T3PO Uyasebenza

I-T3PO iyimodeli yokuhumusha ngesikhathi esisodwa esakazwa ngokuqhubekayo, evulelekile yomthombo, eyakhiwe yi-NetEase Youdao. Igama layo eligcwele lithi “simulTaneous Translation via pareTo Policy Optimization”. I-T3PO igxile ekuhumusheni ngesikhathi umuntu esakhuluma: ihlola ngokuguquguqukayo ukuthi umongo wanele yini, iqale ukuhumusha ngokushesha uma wanele, noma ilinde kancane uma unganele, isebenzisa umqondo we-Pareto ukulinganisa ikhwalithi nokubambezeleka. Uma i-T3PO ixhunywe ngemva kwe-R2T2, kwakheka uchungechunge lokukhuluma, ukulalela nokuhumusha ngesikhathi esisodwa, olulungiselelwe izimo ezinjengokuzilungisa komkhulumeli, izinombolo emibikweni yezimali, nokuhlunga imishwana yesiNgisi esetshenziswa ngomlomo.

Ukuhumusha ngesikhathi esisodwa okusakazwayo: Lapho ulimi lomthombo luqhubeka lungena, umbhalo ohunyushiwe ukhiqizwa ngesikhathi esifanayo ngaphandle kokulinda umusho ophelele. 
Isinqumo sokufunda noma sokubhala: Uma umongo wanele, kukhethwa u-WRITE ukuze kukhishwe ukuhumusha; uma unganele, kukhethwa u-READ ukuze kuqoqwe ulwazi olwengeziwe, kulinganiswa ngokuqhubekayo phakathi “kwejubane” “nokunemba”.
Ukulungiswa kwekhwalithi nokubambezeleka: Kusetshenziswa umqondo we-Pareto ukuze kuthuthukiswe ngokuhlukene ikhwalithi yokuhumusha nokubambezeleka, isu elifanayo lingabhekelela ukubambezeleka okuphansi noma ukwethembeka okuphezulu.
Umjikelezo ophelele wokuhumusha inkulumo ngesikhathi esisodwa: Ngemva kokuxhuma ku-R2T2, umbhalo owengezwayo wokubonwa kwenkulumo osuvele “uqinisekisiwe ngokuphelele” usetshenziswa njengolimi lomthombo, kwakheka inqubo “yokukhuluma, yokulalela neyokuhumusha ngesikhathi esisodwa”.
Ukujwayela izimo zangempela zenkulumo: Ibhekene nezimo ezifana nokuzilungisa kwesikhulumi, izinombolo emibikweni yezimali, namagama asetshenziswa kaningi esiNgisini, kusetshenziswa amasu okucubungula asondelene kakhulu nokuhumusha kwabantu ngesikhathi esisodwa.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI anomthombo ovulekile
Ukumodela umsebenzi: I-T3PO ibheka ukuhumusha ngesikhathi esisodwa njengenqubo yokwenza izinqumo ngokwandayo: esinyathelweni ngasinye imodeli ibheka “isiqalo solimi lomthombo esesizwakele + isiqalo senguqulo esesikhiqiziwe”, bese inquma ukuthi i-READ iqhubeke ilinde ulimi lomthombo, noma i-WRITE ikhiqize igama elilandelayo lolimi okuqondiswe kulo, kuze kuphele okukhiphayo.
Ukuhlukanisa ikhwalithi nokubambezeleka: ikhwalithi yokuhumusha nesikhathi sokuphendula kukalwa ngokwehlukana, bese kukhethwa indawo yokusebenza emngceleni we-Pareto; imodeli efanayo ingasebenzela izidingo ezimbili: “kungcono kube kancane kodwa kunembile” kanye “khipha uhlaka kuqala bese ululungisa”.
Ukuthuthukisa nge-Pareto DPO: ohlangothini lokuqeqesha kusetshenziswa izindlela zokuthuthukisa okuncamelwayo ezifana ne-Pareto DPO/PDPO, ukuze isu lithande kakhulu izenzo ze-READ/WRITE “ezinekhwalithi ephezulu ngaphansi kokubambezeleka okufanayo, noma ukubambezeleka okuphansi ngaphansi kwekhwalithi efanayo”.
Ukuqopha kwe-autoregressive ngokwandayo: ngesikhathi sokucabanga, i-WRITE ikhiqiza ama-token olimi okuqondiswe kulo ngokulandelana, kanti i-READ idla umongo owengeziwe wolimi lomthombo; isiqalo esiqondiswe kuso sisebenza ngesikhathi esisodwa “njengokuqukethwe okuhumushiwe” kanye “nomkhawulo olandelayo”, futhi okukhiphayo kuqalwa kuphela lapho incazelo yesiqalo somthombo isanele, ngaleyo ndlela kwehliswe ukuhumusha kabi kwemisho engakapheleli.
Ukuxhumanisa ne-ASR: i-T3PO uqobo igxile ekuhumusheni ngesikhathi esisodwa kusuka kumbhalo kuya kumbhalo; lapho uhlangothi lwangaphambili luxhunywa ku-ASR ye-append-only efana ne-R2T2, njalo lapho uhlangothi lokuqaphela lukhipha ingxenye yombhalo ezinzile, uhlangothi lokuhumusha luqala ngokushesha i-READ/WRITE.

