Ukufingqa komhlangano kanye nokukhishwa kwento yesenzo nge-Amazon Nova
Imihlangano idlala indima ebalulekile ekwenzeni izinqumo, ukuxhumanisa kwephrojekthi, kanye nokubambisana, kanye nemihlangano ekude kuvamile ezinhlanganweni eziningi. Kodwa-ke, ukuthwebula nokuhlelwa kwehlela ukubekezelela okuphambili kulezi zingxoxo kuvame ukungasebenzi futhi akuhambelani. Ukufingqa imihlangano noma ukukhishwa kwezinto zesenzo kudinga umzamo obalulekile futhi kuthambekele ekushisweni noma ekuchazeni kabi.
Amamodeli amakhulu olimi (i-LLMS) anikele ngesixazululo esiqine ngokwengeziwe ngokuguqula okubhaliwe okungekho emthethweni kube izifingqo ezihleliwe nezinto zesenzo. Lokhu kunamandla kusebenza kakhulu ekuphathweni kwamaphrojekthi, ukusekelwa kwamakhasimende kanye nezingcingo zokuthengisa, ezomthetho nokuhambisana, nokuphathwa kolwazi lwebhizinisi.
Kulokhu okuthunyelwe, sethula uphawu lwamamodeli wokuqonda ahlukile avela emndenini we-Amazon Nova etholakala e-Amazon Bedrock, ukuhlinzeka ngokuqonda ukuthi ungayikhetha kanjani imodeli engcono kakhulu yomsebenzi wokufingqa umhlangano.
I-LLMS ukukhiqiza ukuqonda okuhlangene
I-LLMS yanamuhla isebenza kakhulu ekufingwini nasekukhishweni kwento yesenzo ngenxa yekhono labo lokuqonda, ubudlelwano besihloko esine-infer, bese ukhiqiza imiphumela ehlelekile. Kula macala okusebenzisa, ubunjiniyela obusheshayo buhlinzeka ngendlela esebenza kahle futhi ekhubazekile uma kuqhathaniswa nemodeli yendabuko yokuhleleka kwemodeli yendabuko noma ukwenza ngokwezifiso. Esikhundleni sokushintsha ubuciko bemodeli obuyisisekelo noma ukuqeqeshwa kumininingwane emikhulu ebiyelwe, ubunjiniyela obusheshayo busebenzisa imibuzo eyenziwe ngokucophelela ukuqondisa indlela yokuziphatha yemodeli, inomthelela ngqo ifomethi nokuqukethwe. Le ndlela ivumela ngokwezifiso okusheshayo, okuqondene nesizinda ngaphandle kwesidingo sezinqubo zokubuyiselwa kwezinsizakusebenza. Ngemisebenzi efana nokufingqa komhlangano kanye nokukhishwa kwento yesenzo, ukuhwebisisa okusheshayo kwenza amandla okulawula okuqondile ngokuphuma okukhiqizwayo, uqiniseke ukuthi ahlangabezana nezidingo ezithile zebhizinisi. Ivumela ukulungiswa okuguquguqukayo kwezikhuthazo ukuze kuvumelane namacala okusebenzisa, okwenza kube yisixazululo esifanelekile sezindawo ezinamandla lapho izimo zemodeli zidinga ukuguqulwa kabusha ngaphandle kokudlula kwemodeli enhle.
Amamodeli we-Amazon Nova ne-Amazon Bedrock
Amamodeli we-Amazon NOVA, angembulwa ku-AWS Re: Ukusungula ngoDisemba 2024, kwakhiwa ukuletha ubuhlakani obuphambili ekusebenzeni kwamanani entengo embonini. Baphakathi kwamamodeli asheshayo futhi abiza kakhulu ama-tiers abo e-intelligence, futhi alungiselelwe izicelo zebhizinisi le-Power Enterprise ngendlela ethembekile, evikelekile nengabizi.
