Iris Uyasebenza

Iris iyi-ejenti yokusesha ekhishwe ngaphansi kwelayisensi evulekile yiqembu le-AllSpark lase-Xiaohongshu, enezinguqulo ezimbili: i-35B (Iris-mini) ne-397B (Iris-pro). U-Iris anganquma ngokuzimela ukuthi usesheni, afunde kanjani, nokuthi aqede nini ukuphendula kuwebhu yangempela. Wakha idatha yokuqeqesha enzima kodwa exazululeka ngokusebenzisa izixhumanisi zamakhasi ewebhu, futhi usebenzisa indlela yokuqeqesha ye-SFT–RL Climbing eshintshanisa i-SFT nokuqinisa ukufunda. Kumikhiqizo emine i-BrowseComp, DeepSearchQA, HLE neminye, uthole imiphumela ehamba phambili phakathi kwamamodeli avulekile anosayizi ofanayo. Inguqulo ye-35B isisondele kumamodeli anamapharamitha ayitriliyoni, kanti amandla okusesha angadlulela emisebenzini efana nokubiza amathuluzi ajwayelekile.

Ukusesha okuzenzakalelayo ku-inthanethi: Unquma ngokwakhe ukuthi usesheni, afunde kanjani imiphumela yokusesha, nokuthi aqhubeke noma ame nini, kuyilapho ecabanga futhi ephendula.
Ukucabanga kwezinyathelo eziningi emakhasini ahlukahlukene: Angakwazi ukunciphisa kancane kancane usesho lwamakhasi amaningi aze athole impendulo eyodwa, kunokuthembela ekufaniseni amagama angukhiye.
Ukusesha okuyinkimbinkimbi ngesiShayina nangesiNgisi: Ku-BrowseComp, BrowseComp-ZH nakweminye imisebenzi yokusesha amakhasi esiShayina nesiNgisi, uthole imiphumela ehamba phambili phakathi kwamamodeli avulekile anosayizi ofanayo.
Imibuzo nezimpendulo ezihlanganisa ubufakazi: Lapho ephendula, uhlanganisa ngokuphelele uchungechunge lobufakazi, futhi wenza kahle kakhulu ku-DeepSearchQA.
Ukuxazulula izinkinga ezisezingeni lochwepheshe: Angakwazi ukuphatha izinkinga ezinzima ezifana ne-HLE ezidinga ukucabanga kwezinga lochwepheshe.
Ukuphathwa komongo: Usekela amasu okuphatha umongo njengokunqampuna nokuminyanisa, ukuze usesho olude lungaphazamiseki; amamodeli amancane azuza kakhulu kulokhu.
Ukudlulela kwamanye amakhono: Ngaphandle kokuqeqeshwa okukhethekile, angakwazi ukwenza imisebenzi efana nokubiza amathuluzi ajwayelekile nokusebenzisana emsebenzini.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI avulekile
