书生-A1.5 Uyasebenza

书生-A1.5 (Sage-A1.5) iyi-ejenti yesayensi ehlanganisa amakhono ajwayelekile nawakhethekile enama-35B, eyethulwe yi-Shanghai AI Laboratory. Yakhelwe phezu kwamakhono esikhathi eside e-Agents-A1 kanye neplathifomu i-书生·端砚, futhi iqondiswe ekuqinisekiseni ucwaningo lwangempela lwe-RSI: ihlukanisa imisebenzi ngokuzenzakalela, ihlela ama-ejenti amancane amaningi ukuthi enze ucwaningo ngokuhambisana, bese iqala ukuqinisekiswa ukuze ibone izikhala ebufakazini futhi ilungise imicabango. 书生-A1.5 ihlanganisa izibonelo zesayensi yobuchopho, i-electrochemistry kanye nokubuyekezwa kwesayensi yasemkhathini. Iqhubeka ithuthuka ngokuhlolwa kwesayensi kwe-MVSP kanye nokubuyiselwa kwemiphumela ye-SFT/RL, futhi isekela ukuthunyelwa okulinganisiwe ku昇腾, 沐曦 kanye ne-AMD, igcizelela ucwaningo oluzimele lwesikhathi eside olulandelekayo nolungaqinisekiswa kabusha.

Ukuhlelwa kwama-ejenti: Ihlukanisa ngokuzenzakalela imisebenzi yocwaningo eyinkimbinkimbi, isebenzise ama-ejenti amancane amaningi ukucinga, iqhathanise ubufakazi futhi ihlanganise iziphetho ngokuhambisana.
Ukuqinisekiswa kwama-ejenti: Inquma ngokuzenzakalela isikhathi nento okufanele kuqinisekiswe, ihlukanise amaqiniso aqinisekisiwe emicabangweni esadinga ukuqinisekiswa, bese ilungisa ngokushesha indlela yokucinga noma isu locwaningo.
Ukwenza ucwaningo lwesikhathi eside: Ihlanganisa ukucinga, imicabango, ikhodi nokuqinisekisa kube umjikelezo ophelele, ilethe ikhodi, idatha, amashadi nemibiko okulandelekayo nokungaqinisekiswa kabusha.
Ukunika amandla izimo zesayensi: Iqondiswe emisebenzini efana nesayensi yobuchopho, i-electrochemistry nokubuyekezwa kwesayensi yasemkhathini, isekele imiphumela ehlelekile efana namamethriksi obufakazi, amathebula okuhlola namarekhodi ezibuyekezo.
Ukusekela ikhodi yocwaningo: Iqondiswe ekuthuthukiseni ama-algorithm nasekuthuthukiseni ikhodi yesayensi, ihlanganise izimo zokuhlola ikhodi ezifana ne-SWE-Bench Pro, Multilingual kanye ne-MLE-Bench-Lite.
Ukuthunyelwa kumahardware amaningi: Ihlinzeka ngezindlela eziningi zokulinganisa, isekele ama-NPU e昇腾, 沐曦, ama-GPU e-AMD nezinye, ukuze kube lula ukuhlolwa ngabalingani bocwaningo nabemboni.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo zamaphrojekthi e-AI avulekile
Isisekelo semisebenzi yesikhathi eside: I-Sage-A1.5 isebenzisa amakhono emisebenzi yesikhathi eside e-Agents-A1 njengesisekelo, ithembele epulatifomu yokutholwa kwesayensi i-书生·端砚. Ihlanganisa ukuhlela, ukwenza nokulungisa ezinqubweni zocwaningo ekuqeqeshweni, ukuze imodeli ingagcini ngokusebenzisa amathuluzi kuphela, kodwa ikwazi ukuhlanganisa amakhono akhona emisebenzini emide futhi ilungise ngoku动态kile izenzo ezilandelayo ngokusekelwe ebufakazini obusha.
Umjikelezo we-RSI oqhutshwa yizimo zangempela: Indlela esemqoka ithi “imvelo iyidatha”. Ukuhlolwa kwesayensi kusetshenziswa njengesiqondiso, kanti ukwakhiwa kwemvelo kunikeza impendulo, kwakha indlela eqhubekayo yokuthuthuka lapho ukuhlonzwa kwamakhono, ukwakhiwa kwezindawo zemisebenzi nokuqeqeshwa kwemodeli kukhuthazana. Lokhu kwenza imodeli iqoqe ulwazi lomkhakha futhi ilungise ukuchema kokucabanga phakathi nenqubo yangempela yocwaningo.
Indlela yokuqondanisa nokuhlolwa kwe-MVSP: Ithimba lethule i-MVSP, okuyisiphakamiso esincane sesayensi esiqinisekisekayo, futhi lakha ngokubambisana izisekelo zokuhlola ezisezingeni eliphezulu nezimo zesayensi ezinkulu ezingaphezu kweziyishumi, okuhlanganisa izindlela zobuchopho, i-high-energy physics, ukuklama izinto, ukwakhiwa kwemithi emisha, ubuhlakani bamakhemikhali nokuxilongwa kwezifo. Lokhu kuqondanisa indlela yokuthuthukiswa kwemodeli nezidingo zangempela zososayensi nemibuzo yesayensi engaqinisekiswa.
Ukukhiqizwa kwedatha nezinsiza zokuqeqesha: Iphrojekthi ikhiqiza ngokuqhubekayo idatha ye-SFT nezinsiza zokuqeqesha ze-RL ngokulingisa izindawo zokusebenza, ngokuthumela amaketanga amathuluzi okubala kwesayensi kanye nezindawo ze-sandbox zobungcweti. Ngesikhathi esifanayo, isebenzisa i-AI for AI ukuhlonza izithiyo ochungechungeni locwaningo, ithuthukise ukuhlelwa kwamathuluzi namasu okuqinisekisa e-Scientific Harness, futhi yandise ukusebenza kahle nokuzinza kwemisebenzi yesikhathi eside.
Indlela yokuqeqesha ezindaweni ezihlukahlukene nezinhlobonhlobo: Iqondene nokwehluka kokusatshalaliswa kwemisebenzi yezizinda ezihlukahlukene, izindlela zokucabanga kanye nezinhlaka zama-ejenti. Isebenzisa indlela yezigaba ezimbili ethi “ukulungiswa okuncane okugadiwe kwezizinda eziningi + ukuqeqeshwa kwangemuva okuhlanganisiwe”, ihambisana nepayipi elisebenza ngokungavumelani ngokuphelele kanye nendlela yokusampula ezivumelanayo, ukuze iqinise kanyekanye ukuhlola kwesayensi nekhono lokubhala ikhodi yobunjiniyela.

