OXR Uyasebenza

I-OXR iyimodeli yokuhlaziya amadokhumenti engu-1.3B ekhishwe njenge-open source yi-SoMark yakwaShoukou Technology, eyenzelwe ama-PDF nezithombe zamadokhumenti, futhi ikhiphe i-Markdown/JSON elungele ama-Agent. Iyakwazi ukubona izinhlobo ezingu-13 zezinto, okuhlanganisa umbhalo, amatafula, amafomula, izibalo/izakhiwo zamakhemikhali nekhodi, futhi inikeze izixhumanisi zendawo ekhasini kanye nokulandelana kokufundwa. Ukuqeqeshwa kwangemva kwe-OXR kufaka umjikelezo wokuzithuthukisa we-RSI; ihamba phambili kumabhentshimakhi amathathu i-OmniDocBench, i-olmOCR-Bench ne-DocParsingBench, futhi ihamba phambili ngesivinini sokuhlaziya.

Ukuguqula amadokhumenti abe yidatha ehlelekile: ihlaziya ama-PDF nezithombe zamadokhumenti ibe yi-Markdown/JSON, ukuze ama-Agent namasistimu alandelayo akwazi ukukusebenzisa ngokuqondile.
Ukuqaphela izinto eziningi: isekela izinhlobo ezingu-13 zezinto zamadokhumenti, okuhlanganisa umbhalo, amatafula, amafomula, izibalo zamakhemikhali, izakhiwo zamakhemikhali namabhulokhi ekhodi.
Ukukhipha ulwazi lwesakhiwo sekhasi: inikeza amabhokisi okubopha emiphumeleni yokuqaphela futhi ihlele okuqukethwe ngokulandelana kokufundwa.
Ukucubungula izakhiwo eziyinkimbinkimbi: isekela amaphepha/ amaphephandaba anamakholomu amaningi, amaphepha ezivivinyo e-A3, umbhalo oqondile nokubhala ngesandla kwesiShayina nesiNgisi.
Ukuhlaziya isakhiwo sekhasi okwenziwa ngokwezifiso: imodeli yesakhiwo eyakhelwe ngaphakathi ingashintshwa ukuze ihambisane nezincazelo ezahlukene zezinhlobo zezinto, imingcele yezindawo namaqembu okuqukethwe.
Izindlela eziningi zokuxhuma: inikeza i-API yamahhala, i-demo eku-inthanethi, kanye nokufakwa kwendawo nge-Docker noma ngekhodi yomthombo, futhi iyahambisana ne-SoMark CLI/SDK.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI open source
Imodeli nokuqeqeshwa: I-OXR iyimodeli enamapharamitha angu-1.3B, eqeqeshwe kusukela ekuqaleni. Esigabeni sokuqeqeshwa kwangemva kwalokho kufakwa umjikelezo we-RSI ogxile ekuhlaziyeni amadokhumenti: amamodeli amaningi ahlaziya amasampula anamaphutha namasampula abuthakathaka, bese ekhiqiza ngokuzenzakalela amasu okuqeqesha nedatha.
Ukwahlukaniswa kwemisebenzi yesakhiwo: uhlelo lwakhiwe imodeli elula ye-Layout Analysis kanye nemodeli eyinhloko ye-OXR. Eyokuqala ithola izindawo zokuqukethwe, kanti eyesibili iqedela ukuqaphela okuqukethwe nokunquma ukulandelana kokufundwa esigabeni esifanayo sokuqaphela, inciphise izindleko zokuhlela okuhlukile futhi isebenzise ubudlelwano bezincazelo ukuthuthukisa ukunqunywa kokulandelana.
Idizayini enama-plug-in: ukuhlaziya isakhiwo sekhasi kuhlukaniswe kwaba imojuli ezimele, engashintshwa ngemodeli yesakhiwo evela komunye umhlinzeki noma eyakhiwe ngaphakathi, kuyilapho kugcinwa amakhono e-OXR okuqaphela okuqukethwe nokunquma ukulandelana kokufundwa, ngaleyo ndlela kwehliswe izindleko zokuzivumelanisa nezizinda ezahlukene.
Ukuthuthukisa ukusebenza kahle: inqubo yokuhlaziya yakhiwe kabusha ezingeni lomsebenzi, ukuze imodeli encane engu-1.3B ithole umthamo omkhulu wokucubungula igcina ukunemba. Ekuqhathaniseni okusemthethweni kokuhlaziya amakhasi angu-100 ngaphansi kwezimo ezifanayo ku-RTX PRO 6000, i-OXR ithatha cishe imizuzwana engu-101, ishesha kune-PaddleOCR-VL-1.6 ne-MinerU2.5-Pro.
Ukuhlanganiswa kwezinhlelo: i-API yokuhlaziya iyahambisana ne-SoMark, futhi ingaphinda isebenzise i-somark-py/somark-js/CLI. Ekufakweni kwendawo, i-Docker all-in-one iyindlela elula kakhulu; ukufakwa ngekhodi yomthombo kudinga ukuqaliswa ngokuhlukana kwesevisi yemodeli ye-OXR, isevisi ye-Layout Analysis kanye ne-API yokuhlaziya.

