Laya Uyasebenza

I-Laya iyimodeli ye-AI yokwenza izinqumo engeyona i-autoregressive futhi ewumthombo ovulekile. Imodeli ayikhiqizi umbhalo; isebenzisa i-encoder eqondisa ezinhlangothini zombili i-ModernBERT/mmBERT, yenza i-forward pass eyodwa kuphela embhalweni kanye nombuzo onezinhlobo, bese ikhipha imiphumela ehlelekile namathuba alinganisiwe cishe ngo-33 ms. Imodeli inendlela eyakhelwe ngaphakathi yokuhambisa izilimi esekela izilimi ezingaphezu kuka-100. Uma iqhathaniswa nemodeli evaliwe i-Jev, i-Laya inenzuzo ekunembeni (76.6% uma kuqhathaniswa no-72.7%), ekusekeleni izilimi eziningi (izilimi ezingaphezu kuka-100) nasezindlekweni (ukusetshenziswa endaweni kumahhala).

Isinqumo nge-forward pass eyodwa: faka isimo kanye neqoqo lemibuzo onezinhlobo, i-forward pass eyodwa ibuyise ngokuqondile imiphumela ehlelekile ngaphandle kokukhiqiza umbhalo, cishe ngo-33 ms.
Izisekelo ezintathu zezinqumo: i-choice ikhetha okukodwa kwezinye izinketho futhi inikeze amathuba enketho ngayinye; i-score inikeza amaphuzu esikalini esihleleke ngokulandelana futhi ibonise ukusabalala; i-noul iphendula imibuzo ethi yebo noma cha futhi ibuyise amathuba okuba impendulo iyiqiniso aphakathi kuka-0–1.
Ukuhlela ulimi ngokuzenzakalelayo: i-Router ithola umbhalo nolimi lokufaka ngaphansi kuka-0.5 ms, bese isabalalisa ngokuzenzakalelayo ku-checkpoint yesiNgisi noma yezilimi eziningi, igweme ukukhetha imodeli engalungile eholela ekuphenduleni ngokuzethemba kodwa ngokungalungile.
Ukulawula ngokuzethemba: amathuba aqeqeshwe nge-RLCD anencazelo yezibalo; izinqumo ezinokuzethemba okuphezulu zingadluliswa ngokuzenzakalelayo ngokomkhawulo, kanti ezinokuzethemba okuphansi zidluliselwe kumuntu.
Izilungiselelo zokusebenza ezakhelwe ngaphakathi: kunikezwa izifanekiso zemibuzo ezilungiselelwe ukuhlela amamodeli, ukuvikela ekufakweni kwemiyalelo, ukuhlola ukuphepha kokuqukethwe nokuhlukanisa amathikithi, ukuze zisetshenziswe ngokushesha.
Ukulungisa imodeli endaweni: kufakwe i-Kaggle notebook esekela yonke inqubo yokulungisa imodeli ngedatha yakho, okuhlanganisa ukwakha idatha, ukuqeqeshwa kwe-RLCD, ukulinganisa izinga lokushisa nokuhlola.

Landela i-WeChat bese uphendula ngokuthi “umthombo ovulekile” ukuze ujoyine iqembu lokushintshisana ngamaphrojekthi e-AI omthombo ovulekile
Ukwakhiwa kwe-bidirectional non-autoregressive: kusekelwe ku-ModernBERT-large (421M) noma ku-mmBERT-base (322M) encoder eqondisa ezinhlangothini zombili, lapho ama-token okufaka ebona umongo wangaphambili nowangemuva ngesikhathi esisodwa; ngesikhathi sokubikezela akukhiqizwa umbhalo uhlamvu ngohlamvu. Kunalokho, inketho ngayinye inikezwa indawo ezimele ye-[MASK], bese i-forward pass eyodwa ikhipha ama-logits azo zonke izinketho, eziguqulwa nge-Softmax zibe ukusabalala kwamathuba. Yilokhu okwenza kube nokubambezeleka kwezinga lama-33 ms.
Ukuqeqeshwa kokuqinisa nge-RLCD: kusetshenziswa imithetho eqinile yokukala iqiniso, efana nokukala ngamaphuzu e-log, ukukala okuyindilinga nokukala amathuba ngokwamazinga, njengomvuzo. Imodeli ithola amaphuzu aphezulu kuphela uma ikhipha amathuba angempela angemva kwesigameko; uma ingaqinisekile, ukuzethemba kwayo kwehla ngokwemvelo, kugwenywe inkinga yamamodeli amakhulu endabuko aphendula ngokuzethemba ngokweqile.
Indlela yokuhlela izilimi: ngaphambi kwe-forward, i-Router isebenzisa i-Python kuphela ukuthola amasethi ezinhlamvu ze-Unicode nolimi, bese ikhetha i-checkpoint ngokwalokho; njengoba i-checkpoint yesiNgisi inganikeza impendulo enembayo engu-0 ngokuzethemba okungu-0.95 emibhalweni engeyona eyesiLatin, ukuzethemba ngokwakho akukwazi ukuxwayisa kusenesikhathi, ngakho isinqumo sokuhlela kufanele senziwe ngaphambi kokubikezela.
Ukulinganisa amathuba: ama-checkpoint aphuma efekthri ajwayele ukuzethemba ngokweqile; izinga lokushisa kufanele lilinganiswe kudatha yakho ngokohlobo lombuzo nenani lezinketho. Ngemva kokulinganisa, i-ECE emaphakathi ingehliswa isuke ku-0.466 iye ku-0.081, lapho amathuba esengasetshenziswa khona ngokuqondile ezinqumweni zokudlulisa ngokuzenzakalelayo.