Umndeni wemodeli yokuqonda unama-tiers amane amamodeli: I-Nova Micro (Umbhalo-kuphela, i-Ultra-esebenza ngokuguquguquka), i-Nova Premier (i-multimodal, imodeli ye-nova (i-multimodal, imodeli egula kakhulu ye-NOVA (i-multimodal, imodeli egula kakhulu ye-NOVA (i-multimodal, imodeli egula kakhulu yeNova (i-multimodal, imodeli egula kakhulu ye-NOVA ( Amamodeli we-Amazon NOVA angasetshenziselwa imisebenzi ehlukahlukene, kusuka ekufingwinisweni kuya esizukulwaneni esihlelekile sombhalo. Nge-distillation yemodeli ye-Amazon Bedrock, amakhasimende angabuyisela ubuhlakani beNova uNova Premier kwimodeli esheshayo neyabiza kakhulu efana ne-Nova Pro noma i-Nova Lite yecala lazo lokusebenzisa noma i-Domain. Lokhu kungatholakala nge-Amazon Bedrock Console nama-APIs anjenge-API eguqukayo futhi acele i-API.
Ukubuka konke
Lokhu okuthunyelwe kukhombisa ukuthi ungawasebenzisa kanjani amamodeli wokuqonda ama-Amazon Nova, atholakala nge-Amazon Bedrock, ngokuzenzakalela okuzenzakalelayo kokuqonda kusetshenziswa ubunjiniyela obusheshayo. Sigxile kokuphuma kokuqala okubili:
- Ukufingqa komhlangano – Isifinyezo esisezingeni eliphakeme esisebenzayo esifinyelela amaphuzu asemqoka engxoxo, izinqumo ezenziwe, nezibuyekezo ezibucayi ezivela kumbhalo womhlangano
- Izinto zesenzo – Uhlu oluhlelekile lwemisebenzi esebenzayo ethathwe engxoxweni yomhlangano esebenza kulo lonke iqembu noma iphrojekthi
Umdwebo olandelayo ukhombisa ukusebenza kwezixazululo.
Izimfuneko
Ukulandela kanye nalokhu okuthunyelwe, ukujwayelana nokubiza ama-LLMS usebenzisa i-Amazon Bedrock kulindeleke. Ngezinyathelo ezinemininingwane ekusebenziseni i-Amazon Bedrock yemisebenzi yokufingqa kombhalo, bheka ukwakha uhlelo lokusebenza lwe-AI SMINTRALIZER nge-Amazon Bedrock. Ngemininingwane eyengeziwe mayelana nokushayela ama-LLMS, bheka ku-API enxusa i-API bese usebenzisa amadokhumenti wereferensi we-API aguqukayo.
Izixazululo Zoxazululo
Sithuthukise izici ezimbili eziyisisekelo zeSifing Sesimo Sezixazululo kanye nokukhishwa kwento yesenzo – ngokusebenzisa amamodeli athandwayo atholakala nge-Amazon Bedrock. Ezingxenyeni ezilandelayo, sibheka okushukumisayo okusetshenziselwa le misebenzi ebalulekile.
Ngomsebenzi wokufingqa umhlangano, sasebenzisa isabelo somuntu, sikhuthaza i-LLM ukukhiqiza isifinyezo ngaphakathi
Ngomsebenzi wesenzo sokukhishwa kwento, sanikeza imiyalo ethile ekwakheni izinto zesenzo ezikhuphukayo futhi zisetshenziselwe ukucatshangelwa ukuthuthukisa ikhwalithi yezinto ezenziwe ngesenzo. Emlayezweni osizayo, isiqalo
Imindeni ehlukene yemodeli iphendula ngokukhuthaza okufanayo ngendlela ehlukile, futhi kubalulekile ukulandela umhlahlandlela wokuphakanyiswa ochazwe yimodeli ethile. Ukuthola eminye imininingwane ngemikhuba emihle kakhulu ye-Amazon Nova ekhuthaza, bheka imikhuba emihle kakhulu yama-Amazon Nova ukuqonda.
