela ezingeni le-SOTA kuma-benchmark amadokhumenti amade afana ne-MPDocBench ne-MarathonDocBench.

Indawo yokugcina ye-GitHub：https://github.com/baidu/Unlimited-OCR
Indawo yamamodeli ye-HuggingFace：https://huggingface.co/spaces/baidu/Unlimited-OCR-2

Ukuhlaziya izincwadi ezinde nemibhalo: Yenza amanoveli aphelele, izincwadi zokufunda nemibhalo yakudala ibe yidijithali ngokushesha, igcine umongo wamakhasi ahlukene.
Ukuhlaziya imibiko yocwaningo nemibiko yezimali: Funda imibiko yocwaningo noma yonyaka enamakhulu amakhasi ngesikhathi esisodwa, ukhiphe imibono, izinkomba nezingozi.
Ukubuyekezwa kwezomthetho nokuhambisana nemithetho: Cubungula amaqoqo emithetho, amafayela ezinkontileka namadokhumenti amacala, usekele ukuhlobanisa izigaba eziphakathi kwamakhasi.
Ukwenza amadokhumenti namarisidi abe yidijithali: Guqula izinqolobane eziskeniwe, amarekhodi ezempilo nezinto zomlando zibe umbhalo ohlelekile ngobuningi, bese uzigcina ohlelweni.
Ukwakha i-RAG/nolwazi: Guqula ama-PDF amade kakhulu abe okuqukethwe okungaseshwa, kusekele imibuzo nezimpendulo, ukusesha nokubizwa kwama-agent ahlakaniphile.