E-Commerce Bench Uyasebenza

I-E-Commerce Bench iyisisekelo sokuhlola ukusebenza kwe-e-commerce kwesikhathi eside, esikhishwe njengomthombo ovulekile ngu-Alibaba Tongyi Qianwen ngokubambisana ne-Taotian Group. I-E-Commerce Bench ivumela i-LLM Agent ukuthi iphathe isitolo esiku-inthanethi izinsuku ezingu-365 ngemali yokuqala engu-100,000 yuan emakethe elingisa umhlaba wangempela, okuhlanganisa ukukhetha imikhiqizo, ukuxoxisana nabahlinzeki, ukubeka amanani, kanye nokuphatha isitoko nokugeleza kwemali. Indawo isekelwe kudatha yesikhulumi sangempela, inabahlinzeki abangu-576, okuhlanganisa abakhwabanisi, amaphromoshini angu-8 kanye nezehlakalo zemakethe ezingu-10. I-kernel enqumayo iqinisekisa ukuphindaphindeka kwemiphumela, kanti amakhono ebhizinisi emodeli ahlolwa ngezindlela eziyisikhombisa, okuhlanganisa inzuzo, ukuxoxisana, ukulwa nokukhwabanisa kanye nokufunda kwesikhathi eside.

Ukulingisa ukuphathwa kwesikhathi eside: I-LLM Agent iqhuba isitolo esiku-inthanethi izinsuku ezingu-365 emakethe eguqukayo ngemali yokuqala engu-100,000 yuan, ihlose ukukhulisa impahla ephelele ekupheleni konyaka.
Ukuphatha izitolo eziningi: Kusekela ukuvula izitolo ezifika kwezingu-4 ngesikhathi esisodwa, ezinhlotsheni ezingu-12 ezahlukene, ukuze kusungulwe ukuhlelwa kwemikhakha okuhlukahlukene.
Ukuxoxisana nabahlinzeki: Kuxoxiswana ngamanani imijikelezo eminingi nabahlinzeki abangu-576 ukuze kutholakale izindleko zokuthenga eziphansi.
Ukuhlonza ukukhwabanisa: Kuhlonzwa futhi kugwenywe izinhlobo ezingu-5 zemikhonyovu ejwayelekile evela kubahlinzeki abakhwabanisayo abangu-152: imali ye-VIP, ukungabikho kwesitoko, ukwehliswa kwekhwalithi, ukuphuthuma okungamanga kanye nezicupho zezaphulelo.
Ukubhekana nemakethe eguqukayo: Amasu okuphatha ayalungiswa phakathi kwamaphromoshini angu-8 nezehlakalo eziphuthumayo ezingu-10, ezihlanganisa izinhlekelele nokuphazamiseka kochungechunge lokuhlinzeka.
Ukuphatha isitoko nendawo yokugcina: Kuphathwa ukuthengwa, ukuthengiswa kanye nesitoko sama-SKU angu-6,886, kanye nezindleko zokugcina ezibalwa ngento nangosuku.
Ukulawula ukugeleza kwemali: Kuphathwa uhlelo lokukhokhelana olunama-akhawunti amathathu kanye nokubambezeleka kwezinsuku ezingu-9 ekubuyeni kwemali, kugwenywe isimo se-profitable-but-bankrupt, lapho ibhizinisi linenzuzo kodwa liqothuka.
Ukuhlola okuphelele ngezindlela eziyisikhombisa: Amaphuzu anikezwa ngokwehlukana ngenzuzo, ukuxoxisana, ukulwa nokukhwabanisa, amandla okukhokha izikweletu, ukusebenza kahle, ukwenziwa kanye nokufunda kwesikhathi eside.

Landela i-WeChat bese uphendula ngokuthi “umthombo ovulekile” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI omthombo ovulekile
Ukwakhiwa ngezendlalelo ezine: Uhlelo lwakhiwe ngokwehla kusukela kungqimba lomjikelezo we-Agent, ungqimba lwamathuluzi, ungqimba lwendawo kanye nongqimba lwedatha, kwakheka umjikelezo ophelele wesinqumo, ukwenziwa kanye nempendulo yendawo.
I-kernel enqumayo yokuxoxisana: Zonke izipesheli, ukwehliswa kwamanani kanye nezinqumo zokwamukela noma ukwenqaba zabahlinzeki zibalwa yi-kernel ye-algorithm engaguquki. I-LLM ibhekele kuphela ukuguqula izinqumo zezinombolo zibe izingxoxo zolimi lwemvelo. Lokhu kugcina umuzwa wangempela wokuxoxisana ngemijikelezo eminingi, kuqede ngokuphelele ukungahleliwe kokusampula, kuvimbele imodeli ekutholeni amanani angaphansi komugqa wezindleko ngokusebenzisa i-jailbreak noma inkohliso, futhi kuqinisekise imiphumela ephindaphindekayo kuma-Agent ahlukene uma kusetshenziswa isu elifanayo.
Injini yezomnotho eguqukayo: Isidingo sabasebenzisi asisampulwa ngokungahleliwe. Sib算wa ngokusekelwe esidingweni esiyisisekelo somkhakha, ekuguquguqukeni kwentengo, ezenzakalweni zemakethe kanye nedumela lesitolo; ukungahleliwe okuphindaphindekayo kulawulwa kuphela yimbewu ye-episode.
Ukukhokhelana ngama-akhawunti amathathu nengcindezi yokugeleza kwemali: Izindleko zikhishwa ku-akhawunti yasebhange ngalolo suku ezenzeka ngalo, kodwa imali yokuthengisa ibuyela ebhange kuphela ngemva kokuthi “ukuthunyelwa → isikhathi sokuqinisekisa sesikhulumi sezinsuku ezingu-9 → isikhwama sesikhulumi → ukukhipha imali ngesandla”. Lokhu kudala ngamabomu igebe lemali yokusebenza, kuphoqe i-Agent ukuthi ilawule imali etholakalayo futhi kugweme ukuziphatha okufushane okubheka inzuzo kuphela kunganakwa imali esesandleni.
Ukuhlola ukufunda kwesikhathi eside (AnchorRatio): Kukala ukuthi i-Agent iyakwazi yini ukuthenga umkhiqizo ofanayo ngentengo eyehla ngokuphindaphindiwe. Isetha ukuhlelwa kabusha okungahleliwe kwezingcaphuno zomlando ze-Agent uqobo njenge-null hypothesis; uma inani lehla kakhulu ngaphansi kuka-1.0, lokhu kubonisa ukuthi imodeli iyaqhubeka nokuthuthukisa isu layo lokuthenga ngokufunda kokuhlangenwe nakho kwangaphambilini kokuxoxisana.