Isilinganiso sakwaDataset
Ukuze sihlole ikhambi, sasebenzisa amasampula wedathashi le-qmum. I-QMSUM Dataset iyinhlangano ebhekwe ekufingqiwe komhlangano, okubandakanya ukuhanjiswa kwezilimi zesiNgisi kusuka ezingxoxweni zezemfundo, zebhizinisi, kanye nezingxoxo zokubusa ezishonile. Ihlola i-LLMS ekwakheni izifingqo ezihlelekile, ezibumbayo ezingxoxweni eziyinkimbinkimbi futhi ezinemibala eminingi, okwenza kube yisisetshenziswa esibalulekile sokufingqiwe okukhona kanye nenkulumo ukuqonda. Ngokuhlola, sasebenzisa imihlangano engama-30 esampula ngezikhathi ezithile kusuka ku-QMUM Dataset. Umhlangano ngamunye wawuqukethe okubhaliwe okungama-2-5 ngezihloko – futhi kuqukethe amathokheni acishe abe ngu-8,600 ngombhalo ngamunye ngokwesilinganiso.
Uhlaka Lokuhlola
Ukuthola ukuphuma kwekhwalithi ephezulu kusuka ku-LLMS ekufingwini komhlangano kanye nokukhishwa kwento yesenzo kungaba ngumsebenzi oyinselele. Ama-metric ezokuhlola endabuko afana ne-rouge, i-bleu, kanye ne-meteor ukugxila ekufananeni okungaphezulu kweleveli phakathi kwezifinyezo ezikhiqizwayo nezifinyezo zereferensi, kepha zivame ukwehluleka ukuthwebula ama-nuances afana nokunemba, ukuhambisana, kanye nokusebenza. Ukuhlolwa komuntu kuyindinganiso yegolide kepha kuyabiza, kudla isikhathi, futhi akunakwena. Ukubhekana nalezi zinselelo, ungasebenzisa i-LLM-a-A-AMaji, lapho enye i-LLM isetshenziselwa ukuhlola ikhwalithi yemiphumela ekhiqizwayo esekelwe kwimibandela echazwe kahle. Le ndlela inikezela ngendlela ehlelekile futhi engabizi kakhulu yokushintsha ukuhlola ngenkathi ugcina ukunemba okuphezulu. Kulesi sibonelo, sasebenzisa i-anthropic's Claude 3.5 sonnet v1 njengemodeli yejaji ngoba sikuthole kuhambisana kakhulu nokwahlulela komuntu. Sisebenzise ijaji le-LLM ukuthola izimpendulo ezikhiqizwayo kuma-metric amathathu amakhulu: ukuthembeka, ukufingqa, nombuzo wokuphendula (QA).
Ukuthembeka kwesikolo kulinganisa ukwethembeka kwesifinyezo esikhiqizwayo ngokulinganisa ingxenye yezitatimende ezihlanganisiwe ngesifinyezo esisekelwa ngomongo onikezwe (ngokwesibonelo, umbhalo womhlangano) maqondana nenani lezitatimende.
I-Sumfialization Score yinhlanganisela ye-QA Score kanye ne-Consuseness Score enesisindo esifanayo (0.5). Isikolo se-QA silinganisa ukubikwa kwesifinyezo esikhiqizwayo kusuka kumbhalo womhlangano. Kuqala ukukhiqiza uhlu lombuzo futhi uphendule ngazimbili kusuka ekubhaleni umhlangano futhi kunezinyathelo ingxenye yemibuzo ebuzwayo lapho isifinyezo sisetshenziswa njengomongo esikhundleni somhlangano. Isikolo se-QA siyancoma ekuthembekeni kokwethembeka ngoba amamaki okuthembeka awalinganise ukumbozwa kwesifinyezo esikhiqizwayo. Sisebenzise kuphela i-QA Score ukukala ikhwalithi yesifinyezo esikhiqiziwe ngoba izinto zesenzo akufanele zimboze zonke izici zombhalo womhlangano. Isikolo esinciphisayo silinganisa isilinganiso sobude besifinyezo esikhiqizwayo esihlukaniswe ngobude bemibhalo ephelele yomhlangano.
