Piccolo Uyasebenza

I-Piccolo iwuchungechunge lwamamodeli okwenza izinqumo olukhishwe njenge-open source yithimba le-AgentScope lika-Alibaba, olunezilinganiso ezintathu: 27B, 9B no-4B. I-Piccolo ikhipha ngokuqondile imiphumela yezinqumo ehlelekile namathuba, futhi ingafakwa ekuhambeni komsebenzi kwama-agent ekukhetheni amathuluzi, ekuhloleni izimvume, nasekubuyiseleni emuva ngemva kokwehluleka; ingasetshenziswa ekucubunguleni idatha nasezinqumweni zesikhathi sangempela emidlalweni. I-27B ifinyelela ku-92.21% ku-JevBench Public naku-68.59 ku-Decision Index yombhalo. Ithimba lakhe futhi lakhulula izilinganiso zezimo ezimbili, i-HarnessJevBench ne-DataJevBench. Ucwaningo nentuthuko kusebenzisa uhlaka lokuzithuthukisa oluphindaphindekayo, lapho ama-Agent ocwaningo ehlaziya imibuzo ephendulwe kabi futhi ephindaphinda idatha ukuze kuqhubeke ukuthuthuka.

Ukukhetha ithuluzi: Ngokusekelwe emsebenzini, ebufakazini nasezithiyweni, inquma ukuthi yiliphi ithuluzi elizobizwa phakathi kwamathuluzi akhethiwe, njengokukhetha uhlelo lwama-oda lapho kubuzwa usuku lokulethwa.
Ukuqondisa amamodeli: Uma amamodeli amaningi engawuqeda umsebenzi, ikhetha insiza ezowenza ngokucabangela isabelomali nezinye izici.
Ukuhlunga umongo nenkumbulo: Igcina ulwazi olubalulekile futhi ihlunge okuqukethwe okuphelelwe yisikhathi noma okungahlobene.
Ukuhlola izimvume nokwethembeka: Yenza izinqumo ngaphakathi kwemingcele yezimvume, njengokukhetha ukugcina uhlaka esikhundleni sokuluthumela ngokuqondile uma kuvunyelwe ukulugcina kuphela.
Ukuqinisekisa ukuqedwa nokubuyisela ngemva kokwehluleka: Iyaqhubeka nokuqinisekisa lapho ubufakazi bunganele, futhi inquma ukuthi izophinda izame yini lapho kuba nokuphazamiseka kwesikhashana.
Izinqumo zokucubungula idatha: Inquma ukulandelana kokuhlanza nokususa okuphindaphindiwe, ikhetha indlela yokufingqa, icela ulwazi olungekho, futhi inquma ukuthi idatha idinga yini ukufihlwa ngaphambi kokuthunyelwa ngaphandle.
Izinqumo zesikhathi sangempela emidlalweni: Ngokusekelwe esithombeni somdlalo, inikeza ukuqondisa kwezinga lokhiye kwezenzo eziqhubekayo njengokunyakaza, ukuhlasela nokubalekela, futhi isekela imidlalo yezenzo, yamaphazili neyebhodi.

