Index-Translate Uyasebenza

I-Index-Translate iwumndeni wamamodeli okuhumusha ngezilimi eziningi ovulekile womphakathi, owakhiwe ithimba le-Index LLM le-Bilibili, wakhiwe phezu kwe-Qwen3.5 futhi usebenzisa ilayisensi ye-Apache-2.0. Imodeli yombhalo isekela izilimi ezingu-150, itholakala ngosayizi abathathu: 2B/9B/35B-A3B, futhi isekela imiyalelo yokuhumusha amagama-mqondo, amafomethi nokuqukethwe okufanele kugcinwe. Umndeni we-Index-Translate uhlanganisa ne-Index-Echo yokuhumusha ngezwi, i-Index-Homura yokuhumusha okulawulwa ngenani lamasilabhili, kanye ne-Index-NativeLong yokuhumusha amadokhumenti amade.

Ukuhumusha umbhalo nge-Index-Translate: Kuhumusha umbhalo nokuqukethwe okuhlelekile ngezilimi ezingu-150, kuyilapho kusebenza ngesikhathi esisodwa imiyalelo yokuhlanganisa amagama-mqondo, ukugcina ifomethi nokuhumusha izinkambu ezithile.
Ukuhumusha ngezwi nge-Index-Echo: Isekela izwi liye ezwini (S2ST dubbing) kanye nezwi liye embhalweni (S2TT imibhalo engezansi), futhi lapho ikhiqiza i-dubbing ibhekisa ezicini zethoni yezwi yesikhulumi somsindo wokuqala.
Ukuhumusha okulawulwa ngamasilabhili nge-Index-Homura: Yamukela inani eliqondiwe lamasilabhili njengokufakwayo, ilungise ngokuzenzakalelayo ukukhethwa kwamagama nobude bokuhumusha, inikeze umbhalo ohambisana nobude besikhathi bevidiyo.
Ukuhumusha amadokhumenti amade nge-Index-NativeLong: Ifaka idokhumenti yonke kanye bese ikhiqiza ukuhumusha ngokuqhubekayo, isebenzisa umongo omude ukugcina ukuvumelana kwamagama abantu, amagama-mqondo namazinga emiqondo.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxoxa ngamaphrojekthi e-AI avulekile
Isisekelo nesakhiwo esihlanganisiwe: Wonke umndeni wakhiwe phezu kwe-Qwen3.5. Ngokulungiswa kusetshenziswa imibhalo yezilimi eziningi nedatha yemiyalelo yokuhumusha, imodeli yombhalo ifunda ngesikhathi esisodwa ikhono lokuhumusha kanye nokulandela imikhawulo ngaphakathi kwesakhiwo esisodwa; ekucatshangweni kusetshenziswa i-vLLM, futhi imodeli yombhalo isebenzisa ngokuzenzakalelayo i-greedy decoding ukuze kuqinisekiswe okukhiphayo okuzinzile nokuphindaphindekayo.
Ukuhumusha ngezwi nge-Index-Echo: Kusetshenziswa ipayiphi elandelanayo “ukuqaphela→ukuhumusha→ukuhlanganisa”. Ku-S2ST, i-CosyVoice speech synthesis ilandela ukuhumusha, bese kukhishwa i-embedding yesikhulumi somsindo wokuqala futhi ifakwe kumojuli yokuhlanganisa ukuze kwenziwe i-voice cloning; umsindo uhlukaniswa ube amafasitela amasekhondi angu-60, kugcinwe umongo womlando wamafasitela ama-5 okuqala ukuze kugcinwe ukuhambisana kwenkulumo.
Ukulawula amasilabhili nge-Index-Homura: Inani eliqondiwe lamasilabhili lifakwa ku-prompt njengesimo esicacile. Ngokuhlanganisa ne-sampling decoding, imodeli ikhetha amagama futhi ihlele ubude ngesikhathi esifanayo sokuhumusha, ihlele kabusha isakhiwo semisho ukuze ifinyelele kwisabelomali samasilabhili; ku-SandGlass, u-81.92% wokukhiphayo unokuphambuka kwamasilabhili okungeqi ku-10%.
Ukuhumusha amadokhumenti amade nge-Index-NativeLong: Kusetshenziswa indlela yokufaka idokhumenti yonke ngokwendabuko, ifunde idokhumenti ephelele kanye bese ikhiqiza ukuhumusha ngokuqhubekayo. Isebenzisa iwindi lomongo elide kakhulu ukuze “ikhumbule” amagama abantu nokusetshenziswa kwamagama-mqondo engxenyeni yangaphambili ngesikhathi sokukhiqiza, ngaphandle kwesidingo sohlu lwamagama lwangaphandle; lokhu kuhlukile ekuhumusheni okuhlukanisa idokhumenti ngezingxenye, okuthembele kumongo oseduze nohlu lwamagama oluzenzakalelayo.
Indlela yokuhlola: Ngaphandle kwamabhentshimakhi ajwayelekile afana ne-FLORES ne-WMT26, ithimba lizakhele amasethi emisebenzi emine: instTrans, MEME, SandGlass, kanye nokuhlola amadokhumenti amade ngobude obuhlukahlukene. Ikhwalithi yokuhumusha nokulandela imiyalelo kunikezwa amaphuzu ngokwehlukana, okusondelana kakhulu nezimo zokusetshenziswa zangempela.