Ukwakhiwa kwedatha ngokudala imibuzo emuva: Kusetshenziswa ikhasi lembewu njengempendulo, izixhumanisi zalo zilandwa ukuze kwakhiwe igrafu yezinto ezihlobene. Kule grafu kwakhiwa imibuzo edinga ukucabanga ngezinyathelo eziningi emakhasini ahlukahlukene. Izinto ezikumbuzo ezingeyona impendulo ziguqulwa zibe izinkomba ezichazayo ukuze kususwe imikhondo efanisekayo. Ekugcineni, imodeli eyireferensi yenza ukuhlunga okukabili; kugcinwa kuphela imibuzo imodeli engakwazi ukuyiphendula ngaphandle kwencwadi kodwa ekwazi ukuyiphendula ngokunembile uma inikezwe ubufakazi. Yonke le nqubo izenzakalela futhi ingakhula ngokuhambisana nobukhulu bedatha.
Ukuqeqeshwa okushintshanisa i-SFT–RL Climbing: Esigabeni se-SFT, imodeli kathisha enamandla ikhiqiza izindlela zokusebenza isebenzisa amathuluzi okusesha angempela, bese kuqeqeshwa ngemva kokuhlunga imiphumela nokuhlunga kwesinyathelo esisodwa okwenziwa “umahluleli”. Esigabeni se-RL, imodeli ifunda ngenkathi isesha kuwebhu yangempela. Ekupheleni komjikelezo ngamunye, izindlela ezinzima kakhulu ukuzilungisa nezisebenza kahle kakhulu zibuyiselwa ku-SFT yomjikelezo olandelayo. Njengoba imodeli iba namandla, imibuzo iba nzima ngokuzenzakalelayo, kwakheke ikharikhulamu ezivumelanayo.
Ukuhlanganiswa kobunjiniyela bokuqeqesha: Ukumaka amaphuzu nokwenza izifinyezo kwenziwa ngamamodeli akhelwe ngaphakathi kweqoqo lokuqeqesha, ngaphandle kokuncika kunoma iyiphi i-API yangaphandle. Lokhu kugwema ukuphazamiseka kokuguquguquka kwenethiwekhi emjikelezweni wokuqeqesha. Kusetshenziswa nesifinyezi esifanayo ngesikhathi sokuqeqesha nangokusetshenziswa ku-inthanethi, kususwe umehluko phakathi kokuqeqesha nokucabanga.
Ukuqhubeka kusukela endaweni yokuma emizileni emide kakhulu: Izikhathi ezimbalwa ezinde kakhulu azisalahlwa ziphelele. Inqubo iyaphazanyiswa ezingeni lesicelo, bese isinyathelo esilandelayo siqhubeka kusukela esiqeshini esivele sithunyelwe, okugcina ukusetshenziswa kwe-GPU kuphezulu ngaphansi kokuhlela okuvumelanayo.
Umthetho wokuhlola ohambisana nezimo: Zonke izinhlelo eziqhathaniswayo zihlolwa ngamathuluzi afanayo, umongo ofanayo, imodeli efanayo yokumaka amaphuzu, kanye nesu elifanayo lokuphatha umongo, kusetshenziswa i-ejenti eyodwa ye-ReAct kuphela. Lokhu kususa amaqhinga afana nama-ejenti abaningi nokuziqinisekisa ngesikhathi sokuhlola, ukuze amaphuzu abonise ngempela amakhono emodeli.