Landela i-WeChat ukuze uthole impendulo ethi “开源”, bese ujoyina iqembu lezingxoxo lamaphrojekthi e-AI open source
Indlela yesixhumi esibonakalayo sezinqumo: I-Piccolo iyingxenye ekhethekile yokwahlulela, ethola “isimo, umbuzo, izinketho” njengokufakwayo. Isebenzisa ukunaka kwe-prompt okuqondiswa nhlangothi zombili ukuze umbuzo nezinketho kubhekisise omunye komunye, bese ifunda ngokuqondile amathuba enketho ngayinye kuma-logits ama-token ezinketho, ikhiphe izinqumo ezihlelekile ezingasetshenziswa ngokuqondile uhlelo. Ayikhiqizi umbhalo nhlobo kuyo yonke le nqubo. Isebenzisa i-endpoint ehambisana ne-Jev ethi /v1/systemone ngokwemvelo, futhi ihambisana nokuhlelwa kokufakwayo kwe-flex_twice ukuze kuthuthukiswe indlela yokwenza inferensi.
Ubunjiniyela bedatha obuhlanganisiwe: Idatha yokuqeqesha iqala kumasampula ayizigidi ezingu-6.65 aqoqwe emphakathini we-open source, ahlanganisa ukuhlukanisa, ukusesha, ukucabanga kwezibalo, ukuhlola okuthandwayo nokusebenzisa amathuluzi. Yonke idatha ihlelwa ngendlela yesinqumo ethi “isimo, umbuzo, izinketho, ukusatshalaliswa okuqondiwe”, kanti amalebula aqinile nathambile asebenzisa isixhumi esibonakalayo esifanayo sokufundisa. Ithimba lakha uhlelo lwamalebula encazelo olunezilinganiso ezingu-13 zokumaka namanani achazwe kusengaphambili angu-551 lisebenzisa i-Data-Juicer. Yonke idatha iyasuswa okuphindaphindiwe ngaphambi kokungena echibini, futhi ihlolwe ukuze kukhishwe ukunqwabelana nezisekelo zokuhlola, kuqinisekiswe ukuhlukaniswa kokuqeqesha nokuhlola.
Umjikelezo wokuzithuthukisa ophindaphindekayo: Ucwaningo nentuthuko luqhutshwa yi-Research Agent, kanti uphawu oluyinhloko ukulahleka kwe-KL kombuzo ngamunye. Ngemva kokuhlanganisa izingcezu ngamalebula encazelo, idatha enokuphambuka okuphezulu okuqhubekayo iyabuyekezwa izimpendulo, kuseshwe izinto ezifanayo noma kwenziwe izinhlanganisela zezimo ezengeziwe ngokuqondile. Amasampula afundwe ngokwanele nacishe afane ayehliswa ngesampula noma kugcinwe aphindwe njengabameleli. Ukukhethwa kwezilinganiso, ukuhlolwa kokubhalwa kabusha, ukwakhiwa okuqondisiwe nokwehliswa kwesampula kuhlanganiswa kube izinhlelo zokuhlola ezahlukene. I-Agentic Data Engineering ihlela ama-Agent amaningi ukuthi aqhubeke ngokulandelana “ukuklama uhlelo→ukwakha ukuhlola→impendulo yemiphumela→ukubuyekeza ukuxilonga”. I-DataJuicer-Sandbox ihlela ukucutshungulwa kwedatha, ukuqeqesha nokuhlola kube ukuhamba komsebenzi okusebenzisekayo; umuntu ubeka kuphela umgomo, isabelomali nobubanzi bedatha.

Ukuhlola ku-inthanethi: Vakashela iwebhusayithi esemthethweni ethi https://decisionplayground.agentscope.io/ ukuze uyizame ngokuqondile. Akudingeki ulande lutho ukuze uzwe ukusebenza kwe-Piccolo emisebenzini ehlukahlukene yokwahlulela.
Ukulanda imodeli: Ku-Hugging Face, khetha ngokwesidingo izisindo ezigcwele ze-4B/9B/27B noma i-GGUF yamazinga amathathu i-BF16, Q8_0 ne-Q4_K_M. I-GGUF idinga ukuthi zonke izingcezu zilandwe ziye kufolda efanayo.
Ukusebenza endaweni: Faka i-QwenPaw kanye ne-runtime ye-llama.cpp ehambisana ne-Piccolo usuka kukhodi yomthombo. Ku-Models → QwenPaw Local → Decision models, landa i-GGUF bese uchofoza u-Start; bese ku-Default models uqinisekisa ukuthi imodeli ezenzakalelayo yezinqumo ikhombe kuyo.
Ukuvula ukuhlolwa okuzenzakalelayo: Kumenyu yemodeli yeseshini, vula i-Use decision model, bese kuzilungiselelo zokugunyazwa kwamathuluzi ukhetha i-Decision Model Mode. I-Piccolo izohlola ngokuzenzakalelayo ukubizwa kwethuluzi ngakunye, kugcinwe nokubuyela ekuvunyweni ngumuntu.
Ukuvula ukucindezelwa komongo: Ngemva kokuvula i-Context Compaction, vula i-Decision Model Compaction. I-Piccolo izonquma ngaphansi kwengcindezi yomongo ukuthi yimiphi imiphumela yamathuluzi enokuhlobana okuphansi futhi engacindezelwa.
Ukuhlanganisa nge-API: Uma udinga ukuyifaka ekuhambeni komsebenzi ozakhele kona, bheka ukuqaliswa kwereferensi kokwenza inferensi kufolda ye-inference/ kanye nesibonelo sesicelo esincane ku-examples/. Biza i-endpoint yomdabu ethi /v1/systemone, uthumele izicelo ngefomethi ethi “isimo, umbuzo, izinketho”, bese uhlaziya izinqumo namathuba ahlelekile.
Ukubuka okurekhodiwe kwemiboniso: Ngemva kokuklona inqolobane, sebenzisa i-Arena esebenzisanayo ngaphansi kwe-arena/. Ungamisa, uhudule futhi ubuke ngamunye amathuba ezinketho azo zonke izinqumo ezirekhodiwe eziyisishiyagalolunye, okuhlanganisa izimpi ze-《绝区零》.