Clone inqolobane: sebenzisa i-git clone https://github.com/bilibili/Index-Translate.git bese ungena kufolda yephrojekthi.
Faka okudingekayo: sebenzisa pip install -U vllm kanye ne-pip install -r inference/llm/requirements.txt ukuze ulungise indawo ye-vLLM esekela i-Qwen3.5.
Qalisa isevisi: sebenzisa vllm serve IndexTeam/Index-Translate-2B --host 127.0.0.1 --port 8000 --max-model-len 4096 ukuze uqalise isevisi yemodeli kuphothi yendawo engu-8000.
Yenza ukuhumusha: vula itheminali entsha bese usebenzisa python inference/llm/translate.py "你好，世界" --target en --model IndexTeam/Index-Translate-2B ukuze uthole ukuhumusha.
Ukuhlanganisa ngendlela oyifisayo: Iklayenti ngokuzenzakalelayo licela ku-http://127.0.0.1:8000/v1. Ungasebenzisa i---base-url/--api-key noma okuguquguqukayo kwemvelo i-OPENAI_BASE_URL/OPENAI_API_KEY ukuze uxhumane nesevisi yakho.
Ukuhumusha okulawulwa ngamasilabhili: shayela i-inference/llm/syllable_translate.py, usebenzise i---syllables ukudlulisa inani eliqondiwe lamasilabhili.
Ukuhumusha amadokhumenti amade: shayela i-doc_translate.py, usebenzise i---direction ukucacisa izikhombisi-ndlela ezimisiwe ezifana ne-zh-en / zh-ja.
Ukuhumusha ngezwi: landela umhlahlandlela wemibhalo engezansi ye-S2TT noma umhlahlandlela we-dubbing we-S2ST osenqolobaneni, usebenzise i-s2tt.py noma i-dub.py ukucubungula umsindo nevidiyo.
I-demo eku-inthanethi: vakashela ngokuqondile ku-https://index-translate.bilibili.com/ ukuze uzame izinhlobo ezine zokuhumusha: umbhalo, izwi, amasilabhili namadokhumenti amade.
Isandiso sesiphequluli: Ngemva kokufaka imodeli endaweni, ungasebenzisa izandiso ze-Chrome/Edge/Firefox ukuhumusha amakhasi ewebhu.