Thola ikhodi: Klona inqolobane yomthombo ovulekile ye-E-Commerce Bench ku-GitHub uyibeke kukhompuyutha yangakini.
Ilungise indawo: Faka okuncikile kwephrojekthi futhi ulungise indawo yokusebenza ye-Python.
Xhuma imodeli: Faka i-API Key negama lemodeli ye-LLM ezohlolwa efayeleni lokumisa.
Qalisa ukulingisa: Qalisa uhlelo oluyinhloko; i-Agent izoqala ngokuzenzakalelayo umjikelezo wokuphatha i-e-commerce wezinsuku ezingu-365.
Buka imiphumela: Ekupheleni kokulingisa, uhlelo luzokhipha ngokuzenzakalelayo umbiko wokuhlola ngezindlela eziyisikhombisa kanye nejika lokushintsha kwempahla yonyaka wonke.

Ukuhlola ukuphathwa kwesikhathi eside okungokokuqala: Ngokokuqala ngqa, ukuhlolwa kwe-LLM Agent kunwetshwa emisebenzini emifushane kuye endaweni yebhizinisi eguqukayo yokuphatha okuqhubekayo izinsuku ezingu-365.
Kuqhutshwa yidatha yemakethe yangempela: Imikhiqizo engu-6,886, imikhakha engu-60 kanye nabahlinzeki abangu-576 kwakhiwe ngedatha yangempela, ehlanjululwe, yezikhulumi ze-e-commerce.
I-kernel enqumayo iqinisekisa ukuphindaphindeka: Izingcaphuno zokuxoxisana nesidingo semakethe kubalwa i-kernel ye-algorithm engaguquki; i-LLM ibhekele kuphela ukuguqula izingxoxo, kuqedwe ngokuphelele umsindo wokusampula okungahleliwe.
Uhlelo lokuhlola olunezindlela eziyisikhombisa: Alubheki kuphela impahla ekupheleni konyaka, kodwa luhlola ngokuzimela ukuxoxisana, ukulwa nokukhwabanisa, ukugeleza kwemali, ukusebenza kahle, ukwenziwa kanye nokufunda kwesikhathi eside.
Ukulingisa ingcindezi yangempela yebhizinisi: Ukukhokhelana okubambezelekile ngama-akhawunti amathathu kanye nezindleko zokugcina ezibalwa ngosuku kuphoqa i-Agent ukuthi iphathe imali yokusebenza, hhayi ukuphishekela inzuzo ebhaliwe kuphela.
Ukulinganisa ikhono lokulwa nokukhwabanisa: Kuhlanganiswe abahlinzeki abakhwabanisayo abangu-152 kanye nezinhlobo ezingu-5 zemikhonyovu ejwayelekile, ukuze kukalwe ngokunembile ikhono lemodeli lokuhlonza ubungozi nokuvikela imali.

Iwebhusayithi esemthethweni yephrojekthi：https://ecbench.github.io/
Inqolobane ye-GitHub：https://github.com/QwenLM/E-CommerceBench
Iphepha lobuchwepheshe le-arXiv：https://arxiv.org/pdf/2608.30730

Ukuhlola amakhono ebhizinisi amamodeli amakhulu: Ihlinzeka ngamamodeli aphambili afana ne-GPT, Claude kanye ne-Qwen ngokuqhathanisa okujwayelekile kwamakhono okuphatha esikhathi eside.
Ukuphindaphinda amasu e-e-commerce Agent: Isebenza njengendawo yokuqeqesha ye-reinforcement learning nezakhiwo ze-Agent, ukuze kuqinisekiswe ngokushesha amasu entengo, okuthenga kanye nesitoko.
Ucwaningo lwama-algorithm okulawula ubungozi bochungechunge lokuhlinzeka: Indawo elingisiwe enabahlinzeki abakhwabanisayo isetshenziselwa ukuhlola nokuthuthukisa amamodeli okukhetha abahlinzeki nokuhlonza ubungozi.
Ukuqinisekisa ama-algorithm entengo eguqukayo: Kusetshenziswa idatha yangempela yokuguquguquka kwentengo ukuze kuhlolwe ukusebenza kwenzuzo kwamasu entengo ahlukene ngesikhathi samaphromoshini nokuguquguquka kwezinkathi zonyaka.
Ukuqeqesha ngokulingisa izinqumo zebhizinisi: Njengomlingisi wokuphatha wesikole sebhizinisi noma wenkampani, kuqeqeshwa ikhono labaphathi lokwenza izinqumo ngaphansi kwengcindezi yokugeleza kwemali nokungaqiniseki kwemakethe.