Clone ikhodi yekhosombe: Sebenzisa git clone https://github.com/AllSpark-Research/Iris ukuze uthole ikhodi yokucabanga nokuhlola.
Landa izisindo zemodeli: Kuqoqo le-HuggingFace huggingface.co/collections/AllSpark-Research/iris, landa izisindo ze-Iris-mini (35B) noma i-Iris-pro (397B).
Lungisa indawo yokusebenza: Faka okuncikile okudingekayo ngokwezidingo zekhosombe, njengezinhlaka zokucabanga i-vLLM/SGLang, bese ulungisa isixhumi samathuluzi okusesha. Ukusesha okuxhumeke ku-inthanethi kuyisidingo esibalulekile se-Search Agent.
Layisha imodeli bese uyixhuma kumathuluzi: Layisha imodeli ngendlela ye-ejenti ye-ReAct, bese ubhalisa injini yokusesha njengethuluzi elingabizwa yimodeli.
Buza umbuzo bese uqala ukucabanga: Faka umbuzo; imodeli izohlela ngokuzenzakalelayo imibuzo yokusesha, ifunde amakhasi ewebhu, icabange ngezinyathelo eziningi, bese inikeza impendulo enobufakazi.

Imiphumela evulekile ehamba phambili ngosayizi ofanayo: Ku-BrowseComp, BrowseComp-ZH, DeepSearchQA ne-HLE, u-Iris uhola amamodeli avulekile anosayizi ofanayo kuzo zonke izilinganiso zokusesha. Kwezinye izivivinyo, i-Iris-pro engu-397B idlula ngisho namamodeli anezinga lamapharamitha ayitriliyoni.
Usayizi omncane osusondele kumamodeli amakhulu kakhulu: I-Iris-mini engu-35B ithole amaphuzu angu-82.2 ku-BrowseComp, isisondele kumamodeli anenani lamapharamitha eliphindwe kaningi, njenge-Kimi-K2.6, okwenza amamodeli amancane akwazi ukuphonsela inselelo amakhulu emikhakheni ethile.
Ukwakhiwa kwedatha okuzenzakalelayo nekhwalithi ephezulu: Imibuzo idalwa emuva kusuka ezixhumanisini zamakhasi ewebhu, imikhondo efanisekayo iyasuswa, bese kwenziwa ukuhlunga okukabili ukuze kuqinisekiswe ukuthi imibuzo inzima kodwa iyaxazululeka. Yonke le nqubo ayidingi ababhali bemibuzo futhi ingakhula ngokungenamkhawulo ngokuhambisana nedatha.
Ubunjiniyela bokuqeqesha obuzinzile nobusebenza kahle: Ukumaka amaphuzu nokwenza izifinyezo kuhlanganiswe eqoqweni lokuqeqesha, kususwe ngokuphelele ukuncika kuma-API angaphandle. Ukuqhubeka kwezindlela ezinde kusukela endaweni yokuma kugcina ama-GPU esebenza kakhulu futhi kunciphise kakhulu ukungaqiniseki kokuqeqesha.
Ukuvumelana phakathi kokuqeqesha nokusetshenziswa ku-inthanethi: Isifinyezi esicindezela okuqukethwe kokusesha ngesikhathi sokuqeqesha, ekuhloleni nasekusetshenzisweni ku-inthanethi siyafana, ngakho awukho umehluko phakathi kokuqeqesha nokusebenzisa.
Ukudlulela kwamanye amakhono nokusebenza okubanzi: Amandla okusesha adlulela emisebenzini engazange iqeqeshwe ngokukhethekile, efana nokubiza amathuluzi ajwayelekile nokusebenzisana emsebenzini. I-Search iyikhono eliyisisekelo elingasetshenziswa kabusha.

Ikhosombe le-GitHub: https://github.com/AllSpark-Research/Iris
Iqoqo lamamodeli e-HuggingFace: https://huggingface.co/collections/AllSpark-Research/iris
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2609.04304

Ucwaningo olunzulu nemibiko yocwaningo: Usesha ngokuzenzakalelayo imizuliswano eminingi, acabange ngezinyathelo eziningi emakhasini ahlukahlukene, futhi agcwalise uchungechunge lobufakazi. Lokhu kulungele imikhiqizo yocwaningo olunzulu efana nocwaningo lwezimboni, ukuhlaziywa kwezimbangi kanye nocwaningo lwezemfundo.
Ukuqinisekisa amaqiniso ayinkimbinkimbi: Impendulo idinga ukulandelela amakhasi amaningi ngezinyathelo eziningana ukuze kutholakale umphumela owodwa. Lokhu kulungele ukuphikisa ulwazi olungamanga, ukuqinisekisa amaqiniso nokuhlola ngokucophelela, lapho “kudingeka kutholakale umthombo”.
Ukusesha ngezilimi ezimbili, isiShayina nesiNgisi: Ukusebenza okuhle ku-BrowseComp naku-BrowseComp-ZH kwenza u-Iris afaneleke ngokwemvelo ekusesheni ulwazi oluwela imingcele olusebenzisa idatha exubile yesiShayina nesiNgisi.
I-ejenti ejwayelekile yokubiza amathuluzi: Amakhono adlulela emisebenzini ye-General Tool Use efana ne-BFCL ne-τ-bench, ngakho u-Iris angasetshenziswa njengesisekelo esijwayelekile semikhiqizo yama-Agent.
Ukusebenzisana emsebenzini (Cowork): Angaqala ukusebenza emisebenzini yasehhovisi efana ne-OfficeQA ne-APEX ngaphandle kokuqeqeshwa okukhethekile, futhi afanele ukushumekwa kumsizi wasehhovisi ukuze asingathe imibuzo yezincwadi nemisebenzi yezinqubo.