Finyelela indawo yokuzama: Vakashela ikhasi lokuzama leplathifomu i-书生·端砚 ku-https://discovery.intern-ai.org.cn/token-plan/home?tabIndex=0, ungene ngemvume bese ulandela imiyalelo yekhasi ukuze ufake isicelo noma usebenzise isabelo.
Khetha uhlobo lwemodeli: Khetha i-Sage-A1.5 eku-inthanethi noma inguqulo elandelayo elinganisiwe/evulekile ngokwamandla ekhompyutha; ezimweni ezizimele, khetha iphakheji yokuthunyelwa ehambisana ne昇腾 NPU, 沐曦 noma i-AMD.
Dala umsebenzi wocwaningo: Dala iphrojekthi entsha endaweni yokusebenza, ugcwalise umbuzo wocwaningo, izimo ezibophezelayo, umphumela olindelekile nokuthi kuyadingeka yini uchungechunge lobufakazi olulandelekayo.
Faka izinto nendawo: Layisha izincwadi, idatha, ikhodi noma amarekhodi okuhlola, ulungise uchungechunge lwamathuluzi okubala kwesayensi, i-sandbox kanye nemininingwane yolwazi noma amathuluzi angaphandle adingekayo.
Qalisa ukuhlelwa nokwenza: Vumela i-Sage-A1.5 ihlukanise umsebenzi ngokuzenzakalela, ihlele ama-ejenti amancane ukuthi enze ucwaningo ngokuhambisana, bese uvula ukuqinisekiswa kwama-ejenti ukuze ihlonze ngokuzenzakalela izikhala ebufakazini.
Ngenelela futhi ulungise phakathi nomsebenzi: Buka indlela yokucinga, izinguquko emicabangweni namarekhodi okuqinisekisa, unikeze imiyalelo yabantu mayelana nesiqondiso esiphambukile, imicabango ebalulekile noma amaphuzu okuhlola asadinga ukuqinisekiswa.
Buyekeza imiphumela: Hlola ikhodi, idatha, amashadi, imibiko kanye namathebula okuhlola/amamethriksi obufakazi/namarekhodi ezibuyekezo; uma kudingeka, cela ukuphindwa kokusebenza noma ukwengezwa kokuqhathanisa.
Thumela izinto zokulethwa: Khipha ikhodi ephindaphindekayo, imiphumela ephakathi, amashadi nemibiko, bese kwenziwa ukubuyekezwa kochwepheshe ngaphakathi kweqembu locwaningo ngaphambi kokusetshenziswa ephepheni noma ekuthathweni kwezinqumo zokuhlola.