Ukumbozwa kwezilimi okubanzi: Imodeli yombhalo isekela ukuhumushana kwezilimi ezingu-150, futhi isebenza kahle ezilimini ezingasetshenziswa kakhulu; i-35B-A3B ine-off-target rate engu-2.4% kuphela ku-FLORES_minor_pair.
Ukulandela imiyalelo okuqinile: Ngesikhathi sokuhumusha isebenza ngesikhathi esisodwa izidingo zokuhlanganisa amagama-mqondo, ukugcina ifomethi nokucacisa izinkambu. I-9B ithole u-0.7725 ekulandeleni imiyalelo ku-instTrans_minor, okuyiphuzu eliphakeme kunawo wonke phakathi kwamamodeli aqhathanisiwe.
Ukuqonda ama-meme omphakathi: Ukuhlolwa kwe-MEME kwenzelwe izinkulumo zomphakathi ezifana namagama azwakala ngokufana, izifinyezo namahlaya, futhi ikwazi ukuhumusha ngokunembile nangokwemvelo ama-meme afana no-“结芬” no-“狒瘾”.
Ukuhumusha ngezwi kugcina ithoni: Lapho i-Index-Echo ikhiqiza i-dubbing ngolimi oluqondiwe, ibhekisa ezicini zesikhulumi somsindo wokuqala ukuze yenze i-dubbing yezilimi ezahlukene egcina isitayela sezwi sokuqala.
Ukulungiselela i-dubbing ngokulawula amasilabhili: I-Index-Homura yamukela inani eliqondiwe lamasilabhili futhi ihlele kabusha imisho ngokwesabelomali. U-81.92% wokukhiphayo unokuphambuka kwamasilabhili okungeqi ku-10%, okuqondiswe ngqo ezimweni ze-video dubbing.
Ukuvumelana kuwo wonke umbhalo omude: I-Index-NativeLong ihumusha idokhumenti yonke ngokwendabuko, igcina ukuvumelana kwamagama abantu namagama-mqondo emibhalweni emide ngaphandle kohlu lwamagama lwangaphandle, futhi ihlukanisa kahle amazinga emiqondo.

Inqolobane ye-GitHub: https://github.com/bilibili/Index-Translate
Iqoqo lamamodeli le-HuggingFace: https://huggingface.co/collections/IndexTeam/index-translate
Umbiko wezobuchwepheshe: https://github.com/bilibili/Index-Translate/blob/main/docs/Index_Translate_Series_Technical_Report.pdf

I-video localization ne-dubbing: Sebenzisa i-Index-Echo ukuze uhumushe ividiyo ibe yi-dubbing ngolimi oluqondiwe, ugcine ithoni yesikhulumi sokuqala. Hlanganisa nokulawula amasilabhili kwe-Index-Homura ukuze imigqa ihambisane nobude besikhathi sezigcawu, okulungele uchungechunge lwamafilimu, izifundo nokusakazwa kwama-Vlog kwamanye amazwe.
Ukukhiqiza imibhalo engezansi ngezilimi eziningi: Sebenzisa ikhono le-S2TT lokuguqula izwi libe umbhalo ukuze ukhiqize ngokuzenzakalelayo imibhalo engezansi yesiShayina, isiJapane, isiNgisi, iSpanishi nezinye izilimi eziningi ezinkulumweni, ekusakazweni bukhoma nakumavidiyo amafushane, kusekelwe ukusatshalaliswa kokuqukethwe ngezilimi ezahlukene.
Ukuyisa okuqukethwe komphakathi kwamanye amazwe: Lapho uhumusha ama-bullet comments, ama-updates nama-post, gcina incazelo yama-meme omphakathi, njengokuthi “狒瘾”→“FFXIV itch”, kanye namathegi ezihloko nesakhiwo sama-emoji, ukuze okuqukethwe kwemidlalo namasiko e-anime kugcine umongo ezimakethe zakwamanye amazwe.
Ukuhumusha amadokhumenti amade namanoveli aku-inthanethi: Sebenzisa i-Index-NativeLong ukuhumusha inoveli yonke, i-glacier史 namanye amadokhumenti amade. Ngaphandle kokugcina ngesandla uhlu lwamagama, igcina ukuvumelana kwamagama abantu, amagama-mqondo namazinga emiqondo, okulungele okuqukethwe kwewebhu nokushicilelwa kwamanye amazwe.
Ukwenziwa kwasendaweni kwedatha ye-e-commerce nemikhiqizo: Ngokwemiyalelo, humusha kuphela amakholomu athile e-CSV yomkhiqizo, ugcine izihloko nezinombolo, noma uhumushe ngesakhiwo isimemezelo se-JSON ukuze ukuhumusha kuxhunywe ngokuqondile ezinhlelweni zebhizinisi ezilandelayo, kwehliswe izindleko zenqubo yokwenziwa kwasendaweni.