Ukufaka: pip install laya, iphakheji ye-PyPI; ezinsukwini ezimbili ishintshe isuka ku-0.1.0 yaya ku-0.3.4, qaphela ukuthi i-interface isashintsha.
Ukulayisha imodeli: i-Router(preload=True) ilayisha kusengaphambili ama-checkpoint amathathu ukuze igweme ukwakha kabusha imodeli njalo lapho kushintshwa ulimi ngesikhathi sokuqala esibandayo.
Ukulungisa isimo: hlanganisa umbhalo ozohlaziywa ube yi-dict, njengokuthi {"from": ..., "subject": ..., "body": ...}; kusekela noma yiluphi ulimi noma isakhiwo se-JSON.
Ukuchaza imibuzo: sebenzisa i-JSON ukubhala iqoqo lemibuzo onezinhlobo, ucacise i-type (choice / score / noul), i-instructions kanye ne-criteria (izinketho noma isikali).
Ukwenza ukubikezela: shayela i-router.predict(state, questions); i-forward pass eyodwa ibuyise zonke izimpendulo, isiNgisi sisebenzisa i-laya, kanti isi-Hindi nezinye izilimi zisebenzisa i-multilingual ngokuzenzakalelayo.
Ukufunda imiphumela: ku-res["answers"] thatha ilebula le-choice, inani le-score noma amathuba e-noul; i-res["routing"] izochaza ukuthi kungani kukhethwe leyo checkpoint.
Ukulawula ukuzethemba: sethela i-confidence umkhawulo, njengokuthi 0.85; okuzethemba kakhulu kucutshungulwe ngokuzenzakalelayo, okuzethemba kancane kudluliselwe kumuntu. Lokhu kufanele kwenziwe kuphela ngemva kokulinganisa izinga lokushisa kudatha yakho.

Ukubikezela okusheshayo: i-forward pass eyodwa cishe ngo-33 ms (T4), kanti ekucubunguleni iqoqo kwehla kuye ku-7.2 ms/ngombuzo, cishe kushesha izikhathi ezingu-7.8 kune-TypeSafe Jev evaliwe.
Ayikho ingozi yokusungula okungamanga: ayikhiqizi umbhalo, ikhipha izinqumo ezihlelekile kuphela, ngakho ayikho imithwalo yokuhlaziya futhi ayikho inkinga yokuqamba okuqukethwe.
Amathuba ayalinganiswa: iqeqeshwe ngemithetho eqinile yokukala iqiniso nge-RLCD, ngakho ukuzethemba kunencazelo yezibalo futhi kungashayela ngokuqondile ukuhamba ngokuzenzakalelayo noma inqubo yokubuyekezwa ngumuntu.
Ukunemba kokulinganisa kuhamba phambili: ngemva kokulinganisa izinga lokushisa, i-ECE ingu-0.081 kuphela, ingcono kune-0.246 ye-Jev; ukunemba kwe-typed-decisions okungu-0.766 nakho kudlula u-0.727 we-Jev.
Ukuhlela izilimi eziningi: i-Router eyakhelwe ngaphakathi ibona izilimi ezingaphezu kuka-100 ngaphansi kwe-millisecond futhi ikhethe ngokuzenzakalelayo i-checkpoint, igweme imodeli yesiNgisi ephendula ngokuzethemba kodwa ngokungalungile emibhalweni engeyona eyesiLatin.
Umthombo ovulekile ngokuphelele futhi mahhala: ilayisense ye-Apache-2.0, izisindo zivulekile futhi ingasingathwa nguwe; uma iqhathaniswa nezindleko ze-Jev API ezingu-$0.042 ngamathokheni ayisigidi, izindleko zokusetshenziswa isikhathi eside zinguziro.
Okukhiphayo onezinhlobo: izisekelo ezintathu, i-choice / score / noul, zihlanganisa iningi lezimo zokwenza izinqumo ezihlelekile, okuhlanganisa ukuhlukanisa, ukukala nokwahlulela yebo noma cha.

Indawo yokugcina ye-GitHub: https://github.com/NandhaKishorM/laya
Indawo yamamodeli ye-HuggingFace: https://huggingface.co/convaiinnovations/laya

Ukuhlukanisa ngokuzenzakalelayo amathikithi esevisi yamakhasimende: nquma umnyango ofanele (billing/technical/sales), izinga lokuphuthuma, ukukhungatheka komsebenzisi nobungozi bokulahlekelwa ikhasimende; amathikithi anokuzethemba okuphezulu anikezwe ngokuzenzakalelayo, kanti anokuzethemba okuphansi adluliselwe kumuntu.
Ukuphepha nokuqapha okuqukethwe: thola okuqukethwe okunobuthi, ukuhlukumeza nezinsongo; isisekelo se-noul sikhipha amathuba okwephulwa kwemithetho, okulungele ukuqapha ngesikhathi sangempela emiphakathini nasemapulatifomu e-UGC.
Ukuvikela ekufakweni kwemiyalelo (Guardrails): thola ukuhlasela okufana nokweqa imikhawulo, ukufakwa kwemiyalelo nokudalulwa kokhiye bemfihlo emnyango wohlelo lwe-Agent; ukubambezeleka kwezinga lama-33 ms kwenza kube lula ukukufaka kuwo wonke uchungechunge lwezicelo.
Ukuhlela amamodeli ahlakaniphile: esakhiweni samamodeli amakhulu axubile, nquma ukuthi isicelo kufanele siye kumodeli encane noma kumodeli ehamba phambili, wenze isinqumo sokuhlela ngezindleko eziphansi kakhulu futhi unciphise izindleko zokubikezela.