SemIf Uyasebenza

SemIf (eyayibizwa ngokuthi OpenJev) iyimodeli evulekile ye-AI yokwenza izinqumo, esebenzisa amamodeli avulekile ukuphindaphinda iphethini yesixhumi esibonakalayo sezinqumo zencazelo yesevisi evaliwe i-Jev ye-TypeSafe. I-SemIf ifunda ngokuqondile amaphuzu e-logits ahambisana nezinketho ezingamakhandidethi ekudluleni okukodwa okuya phambili, bese iseva yakha umphumela ohlelekile; uma imibuzo eminingi yabelana ngesimo, isiqalo singasetshenziswa kabusha, okusheshisa ukusebenza cishe ngokuphindwe kayishumi. I-SemIf isebenza ku-RTX 3090, inikeza umbukiso wesiphequluli, futhi ihlanganisa ibhenchimakhi ephindaphindeka umugqa ngomugqa, ukulinganiswa kwezinga lokushisa kanye nokuxhunywa kwemodeli engu-27B.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI avulekile
Ukulinganisa okuqondile kwe-logit: ukudlula okukodwa okuya phambili kufunda ama-logits ezinketho ezingamakhandidethi ukuze kutholakale amathuba, ngaphandle kokukhiqiza umbhalo wempendulo bese uwuhlaziya; kushesha cishe izikhathi ezi-5 kunokukhiqiza okuzenzakalelayo.
Ukusetshenziswa kabusha kwesimo esabiwe: lapho izinqumo eziningi zabelana ngomongo ofanayo, isiqalo sigcwaliswa kanye, bese amazinga ahlukahlukene ehlolwa ngokulandelana noma ngokufana; umthamo wokucubungula ungakhuphuka cishe ngokuphindwe kayishumi.
Izinqumo ezichazwa ngesikhathi sokusebenza: izindinganiso nezincazelo zezinketho zidluliswa nesicelo, ngaphandle kokulungisa imodeli noma ukuchaza imisebenzi kusengaphambili; lokhu kulungele izinqumo ezincane zama-agent njengokuzulazula nokunquma ukuphinda uzame.
Ukuhlolwa okuphelele: i-prompt emisiwe, inguqulo yemodeli, i-hashi ye-prompt, imiphumela yomugqa ngamunye kanye nokwehluleka okwaziwayo konke kugcinwa endaweni yokugcina, okusekela ukuphindaphinda okunembile.
Ukulinganiswa kwezinga lokushisa: kufakwe ukukalwa kokushintshwa kwezinga lokushisa okuqeqeshwe ngokomthwalo womsebenzi, okuthuthukisa ukuhambisana phakathi kokuzethemba kwamathuba nokunemba kwangempela.
Ukusekelwa kwama-backend amaningi: kusekela i-CUDA GPU, i-Apple Silicon (MPS/MLX), i-CPU kanye nombukiso we-WebGPU esipheqululini.

Ukulungiselela imvelo: faka i-Python 3.10+, i-CUDA kanye nekhadi lemifanekiso elikwazi ukusingatha imodeli engu-4B BF16.
Ukufaka iphrojekthi: ngemva kokudala imvelo ebonakalayo, sebenzisa i-pip install -e '.[test]'.
Ukulungisa okokufaka: bhala ifayela lezinqumo elikufomethi ye-JSONL, lapho umugqa ngamunye uqukethe i-id, i-state, i-question kanye nohlu lwezinketho ezingamakhandidethi.
Ukuqalisa ukulinganisa: sebenzisa semif-score --mode direct --model Qwen/Qwen3.5-4B --revision <fixed version> --input examples/decisions.jsonl --output results.jsonl ukuze uthole amathuba ezinketho ngomugqa ngamunye, isikhathi sokucubungula kanye ne-hashi ye-prompt.
Imodi yesimo esabiwe: uma i-state yayo yonke imigqa ifana ngokuphelele, sebenzisa --mode shared, ugcwalise isiqalo kanye bese uhlola zonke izindinganiso ngokufana.
Ukusebenzisa ukulinganiswa: sebenzisa umbhalo ohlinzekwe yiphrojekthi ukuze ulungise ukushintshwa kwezinga lokushisa kumthwalo womsebenzi okuhlosiwe, ukuze ukuzethemba kokuphumayo kuhambisane nokunemba kwangempela.