Ukuthuthuka okuqhutshwa ucwaningo lwangempela: Ngomjikelezo we-RSI othi imvelo iyidatha, iqhubeka iqongelela ulwazi futhi ilungise ukuchema emisebenzini yesayensi yangempela.
Isisekelo esihlanganisa okujwayelekile nokukhethekile: Izuza amakhono emisebenzi yesikhathi eside e-Agents-A1, ngesikhathi iqinisa ukucabanga okukhethekile, ukuphatha ubufakazi nokusebenzisa amathuluzi ezimweni zesayensi.
Ikhono eliqinile lokuhlela ama-ejenti: Ingahlukanisa imisebenzi ngokuzenzakalela, ihlele ama-ejenti amancane amaningi ukuthi enze ucwaningo ngokuhambisana, futhi ilungele ukuhlola ucwaningo oluyinkimbinkimbi olunemikhondo nobufakazi obuningi.
Ukuqinisekisa nokulungisa okuzenzakalelayo: Iyakwazi ukunquma ukuthi kufanele kuqinisekiswe nini nokuthi kuqinisekiswe ini, ihlukanise amaqiniso emicabangweni, ibone izikhala ebufakazini futhi ilungise isu, ngaleyo ndlela inciphise ukuqoqeka kwamaphutha emisebenzini yesikhathi eside.
Ukuqondanisa ukuhlolwa nezidingo zesayensi: Nge-MVSP kanye nokwakhiwa ngokubambisana kokuhlolwa okuqinisekisekayo nezimo zesayensi ezinkulu ezingaphezu kweziyishumi, indlela yokuthuthukiswa kwemodeli isondela kakhulu ezinkingeni zangempela zososayensi.

Iwebhusayithi esemthethweni yephrojekthi: https://alpha-innovator.github.io/Sage-A1.5/
Iqoqo lamamodeli e-HuggingFace: https://huggingface.co/collections/InternScience/sage-a15
Iphepha lobuchwepheshe: https://arxiv.org/pdf/2610.28578
Idemo yokuzama ku-inthanethi: https://discovery.intern-ai.org.cn/token-plan/home?tabIndex=0

Ukuhlaziywa kwedatha yesayensi yezinzwa: Isebenzisa idatha yobuchopho obuphelele begundane ukuhlaziya izinkomba zokuhlukanisa izifunda zobuchopho, i-PCA, ukuhlolwa kwe-permutation kanye nokubonisa, isize ukuthola amaphethini okusabalala kolwazi lomsebenzi noma lokunyakaza.
Ukuthuthukisa ukuhlanganiswa kwe-electrochemistry: Ihlanganisa izincwadi nezivivinyo zangaphambilini, ithuthukise ngokuphindaphindiwe izimo zokusabela, imicabango yendlela yokusabela kanye nokubikezela isivuno, isize emisha yemizila yokuhlanganisa efana ne-Friedel–Crafts alkylation.
Ukubuyekezwa nokuhlolwa kwezifinyezo zesayensi yasemkhathini: Ihlela ama-ejenti amancane amaningi ukuthi ahlole izitatimende zemisebenzi yasemkhathini, ahlukanise amazinga obufakazi, futhi ikhiphe amathebula okuhlola, amamethriksi obufakazi namarekhodi ezibuyekezo.
Ukulungisa amaphutha ocwaningweni lwe-photochemistry nendlela yokusabela: Lapho ukucinga kungahambelani, inciphisa isiqondiso, ihlole izimiso zokusabela futhi imake izinyathelo ezibalulekile ezidinga ukuqinisekiswa ngokuhlolwa.
Ukuhlolwa kwe-high-energy physics, izinto, imithi emisha nokuxilongwa kwezifo: Isebenzisa i-MVSP ukwakha izindawo zesayensi eziqinisekisekayo, iqondanise ukuthuthukiswa kwemodeli nezidingo zangempela zososayensi.