Ukulanda izinsiza: klona i-repository ye-GitHub, bese ulanda izisindo ze-OXR-1.0 BF16 ku-Hugging Face noma ku-ModelScope.
Ukuqalisa nge-Docker: yakha i-docker/Dockerfile.allinone bese ufaka ifolda yemodeli njenge-volume; ngemva kokuqalisa vakashela ku-localhost:8000/demo noma ku-/docs.
Ukufaka ngekhodi yomthombo: qala ngokulandelana isevisi yemodeli ye-OXR, isevisi ye-Layout Analysis, bese uqala i-API yokuhlaziya amadokhumenti futhi ulungise i-config.yaml.
Ukuqalisa ukuhlaziya: shayela i-/v1/parse/sync noma i-interface ye-asynchronous, noma usebenzise i-SoMark CLI/SDK ngokushintsha i-base URL bese uthumela ifayela.
Ukuthola imiphumela: khetha ifomethi yokukhipha efana ne-markdown/json, landa imiphumela yokuhlaziya bese uhlola izixhumanisi kanye nokulandelana kokufundwa.
Ukuyisebenzisa ekukhiqizeni: sebenzisa isevisi yendawo noma isevisi yezohwebo ye-SoMark, bese ushintsha noma ulungise imojuli yokuhlaziya isakhiwo ngokwezidingo zebhizinisi.

Ukunemba okuhamba phambili kumabhentshimakhi amathathu: ikleliswe endaweni yokuqala ku-OmniDocBench, olmOCR-Bench ne-DocParsingBench, futhi inzuzo yayo ibonakala kakhulu ezimweni zamatafula.
Isivinini esiphezulu: imodeli enamapharamitha angu-1.3B ithatha cishe imizuzwana engu-101 ekuhlaziyeni amakhasi angu-100 ngaphansi kwezimo ezifanayo, futhi ibekwe ngokusemthethweni njengeyokuqala phakathi kwamamodeli e-open source ahlolwayo.
Ukumbozwa okubanzi kwezinto: isekela izinhlobo ezingu-13 zezinto zamadokhumenti, okuhlanganisa umbhalo, amatafula, amafomula, izibalo/izakhiwo zamakhemikhali namabhulokhi ekhodi.
Okukhiphayo okulungele ama-Agent: imiphumela iqukethe izixhumanisi zesakhiwo sekhasi nokulandelana kokufundwa, futhi ingaguqulelwa ngokuqondile ku-Markdown/JSON ukuze ingene ekuhambeni komsebenzi.
Ukuthuthuka okuqhubekayo: ukuqeqeshwa kwangemva kwalokho kusebenzisa umjikelezo we-RSI ukuze kukhiqizwe ngokuzenzakalela amasu nedatha kusetshenziswa amasampula anamaphutha namasampula abuthakathaka, kunciphise ukuthembela ekulebuleni okwenziwa ngesandla.
Izindleko eziphansi zokuzivumelanisa: ukuhlaziya isakhiwo kuyimojuli ezimele enama-plug-in, engashintshwa ngemodeli yesakhiwo eyenziwe ngokwezifiso yebhizinisi, kuyilapho kugcinwa amakhono okuqaphela nokuhlela.

I-GitHub repository: https://github.com/SoMarkAI/OXR
Indawo yemodeli ye-Hugging Face: https://huggingface.co/SoMarkAI/OXR-1.0

Ukufaka ezisekelweni zolwazi ze-RAG: guqula ama-PDF/amadokhumenti askeniwe abe yi-Markdown/JSON enezixhumanisi zesakhiwo, ukuze kuthuthukiswe ikhwalithi yokuhlukanisa izingxenye, ukubuyisa nokuqondisa.
Ukuhlaziya ezezimali nemibiko yocwaningo: cwaninga imibiko enamakholomu amaningi, amatafula ayinkimbinkimbi nezincazelo zamagrafu, bese ukhipha ngokuzinzile isakhiwo samatafula nokulandelana kokufundwa.
Ukuguqulwa kwedijithali kwezemfundo nezivivinyo: hlaziya amaphepha ezivivinyo e-A3, okuqukethwe okubhalwe mpo noma ngesandla, ukuze kusekelwe ukwakhiwa kwamabhange emibuzo, ukumakwa nokuhlaziywa kwamaphutha.
Izinqubo zocwaningo nokushicilela: qaphela amafomula amaphepha ocwaningo, izibalo/izakhiwo zamakhemikhali, bese uziguqulela ku-LaTeX, mhchem ne-SMILES ukuze ziphinde zisetshenziswe.
Agent yamadokhumenti ezinkampani: hlinzeka ngokufakwa okuhlelekile kwezinkontileka, ama-CV, amarekhodi namadokhumenti ekhodi; imojuli yokuhlaziya isakhiwo ingashintshwa ukuze ihambisane nezakhiwo zemboni.