Clef Uyasebenza

I-Clef iwumndeni wamamodeli ezinqumo ovulekile we-Cloudflare, obeka umongo kuqala bese unquma isinyathelo esilandelayo. Ngokungafani namamodeli ezingxoxo akhiqiza umbhalo ovulekile, i-Clef ichaza umbuzo okwesikhashana esicelweni, inikeze inketho ngayinye amathuba, ukuze ikhodi ikwazi ukwenza umzila noma ukuvimba ngokusekelwe kulokho, ngaphandle kokuqeqesha kabusha njalo uma kwengezwa isigaba esisha. Emndenini we-Clef, i-Clef engu-27B igxile ekunembeni, kanti i-Clef-flash engu-9B ifinyelela ekubambezelekeni okumaphakathi okungu-38.8 milliseconds, okuyenza ifaneleke ukufakwa endleleni eshisayo ye-Agent.

Izinqumo ezisekelwe emathubeni: faka i-state nombuzo onezinhlobo ezichaziwe, bese uthola amathuba enketho ngayinye; ikhodi ingawasebenzisa ngokuqondile ukwenza umzila, ukuvimba noma ukudlulisela kumuntu.
Izinhlobo ezintathu zemibuzo: isekela i-noul, i-choice ne-score, ngemibuzo efika ku-64 esicelweni esisodwa.
Ukuhlukaniswa okubukwayo: ine-vision encoder eyakhelwe ngaphakathi, engafaka izithombe ezifika kwezi-4 kanye nombhalo ukuze kukalwe ndawonye.
Umongo ongu-64K: ungamukela amalogi amade, amathikithi osizo, isimo se-JSON nokunye okokufaka okuphelele.
Ukuhambisana ne-Jev API: i-schema yesicelo ihambisana ne-Jev, ngakho ekuthutheni kuvamise ukushintshwa kuphela i-endpoint negama lemodeli.
Ukucabanga okungeyona i-autoregressive: ngemva kwe-prefill eyodwa, izinketho ezisemthethweni zikalelelwa ngokufana; akukhiqizwa umbhalo nge-token ngayinye, ngakho ukubambezeleka kuphansi kakhulu kune-LLM evamile.
Ukusingathwa emaphethelweni okunokubambezeleka okuphansi: kusetshenziswa inethiwekhi yomhlaba wonke yama-GPU ye-Workers AI, i-Clef inebambezeleko emaphakathi engu-209ms, kanti i-Clef-flash ingu-38.8ms kuphela, ngakho ingangena endleleni eshisayo.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI omthombo ovulekile
Isakhiwo sezinqumo esingeyona i-autoregressive: i-Clef iqala ngokwenza ukubala kwe-forward kwe-prefill-only kokokufaka, bese ikala ngokufana izinketho ezisemthethweni ze-schema ezichazwe esicelweni, ikhiphe ngokuqondile umphumela ohlelekile ekumeleni kwangaphakathi.
Ukuqondiswa kokunaka ngezigaba ezimbili: inketho ngayinye esemthethweni iqala ikhiphe umongo ohlobene nayo ku-prompt, bese amapharamitha ezinkambu ehlukene enza i-cross-attend futhi axhumane ne-payload yokuqala ukuze kwakhiwe imodeli ehlangene. I-lexical prior igcina ukuvumelana kwencazelo phakathi kwezinketho, bese ubufakazi buhlanganiswa bube amaphuzu ngaphansi kwemikhawulo ye-schema.
Isisekelo esivaliwe + ukulungiswa okulula: isisekelo yi-Qwen3.8-27B ne-Qwen3.5-9B. Ingxenye eyinhloko iyavalwa, futhi kulungiswa ngokuhlanganyela kuphela i-routing head ne-rank-256 LoRA adapter. Lokhu kuthola ikhono lokwenza izinqumo ngamapharamitha ambalwa aqeqeshwayo, kwehlise izindleko zokuqeqesha nokusebenzisa.
Inhloso yokuqeqesha egxile ekulinganiseni: i-post-training isebenzisa i-label-smoothed cross-entropy ukuze kuqinisekiswe ukuthi okukhiphayo kuhlala ngaphakathi kwe-schema esemthethweni, bese yengeza i-Brier loss ekulinganiseni amathuba, ukuze amanani amathuba aphumayo athembeke.
Ukuqinisa nge-RLCD: i-Reinforcement Learning for Calibrated Decisions eyakhiwe ngaphakathi isetshenziswa njengenhloso yesibili yokuthuthukisa. Izinketho eziseduze ngokulandelana zithola amaphuzu ayingxenye, impendulo eqondile ithole amaphuzu aphelele, bese i-reference penalty ivimbela ukushelela kokusatshalaliswa, kuthuthukiswe ukunemba nokusebenza ezimweni ezintsha.
Isu ledatha: idatha yokuqeqesha iyidathasethi yangaphakathi eyenziwe ngokuhlanganisa. Ukuhlelwa kwezinkambu, ukwakheka kwe-prompt ne-schema kuyashintshwa ukuze kwandiswe ukusebenza ezimweni ezintsha, okwenza imodeli ingazweli endleleni umbuzo obhalwe ngayo futhi ikwazi ukujwayela izigaba ezintsha ezichazwa okwesikhashana esicelweni.

