Xing4.0-29B-Intent Uyasebenza

I-Xing4.0-29B-Intent imodeli ekhethekile yokuqonda izinhloso ekhishwe njenge-open source yi-Xingchen Laboratory, eqeqeshwe kabusha kusetshenziswa i-Xing4.0-29B-A4B. I-Xing4.0-29B-Intent isebenzisa isakhiwo sesigaba＋amapharamitha, iguqula ulimi lwemvelo lomsebenzisi lube imiphumela yokuhlukaniswa kwenhloso nokukhishwa kwamapharamitha ngesikhathi esisodwa, futhi isekela imisebenzi emine: ukuqaphela izinhloso, ukuqaphela amabhizinisi, ukukhishwa kobudlelwano, nokukhishwa kwemicimbi. Okuphumayo kwemodeli kungama-token ambalwa kuphela, isikhathi sokuphendula sisezingeni lamakhulu ama-millisecond, futhi ithole amaphuzu aphelele angu-58 ekuhlolweni, yadlula amamodeli ajwayelekile angu-18. Isivele isetshenziswa ezinsizakalweni zamakhasimende ezokuxhumana nakumakhaya ahlakaniphile, futhi ifanele ukusebenza njengongqimba lokuqala lolimi lwe-Agent.

Ukuqaphela izinhloso: kunquma ukuthi umusho womsebenzisi ungowasiphi isigaba sebhizinisi, futhi kusekela inhloso eyodwa noma izinhloso eziningi emshweni owodwa.
Ukukhishwa kwamapharamitha: ngesikhathi sokuqaphela inhloso, kukhishwa amabhizinisi, amanani nezimfanelo ezidingekayo ukuze kuqedwe umsebenzi.
Okuphumayo okuhlelekile okuhlanganisiwe: imiphumela ihlelwa ngefomethi ye-JSON ethi “isigaba＋amapharamitha”; imisebenzi emine yokuqaphela izinhloso, ukuqaphela amabhizinisi, ukukhishwa kobudlelwano nokukhishwa kwemicimbi isebenzisa ifomethi efanayo yokufaka nokukhipha.
Ukuhlaziywa kwemisebenzi eminingi ngokuhambisana: kusekela izinhloso eziningi, amapharamitha amaningi namabhizinisi amaningi emshweni owodwa, futhi kungabuyisa imiphumela eminingi kanye ngesikhathi kusetshenziswa uhlu lwe-JSON.
Ukuvumelanisa okuqhutshwa yi-Schema: onjiniyela bachaza kusengaphambili izigaba ezingakhethwa, incazelo yesigaba nezinkambu ezingakhishwa; imodeli iyaqaphela futhi igcwalise amanani ngaphakathi kohlaka lwebhizinisi olumisiwe.
Ungqimba lolimi lwe-Function Calling: luhlinzeka izinhlelo ze-Agent ngokuqaphela izinhloso nokukhishwa kwamapharamitha okubiza imisebenzi, luxhumanise ulimi lwemvelo nokwenziwa kwezinhlelo zebhizinisi.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo zamaphrojekthi e-AI open source
Isisekelo nokwakheka kwemodeli: imodeli iqeqeshwe kabusha phezu kwesisekelo esijwayelekile se-Xingchen, i-Xing4.0-29B-A4B, futhi isebenzisa ukwakheka kwe-MoE. Inamapharamitha acishe abe yizigidigidi ezingu-29, kuyilapho ku-Token ngayinye kuvulwa kuphela amapharamitha acishe abe yizigidigidi ezingu-4, ngaleyo ndlela igcine amandla okukhipha okuhlelekile, inciphise izindleko zokubala futhi ithole isikhathi sokuphendula esisezingeni lamakhulu ama-millisecond.
Uhlaka lokuhlaziya ulimi oluqhutshwa imodi ehlanganisiwe ye-USP: imodeli isebenzisa uhlaka lwe-USP ukuhlanganisa imisebenzi ehlukahlukene njengokuqonda izinhloso, ukuqaphela amabhizinisi, ukukhishwa kobudlelwano nokukhishwa kwemicimbi ibe yifomethi yokukhipha ye-JSON ehlelekile enezigaba＋amapharamitha. Lokhu kwenza amasistimu angezansi angadingi ukulungisa ngokwehlukana umqondo wokuhlaziya womsebenzi ngamunye, futhi kwehlise kakhulu izindleko zokuthuthela nokuxhuma.
Indlela yokukhiqiza eboshwe yi-Schema: ngesikhathi sokucabanga, onjiniyela kufanele banikeze izigaba ezingakhethwa, izincazelo zezigaba nezincazelo zezinkambu ezingakhishwa. Imodeli iqedela ukuqaphela nokugcwalisa amanani ngaphakathi kwemithetho yezinkambu emisiwe, futhi okuphumayo kuboshwe ngokuqinile esakhiweni esichazwe kusengaphambili. Lokhu kugwema ukungahleliwe kombhalo okhululekile, kuqinisekise izinkambu ezinqunyiwe, ifomethi eqinisekisekayo kanye nokucabanga okumbalwa okungamanga.
Ukuqeqeshwa kokukhipha okuhlelekile: imodeli iqeqeshwe kusetshenziswa cishe amarekhodi okuqeqesha aqondene nalokhu ayizigidi ezingu-2.753, akhiwe kabusha kusuka kumasethi edatha omphakathi angu-61, ahlanganisa imikhakha engu-19 nezilimi ezingu-52. Lokhu kuyenza ikwazi ukuqedela uchungechunge oluphelele lokunquma inhloso nokukhishwa kwamapharamitha ngokukhipha ama-token ambalwa kuphela ngokwesilinganiso.

