Kev Uyasebenza

I-Kev iwumndeni wamamodeli okwenza izinqumo okhishwe njengomthombo ovulekile ngu-Jared Palmer, i-VP yezobunjiniyela kwa-Cognition, osekelwe ku-Qwen3.5/Qwen3.8 futhi onamamodeli amaningana kusukela ku-0.8B kuya ku-27B. I-Kev ayikhiqizi umbhalo; ifaka umbhalo wesimo kanye nombuzo ohlelekile, bese ngokudlula okukodwa ikhiphe ngokuqondile amathuba alinganisiwe. Isekela izinhlobo ezintathu zezinqumo: yebo/cha, ukukhetha izinketho, nokunikeza amaphuzu ezingeni. Iphinde ihambisane ne-TypeSafe System One API.

Ukwenza izinqumo okuhlelekile: faka umbhalo wesimo kanye nemibuzo eminingana ehlelekile; isicelo esisodwa sikhipha ngokuqondile amathuba alinganisiwe enketho ngayinye, ngaphandle kokukhiqiza umbhalo ochazayo.
Izakhi ezintathu zezinqumo: isekela i-noul, i-choice kanye ne-score.
Ukuhlukaniswa kwemibuzo eminingi: imibuzo eminingi yabelana ngesimo esifanayo sokufaka, kodwa ihlukaniswa kusetshenziswa imaski yokunaka; yonke imibuzo iphendulwa ngesikhathi esisodwa ngokudlula okukodwa, ngaphandle kokungcolisana.
Ukusebenzisa endaweni: i-kev.serve inikeza umyalo owodwa wokuqalisa isevisi, iveze i-endpoint ye-POST /v1/systemone, futhi ihambisane ne-CUDA, i-ROCm kanye ne-Apple Silicon (MLX).
Ukulungisa ngemali ephansi: ungasebenzisa idatha yakho enamalebula ukuqhubeka nokuqeqesha kusukela ku-checkpoint esemthethweni; i-kev-finetune skill yenza i-coding agent iqede ngokuzenzakalela ukuhlela idatha, ukuqeqesha, ukuhlola nokuthumela.
Uhlelo lokuhlola olwakhelwe ngaphakathi: lufaka isethi yokuhlola engaguquki, lubike ukunemba, amaphuzu e-Brier, iphutha lokulinganisa kanye nesilinganiso sezinqumo ezizenzakalelayo; i-CI iphinde ihlole ngokuzenzakalela izinombolo ezishicilelwayo.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI omthombo ovulekile
Ikhanda lesikhombi liguqula ukukhiqiza kube ukuhlukanisa: I-Kev igcina izisindo eziyisisekelo ze-Qwen zingaguquki, bese yengeza i-LoRA adapter ye-rank-16 kanye nekhanda lesikhombi. Isimo esifihlekile senketho ngayinye kanye nesimo esifihlekile sombuzo kuyanikezwa amaphuzu, bese i-softmax ikhipha amathuba. Empeleni, lokhu kuguqula “umbuzo wokubhala” wemodeli enkulu ube “umbuzo wokukhetha”, kudlule ukukhiqiza ithokheni ngethokheni.
Ukunaka okuyimbangela ngamabhulokhi kuhlukanisa imibuzo: ukulandelana kokufaka kuhlelwa kube “isimo + amabhulokhi emibuzo eminingi”. Imaski yokunaka ivumela ithokheni ngayinye ukuthi ifunde kuphela isimo nengaphakathi lombuzo wayo; ama-ID ezikhundla aqala ukubalwa kabusha embuzweni ngamunye. Ngakho ukudlula okukodwa kungaphendula imibuzo eminingi engahlobene, yabelane ngokubalwa kwesimo futhi ingavuzi ulwazi phakathi kwemibuzo. Kumamodeli afana ne-Qwen3.5/3.8 ane-Gated DeltaNet exubile, umbuzo ngamunye uhlelwa emgqeni ozimele futhi kusetshenziswe kabusha inqolobane ye-KV yesimo, okwenza ukuhlukaniswa kuqine kakhulu.
Ukuqeqesha nokulinganisa: kusetshenziswa i-cross-entropy kumalebula alungile ukuze kuqeqeshwe i-adapter nekhanda lesikhombi, kuyilapho isisekelo sigcinwa singaguquki. I-checkpoint ngayinye ilinganisela ipharamitha ye-temperature kudatha ebekelwe eceleni; ngesikhathi sokusebenzisa imodeli ikhipha ngokuqondile amathuba asele alinganisiwe. I-temperature ayishintshi ukuhleleka, ilungisa kuphela amanani okuzethemba.