Isivinini esikhulu sokwenza izinqumo: ukufunda ngokuqondile ama-logits ezinketho kususa inqubo yokukhiqiza nokuhlaziya; izinqumo ezingama-21 zebhanari zithatha cishe isekhondi eli-1 kuphela, kushesha ngokuphindwe ka-5.2 kunokukhiqiza i-JSON.
Ukusetshenziswa kabusha kwesiqalo esabiwe: ukugcwaliswa kwangaphambili kanye, ukuhlolwa kwamagatsha kaningi; umthamo wesimo esabiwe ukhuphuka usuka ku-2.33 uye ku-20.03 izinqumo/ngomzuzwana.
Awekho ama-token akhiqizwayo: umbhalo wempendulo awusampulwanga, kuqedwe amaphutha efomethi ye-JSON, izihibe zokukhipha amakhodi kanye nokwehluleka kokuhlaziya okukhiphayo.
Izidingo zehadiwe eziphansi: imodeli engu-4B emisiwe isebenza ku-RTX 3090 eyodwa, futhi isekela i-CPU, i-Apple Silicon kanye ne-WebGPU yesiphequluli.
Umthombo ovulekile futhi ophindaphindekayo ngokuphelele: wonke ama-prompt, izinguqulo zamamodeli, imiphumela yomugqa ngamunye kanye nokwehluleka okwaziwayo kugcinwa endaweni yokugcina, ukuze kuhlolwe futhi kuqinisekiswe into ngayinye.
Ukuzethemba kungalinganiswa: ukushintshwa kwezinga lokushisa ngokomthwalo womsebenzi kwakhelwe ngaphakathi, kuqondanisa ukuzethemba kwamathuba nokunemba kwangempela ukuze okukhiphayo kusetshenziswe ngokuqondile ekwenzeni izinqumo okuzenzakalelayo.

Iwebhusayithi yephrojekthi: https://openjev.com/
Indawo yokugcina ye-GitHub: https://github.com/TheoLeeCJ/SemIf-OpenJev

Ukuqondisa amathikithi esevisi yamakhasimende: ngokusekelwe encazelweni yenkinga yekhasimende nesimo se-akhawunti, amaphuzu abekwe ngokuqondile ezinkethweni zemigqa efana nokuthi “usekelo lwe-akhawunti / usekelo lokukhokha”, ukuze kusatshalaliswe ngokuzenzakalela ngesikhathi esingama-millisecond.
I-firewall yezenzo ze-Agent: ngaphambi kokuba i-Agent yenze isenzo esinobungozi obukhulu, izenzo ezingamakhandidethi zihlolwa ngobhanari; ukunemba okuyinhlanganisela okungu-0.700 kungasekela umugqa wokuqala wokuvikela.
Ukusesha nokuhlela ubufakazi: ekusesheni amakhodi nasekuphenduleni imibuzo kusizindalwazi solwazi senkampani (0.929), kuthatha indawo ye-reranker yendabuko futhi kunikeza amaphuzu amadokhumenti angamakhandidethi ekudluleni okukodwa okuya phambili.
Ukuqinisekisa iziphetho zokucabanga: kusetshenziswa emisebenzini yohlobo lwe-NLI ukuze kunqunywe ukuthi “ubufakazi buyasisekela yini isiphetho”; ngemva kokulinganiswa, ukuzethemba kungaqhuba ngokuqondile ukuzenzakalela okusezansi.