Thola izinsiza: landa izisindo zemodeli, ikhosombe lekhodi namasethi edatha ahambisanayo ku-GitHub noma ku-HuggingFace.
Lungisa indawo: faka isithombe esisemthethweni sokucabanga, noma wakhe isevisi yemodeli usebenzisa i-vLLM noma i-SGLang ngokulandela imiyalelo yemodeli eyisisekelo.
Faka i-SDK: xhuma nge-Python SDK enikezwe yikhosombe, noma shayela i-API yefu ye-Tianyi AI Open Platform ukuze uyisebenzise ngaphandle kokufaka imodeli.
Chaza i-Schema yebhizinisi: cacisa izigaba zezinhloso ezingakhethwa, izincazelo zezigaba nezinkambu ezingakhishwa esigabeni ngasinye.
Hlanganisa umbhalo womyalo: sebenzisa i-SDK ukuhlanganisa okufakwayo komsebenzisi ne-Schema kube isicelo esibonwa yimodeli.
Qalisa ucingo: thumela isicelo kusevisi yemodeli; imodeli ikhipha umphumela we-JSON ohlelekile onama-token ambalwa.
Hlaziya umphumela: funda izinhloso namanani ezinkambu ngokwesakhiwo sesigaba＋amapharamitha, futhi uhlole izinkambu uma kwenzeka iphutha.
Xhuma nebhizinisi: dlulisela umphumela ohlelekile ohlelweni olungezansi ukuze luwenze; lapho kuhileleka imisebenzi ebucayi, qala uhlole izimvume noma uqinisekise nomsebenzisi.

Ukuhlanganiswa kokuqaphela izinhloso nokukhishwa kwamapharamitha: imodeli eyodwa iqedela ukuhlukaniswa kwenhloso nokukhishwa kwamapharamitha ngesikhathi esisodwa, inciphise umsebenzi wokuhlaziya ophindaphindwayo ezinhlelweni ezingezansi.
Isivinini sokuphendula esisheshayo: ngokwesilinganiso ikhipha ama-token ambalwa kuphela, futhi isikhathi sokuphendula somphumela ophelele sisezingeni lamakhulu ama-millisecond, okulungele ukuxhumana kwesikhathi sangempela okuphindaphindeka kakhulu.
Izindleko eziphansi zokushaya: ekwakhekeni kwe-MoE, ku-Token ngayinye kuvulwa kuphela amapharamitha acishe abe ngu-4B; uma kuqhathaniswa namamodeli amakhulu ajwayelekile, umehluko wezindleko ezimweni zokushaya izikhathi eziyisigidi ngosuku wehla ngokuphindwe kaningana.
Okuphumayo okuzinzile nokuthembekile: imiphumela iboshwe ngokuqinile esakhiweni se-JSON se-Schema echazwe kusengaphambili, inezinkambu ezinqunyiwe nefomethi eqinisekisekayo.
Ukuhlaziywa okuyinkimbinkimbi kwezinhloso eziningi: kusekela izinhloso eziningi, amapharamitha amaningi namabhizinisi amaningi emshweni owodwa, futhi ibuyise konke ngohlu lwe-JSON oluhlanganisiwe.
Ifomethi efanayo yemisebenzi emine: ukuqonda izinhloso, ukuqaphela amabhizinisi, ukukhishwa kobudlelwano nokukhishwa kwemicimbi kusebenzisa ifomethi efanayo yokufaka nokukhipha ethi “isigaba＋amapharamitha”, okwenza izindleko zokuthuthela ezinhlelweni ezingezansi zibe phansi kakhulu.

Ikhosombe le-GitHub：https://github.com/XingChen-AGI/Xing4.0-29B-Intent
Ikhathalogi yemodeli ye-HuggingFace：https://huggingface.co/XingChen-AGI/Xing4.0-29B-Intent

Amasevisi amakhasimende ahlakaniphile / ukuqondisa amathikithi omsebenzi: iqaphela inhloso yengxoxo yekhasimende futhi ikhiphe ulwazi olubalulekile, iqede ngokuzenzakalelayo ukuhlukaniswa nokudluliswa kwamathikithi omsebenzi, futhi isetshenziswe ezinsizakalweni zamakhasimende ezokuxhumana.
Ukuxhumana ngezwi namadivayisi: ezimweni zabasizi bezwi, izihlalo zemoto ezihlakaniphile nezilawuli ezimaphakathi ze-IoT, ihlaziya imiyalo yomlomo yomsebenzisi ibe imiyalo ehlelekile yokulawula amadivayisi.
Amarobhothi e-IM ezinkampani: ngesikhathi sokuqala kwebhizinisi, akudingeki idatha enkulu enelebuli; amandla okuqaphela izinhloso nokuqondisa imibuzo nezimpendulo angakhiwa ngokushesha.
Ungqimba lolimi lwe-Agent / Function Calling: luhlinzeka i-Agent ngokuqaphela izinhloso okunokucabanga okumbalwa okungamanga nokukhishwa kwamapharamitha okubiza imisebenzi, luxhumanise ulimi lwemvelo nokusetshenziswa kwamathuluzi.
Ukukhishwa kolwazi kwemisebenzi eminingi: iqedela ngasikhathi sinye ukuqaphela amabhizinisi, ukukhishwa kobudlelwano nokukhishwa kwemicimbi emshweni owodwa, isibonelo: “Shintsha umhlangano uwubeke ntambama ngoLwesihlanu, indawo uyiyise esitezi sesithathu” → ikhiphe ngesikhathi nendawo entsha ngesikhathi esisodwa.