Wonke amasayizi avulekile futhi ilayisensi ivulekile: Amamodeli angu-4B, 9B no-27B, kanye nezisindo zawo, akhishwe ngaphansi kwe-Apache-2.0, futhi kunikezwa izisindo ezigcwele ne-GGUF yamazinga amathathu, okuhlanganisa izimo zokusetshenziswa kusukela efwini kuya endaweni.
Ukusebenza okuvelele ngosayizi omncane: I-4B ifinyelela ku-87.01% ku-JevBench Public naku-58.00 ku-Decision Index, idlula izinga le-9B lezimbangi eziningi futhi ihambisana nezidingo zamamodeli ezinqumo asetshenziswa kaningi nezibiza kancane.
Imiphumela yokuhlola ehamba phambili: I-27B ifinyelela ku-92.21% ku-JevBench Public naku-68.59 ku-Decision Index. Ekuqhathaniseni kwamamodeli angu-11 eqembu, ihamba phambili ku-HarnessJevBench, DataJevBench ne-JevBench.
Isixhumi esibonakalayo sezinqumo esinokubambezeleka okuphansi: Ifunda ngokuqondile ama-logits ama-token ezinketho ukuze ibuyise imiphumela namathuba ahlelekile. Ngokuhambisana nokunaka kwezinhlangothi zombili nokuhlelwa kwe-flex_twice, ifanele ngokwemvelo izigaba zokwahlulela ezisetshenziswa kaningi ekuhambeni komsebenzi.
Ukuhlanganiswa okulula ne-ecosystem ka-Alibaba: Imodeli igcinwa ku-ModelScope, isivele ihlanganiswe nomugqa oyinhloko we-QwenPaw, futhi ingavula ngokuchofoza okukodwa ukuvunywa kwamathuluzi okuzenzakalelayo nokucindezelwa komongo. Uhlaka lwe-AgentScope luvela emthonjeni ofanayo, okwenza indlela yokuqala ibe mfushane.
Ubunjiniyela obugcwele obuvulekile: Ipayipi ledatha lokuzithuthukisa oluphindaphindekayo, i-HarnessJevBench ne-DataJevBench zonke zikhishwe njenge-open source, ukuze umphakathi ukwazi ukuziphindaphinda nokunikela.

Iwebhusayithi yephrojekthi: https://decisionplayground.agentscope.io/
Inqolobane ye-GitHub: https://github.com/agentscope-ai/Piccolo

Ukuvunywa okuzenzakalelayo kwamathuluzi e-Agent: Ifakwa ku-QwenPaw noma ku-runtime ye-Agent ozakhele yona ukuze yahlulele ukubizwa kwethuluzi ngakunye, ivumele imisebenzi yokufunda kuphela futhi ivimbe imiyalo ebhubhisayo njengokubhala phezu kwedatha, esikhundleni sokuvunywa komuntu ngakunye.
Ukucindezelwa komongo nokuhlunga inkumbulo: Lapho ingcindezi yomongo ikhula, inquma ukuthi yimiphi imilayezo yakudala nemiphumela yamathuluzi enokuhlobana okuphansi, bese inquma ukuthi icindezelwe noma ilahlwe, igcine ulwazi olubalulekile.
Izinqumo zokuhamba komsebenzi wokucubungula idatha: Yenza izinqumo ezigabeni zokungena, ukuhlanza, ukufihla nokuletha idatha, njengokunquma ukulandelana kokuhlanza nokususa okuphindaphindiwe, ukucela ulwazi olungekho lapho kungekho nkambu yentela, nokunquma ukuthi i-imeyili idinga ukufihlwa yini ngaphambi kokuthunyelwa ngaphandle.
Izimvume zenqubo yebhizinisi nokuqinisekisa: Ekulawuleni ukwenziwa yi-Agent, inquma ukuthi isenzo sidlula yini izimvume, ukuthi ubufakazi bomsebenzi banele yini ukumemezela ukuqedwa, nokuthi kuphindwe kuzanywe yini uma kuba nokwehluleka kwesikhashana.
Izinqumo zesikhathi sangempela ezindaweni zemidlalo nezisebenzisanayo: Ngokusekelwe esithombeni somdlalo, inikeza ukuqondisa kwezenzo ezingeni lokhiye ngokunyakaza, ukuhlasela nokubalekela komlingiswa. Ingasetshenziswa futhi emidlalweni yamasu efana ne-othello ne-sokoban edinga ukwahlulelwa kwesimo.