Ukulungiselela imvelo: faka i-Python 3.12/3.13 kanye ne-uv, klona i-repository bese usebenzisa i-uv sync --extra serve ukufaka okuncikile kukho.
Ukuqalisa isevisi endaweni: sebenzisa i-KEV_DTYPE=bf16 uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8009 ukuze uqalise imodeli; ekuqaleni izilayisha ngokuzenzakalela izisindo.
Ukuthumela isicelo sesinqumo: thumela i-state nombuzo we-questions ku-POST /v1/systemone; kuzobuyiswa amathuba alinganisiwe ezinketho.
Ukusebenzisa i-Python SDK: sebenzisa i-TypeSafe official SDK ubhekise i-base_url kusevisi yendawo; izinhlelo ezikhona ze-Jev zingashintsha ngaphandle kolunye ushintsho.
Ukuhlola esipheqululini: vakashela i-Hugging Face Space noma uqalise i-playground endaweni; akukho ukufakwa okudingekayo.
Ukuqeqesha ngedatha yakho: hlela idatha enamalebula ngefomethi ye-JSONL, bese usebenzisa i-uv run python -m kev.train --init_from jaredpalmer/kev-4b ... ukuqhubeka nokuqeqesha kusukela ku-checkpoint esemthethweni.
Ukuqeqesha okuzenzakalelayo ngokuphelele nge-Agent: sebenzisa i-npx skills add jaredpalmer/kev@kev-finetune bese uvumela i-coding agent iqede yonke inqubo yokuthola izinkinga, ukudala idatha, ukuqeqesha, ukuhlola nokuthumela, ngaphandle kwe-GPU yendawo.
Ukuthumela efwini: sebenzisa i-pip install modal && modal deploy kev_serve.py ukuze uthumele i-endpoint ye-HTTPS ngokuchofoza okukodwa; lapho ingasetshenziswa iyancipha ngokuzenzakalela ibe ngu-zero futhi ayikhokhisi.
Ukuhlola ukusebenza: sebenzisa i-uv run python -m kev.benchmark --run <imodeli> --suite evals/v4/transfer-v4 ukuze uhlole ukusebenza kusethi yokuhlola engaguquki; hlola kuqala ngaphambi kokuyisa ekukhiqizweni.

Izindleko eziphansi kakhulu: ukudlula okukodwa kukhipha amathuba ngokuqondile, kugweme ukukhiqiza ithokheni ngethokheni; ku-H100 iphendula imibuzo engu-6 ngama-18ms kuphela, futhi ngemva kokuzisingathela akukho mali yokubiza.
Ukuqeqesha kushibhile futhi kuyaphindaphindeka: ukuqeqesha ngokudlulisa amamodeli amathathu asemthethweni kubize cishe ama-dollar angu-95 kuphela, kanti ukulungisa imodeli yomuntu siqu kubiza cishe i-dollar engu-1 ngesikhathi esisodwa; zonke izindleko zivezwa ngokusobala.
Umthombo ovulekile futhi olawulekayo: ngaphansi kwelayisensi ye-Apache-2.0, izisindo, ikhodi yokuqeqesha kanye nesethi yokuhlola engaguquki konke kuyatholakala; ungaziqeqesha, uhlole futhi ugweme ukubanjwa ngumhlinzeki oyedwa.
I-API elungele ukusetshenziswa ngokushesha: ihambisana ngokuphelele ne-TypeSafe System One; izinhlelo ezikhona ze-Jev zingashintsha ngokulungisa i-base_url kuphela.
Imibuzo eminingi ngesikhathi esisodwa: isimo sokufaka siyabelwana, ukunaka okuyimbangela ngamabhulokhi kuyahlukanisa, futhi ukudlula okukodwa kuphendula noma yiliphi inani lemibuzo ngaphandle kokungcolisana.

I-GitHub repository: https://github.com/jaredpalmer/kev

Ukuhambisa amathikithi esevisi yamakhasimende: ngokudlula okukodwa, nquma ngesikhathi esisodwa ukuthi “ithunyelwe kumuphi umnyango, ingabe idinga ukukhushulelwa kumuntu, futhi lingakanani izinga lokucasuka kwekhasimende”; hambisa ngokuzenzakalela ngokusekelwe emngceleni wamathuba, bese udlulisela kumuntu lapho ukuzethemba kuphansi.
Ukuhlola ukuphepha kokuqukethwe: sebenzisa ngokuhambisana ku-UGC umbhalo othi “i-noul: ingabe kwephula umthetho + i-choice: uhlobo lokwephula + i-score: ubukhulu”; thatha isinyathelo ngokuzenzakalela lapho ukuzethemba kuphezulu.
Ukuhlola kokuqala kwezezimali/zesikweletu: nquma uhlobo lwesikhalazo, ukuthi kuyadingeka yini ukungenelela komuntu kanye nezinga lobungozi; ukusatshalaliswa kwamathuba kusiza ekugcineni ubufakazi bokuhambisana nemithetho nasekuhleleni kuqala ukubuyekezwa komuntu.
Ukuhlukanisa imikhiqizo noma izinhloso ze-e-commerce: dlulisela imibuzo yabasebenzisi ezinhlelweni ezilandelayo ezifana nokusesha, izincomo noma isevisi yangemva kokuthengisa, futhi uphathe inkinga yokusatshalaliswa kwamathegi amaningi lapho “igama elilodwa lingena ezigabeni eziningi”.