Qinisekisa i-akhawunti nezimvume: bhalisa i-akhawunti ye-Cloudflare, thola i-Account ID ne-API Token, bese uvula i-Workers AI.
Khetha imodeli: khetha phakathi kwe-@cf/cloudflare/clef noma i-@cf/cloudflare/clef-flash ngokwesidingo.
Yakha umzimba wesicelo: chaza i-model, i-state (umbhalo/isimo se-JSON) kanye ne-questions (imibuzo engu-1–64 yohlobo lwe-noul/choice/score, uchaze ngokucacile imingcele ye-criteria).
Qalisa ucingo: sebenzisa i-REST POST ku-/accounts/{id}/ai/run/@cf/cloudflare/clef, noma uxhume ucingo ku-Worker usebenzisa i-env.AI.run().
Hlaziya imiphumela yamathuba: funda amathuba ezinketho noma amaphuzu anesisindo emibuzweni ngayinye ku-response.answers, bese usebenzisa “inketho ephezulu + umkhawulo wamathuba” ukwenza umzila noma ukuvimba.
Faka izithombe (uyazikhethela): uma kudingeka ukuhlolwa okubukwayo, faka amabhayithi e-PNG/JPEG/WebP esicelweni; ungathumeli i-URL ekude.
Yenza ukuhlolwa komthamo ngaphambi kokwethula: hlola ukubambezeleka nokunemba usebenzisa ubude be-state nenani lemibuzo yakho, bese unquma ukushintshela kubo bonke abasebenzisi noma uhlukanise ithrafikhi ngokobungozi phakathi kwe-flash ne-clef.
Ukulungisa okuthuthukile (uyazikhethela): uma amalebula omkhakha ekhethekile, gcina idatha nge-AI Gateway bese usebenzisa insizakalo yokulungisa i-RL yeqembu le-FDE ukuze wenze imodeli ibe ngeyakho.

Iyashesha: ukukala ngokufana okungeyona i-autoregressive kanye nokusingathwa kuma-GPU asem边边eni kunikeza i-Clef ukubambezeleka okumaphakathi okungu-209ms, kanti i-Clef-flash ingu-38.8ms kuphela; ishesha ngo-2.5×–13× kune-Jev futhi ingangena endleleni eshisayo.
Inembe kakhulu: ikleliswe kuqala ku-Jev Decision Index, yathola amaphuzu aphezulu kumabhentshimakhi ezinqumo angu-7 kwangu-10 asemthethweni, futhi yawina izivivinyo ezi-3 kwezingu-4 ze-Typesafe workflow.
Yengeza izigaba ngaphandle kokuqeqesha kabusha: izigaba zichazwa okwesikhashana esicelweni, ngakho amalebula amasha awadingi ukuqeqeshwa kabusha kwemodeli njengakuma-classifier endabuko.
Amathuba alinganisiwe: i-Brier loss ne-RLCD ekufundiseni kwenza amathuba aphumayo athembeke, futhi umkhawulo ungasethwa ngokuqondile ukuze kwenziwe izinqumo ezizenzakalelayo.
I-multimodal yomdabu: i-vision encoder eyakhelwe ngaphakathi ingahlukanisa izithombe, kanti i-Jev okwamanje isebenza ngombhalo kuphela.
Umongo ophindwe kabili: umongo ongu-64K (i-Jev ino-32K) uvumela amalogi amade nesimo esiningi.
Ukuthutha ngaphandle kwezindleko: ihambisana ngokuphelele ne-Jev API, ngakho ukushintsha i-endpoint negama lemodeli kwanele ekushintsheni okushelelayo.

Iwebhusayithi yephrojekthi: https://blog.cloudflare.com/clef-decision-models/
Umtapo wamamodeli we-HuggingFace: https://huggingface.co/Cloudflare/clef
https://huggingface.co/Cloudflare/clef-flash

Ukuhlukaniswa kwamathikithi osizo lwamakhasimende: faka okuqukethwe kwethikithi nesithombe-skrini, bese esicelweni esisodwa unqume ukuthi kuphuthumayo yini, yiliphi iqembu okufanele kuthunyelwe kulo, kanye nesilinganiso sobunzima; ukuzethemba okuphansi kudluliselwa ngokuzenzakalelayo kumuntu.
Ubuhlakani bezinsongo nokulwa nobugebengu bokweba imininingwane: sebenzisa i-Browser Run ukuze uthwebule amakhasi ewebhu asebenzayo, uhlukanise izizinda noma amakhasi ngamalebula amaningi; ukuhlolwa okusemthethweni kukhombise ukuthi lonke uhlelo luthatha imizuzwana engu-2.2 kuphela.
Ukuhlolwa kwethrafikhi ye-Bot: endleleni eshisayo yesango, sebenzisa i-Clef-flash ukuze unqume ngokushesha ukuthi isiseshi sinobungozi noma sinobungane, usivimbe noma usivumele ngamamilisekhondi.
Ukuhlolwa kwangaphambili kokubizwa kwamathuluzi e-Agent: ngaphambi kokuthi i-Agent yenze isenzo esinobungozi obukhulu, buza i-Clef ukuthi “kufanele kwenziwe yini nokuthi kusetshenziswe liphi ithuluzi”, bese usebenzisa umkhawulo wamathuba ukunquma ukuthi kwenziwe noma kucelwe ukuqinisekiswa.