Finyelela inqolobane: vakashela inqolobane ye-GitHub ethi https://github.com/netease-youdao/Confucius4-T3PO。
Lungisa imvelo: landela i-README ukuze wakhe imvelo ehlukile ye-Python/CUDA futhi ufake okuncikile.
Landa izisindo zemodeli: landa i-Confucius4-T3PO namamodeli angemuva ahambisanayo ku-Hugging Face noma ku-ModelScope.
Ukuhlola okusheshayo: sebenzisa isibonelo esisemthethweni noma i-Web UI ukuze uqinisekise ukuthi kungakhiqiza ukuhumusha okusakazwayo kwe-READ/WRITE.
Xhuma ukuqashelwa kwenkulumo: qala i-R2T2 WebSocket, bese usebenzisa umbhalo ozinzile owengezwa kancane kancane njengokufakwayo kolimi lomthombo ku-T3PO.
Lungisa amazinga okubambezeleka: ngokwezidingo zekhwalithi nokubambezeleka, lungisa amasu okufunda nokubhala kanye namapharamitha ahlobene ne-chunk/ukulinda.
Yenza ukuhlola kokuqina: gxila ekuhloleni ukulungisa inkulumo, izinombolo, amagama akhethekile, ukuxubana kwesiShayina nesiNgisi, kanye nokuhlukaniswa kwemisho emide.
Faka ekukhiqizeni: yipakishe ibe yisevisi yokusakaza, uqaphe ukubambezeleka kuze kuvele uhlamvu lokuqala, ukuhoxiswa/ukuma, ukusetshenziswa kwememori yevidiyo kanye nokuhambisana kwemisebenzi.

Ukukhipha okuqhubekayo okubambezeleka okuphansi: Akulindi ukuthi umusho wonke uphele; yamukela ulimi lomthombo bese ikhiqiza inguqulo ngesikhathi esifanayo, eduze nesigqi sokuhumusha ngomlomo kwabantu.
Ukulinganisela phakathi kwekhwalithi nokubambezeleka: Isebenzisa umqondo we-Pareto ukukhetha phakathi kokuthi “isheshe” nokuthi “inembile” kuye ngesidingo.
Isinqumo esizivumelanisa nezimo sokufunda nokubhala: Uma umongo unganele, iyalinda; iqala ukuhumusha kuphela uma usuwanele, inciphisa ukuhumusha okungalungile kwemisho engagcwalanga nokulungisa umsebenzi kaningi.
Ukuzivumelanisa nolimi olukhulunywayo lwangempela: Isebenza ngokuzinza ezimweni ezinjengokuzilungisa phakathi nenkulumo, izinombolo zemibiko yezimali, ukuxubana kwesiShayina nesiNgisi, namagama agcwalisa inkulumo, inikeze umuzwa osondele ekuhumusheni ngomlomo okusebenzayo.
Kulula ukuyixhumanisa nokuqaphela inkulumo: Uma ixhunywe ne-R2T2, ingakha ngokuqondile uchungechunge lokuhumusha ngomlomo lapho umuntu ekhuluma, elalela futhi ehumusha ngesikhathi esisodwa.

Indawo yokugcina ye-GitHub: https://github.com/netease-youdao/Confucius4-T3PO
Indawo yokugcina yamamodeli e-HuggingFace: https://huggingface.co/netease-youdao/Confucius4-T3PO

Imihlangano nemicimbi yokwethulwa yamazwe ngamazwe: ikhiqiza imibhalo engezansi yesiShayina nesiNgisi ngesikhathi sangempela, ukuze ababambiqhaza abakhuluma izilimi ezahlukene balalele futhi bafunde ukuhumusha ngesikhathi esisodwa.
Izingcingo zemiphumela yezimali nemihlangano yokwethulwa yocwaningo lokutshalwa kwezimali: iqinisekisa kuqala ukunemba kwezinombolo, amayunithi namagama akhethekile, igweme izinqumo zokutshalwa kwezimali ezingalungile ezibangelwa ukuhumusha okungalungile okusakazwayo.
Izinsizakalo zamakhasimende eziwela imingcele nezikhungo zezingcingo: inikeza abenzeli namakhasimende ukuhumusha kwezwi okunokubambezeleka okuncane, futhi isekela ukufakwa nokusebenza kwangasese ukuze kuhlangatshezwane nezidingo zokuthobela imithetho.
Imfundo ye-inthanethi nokuhumusha kanyekanye emakilasini: ihumusha ukufundisa kukathisha ngesikhathi sangempela ibe olimini lwabafundi, futhi isebenzisa uhlu lwamagama obuchwepheshe ukuthuthukisa ukunemba kwamagama akhethekile ezifundweni.
Imibukiso yezohwebo, uhambo nokuxhumana ungaxhunyiwe ku-inthanethi phakathi kwabakhuluma izilimi ezahlukene: inikeza imibhalo engezansi ephathekayo noma ukuhumusha kanyekanye ngamahedfoni ezindaweni ezinomsindo, lapho kuhamba khona abantu noma lapho inethiwekhi ingazinzi, ukuze kube lula ukuxhumana ngokushesha.