Sisebenzise inguqulo eguquliwe ye-Smore yokwethembeka kanye ne-Sumfialization Score eyayine-latency ephansi kakhulu kunokuqalisa ukusebenza kwangempela.
Umphumela
Ukuhlola kwethu amamodeli we-Amazon Nova kulo lonkelelwa ekufenizekisweni kwemisebenzi kanye nemisebenzi yokukhishwa kwento eyeveze amaphethini acacile wokusebenza-latency. Ngokufingqa, uNova Premier wazuza amaphuzu aphezulu okuthembeka (1.0) anesikhathi sokucubungula ama-5.34s, kuyilapho uNova Pro eletha ukwethembeka okungu-0.94 ngo-2.9s. Amamodeli amancane we-NOVA Lite kanye noNova Ngokukhishwa kwento yesenzo, uNova Premier uphinde wahola ekuthembekeni (0.83) nesikhathi sokucubungula esingu-4,94s, kulandelwa nguNova Pro (0.8 Ukwethembeka, 2.03s). Kuyathakazelisa ukuthi, i-nova micro (0.7 Ukwethembeka, 1.43s) Ofterformed Nova Lite (0.63 Ukuthembeka, 1.53s) kulo msebenzi othile naphezu kosayizi wayo omncane. Lezi zilinganiso zinikezela ngemininingwane ebalulekile kwizici zejubane lokusebenza kulo lonke umndeni we-Amazon Nova Model for application. Amagrafu alandelayo akhombisa le miphumela. Lesi sikrini esilandelayo sibonisa umphumela wesampula womsebenzi wethu wokufingqa, kufaka phakathi isifinyezo somhlangano we-LLM okhiqizwayo kanye nohlu lwezinto zesenzo.
Ukugcina
Kulokhu okuthunyelwe, sikhombise ukuthi ungasebenzisa kanjani ukuyeka ukukhumbula ukuqonda okufana nezifingqo zomhlangano kanye nezinto zezenzo ezisebenzisa amamodeli we-Amazon Nova atholakala nge-Amazon Bedrock. Ngokufingqa umhlangano omkhulu womhlangano we-AI-spale, ukwenza kahle i-latency, izindleko, kanye nokunemba kubalulekile. Umndeni wakwa-Amazon Nova we-Light Models (Nova Micro, Nova Lite, Nova Pro, noNova uNova uNova uNova Premier) anikezela ngenye indlela esebenzayo kumamodeli aphezulu, ngcono kakhulu lapho kunciphisa izindleko zokusebenza. Lezi zinto zenza i-Amazon NOVA Ukukhetha okukhangayo kwamabhizinisi ukuphakamisa amavolumu amakhulu emininingwane yokuhlangana esikalini.
Ukuthola eminye imininingwane nge-Amazon Bedrock kanye namamodeli wakamuva we-Amazon NOVA, abhekisele kuMhlahlandlela Womsebenzisi we-Amazon Bedrock kanye nomhlahlandlela womsebenzisi we-Amazon Nova, ngokulandelana. Isikhungo se-AWS SENTRACTION INDENGAS INDLELA ineqembu lama-AWS Science neStrategy Ochwepheshe abanolwazi olubanzi lwe-AI AI, basiza amakhasimende abeke phambili amacala okusebenzisa, ukwakha izixazululo zomgwaqo, futhi ahambise izixazululo ekukhiqizeni. Bheka isikhungo se-Ai Innovation Center ye-AIS yezenye yezindaba zethu zakamuva nezindaba zekhasimende.
Mayelana nababhali
Baishali Chaudhury Ungusosayensi osetshenzisiwe esikhungweni se-Ai Innovation Center adventative aps, lapho agxile khona ekuthuthukiseni izixazululo ze-AI ezikhiqizayo zezinhlelo zokusebenza zangempela zomhlaba. Unesizinda esiqinile embonweni wekhompyutha, ukufunda ngomshini, ne-AI yokunakekelwa kwempilo. UBaishali uphethe i-PhD kwisayensi yekhompyutha evela e-University of South Florida nasePostDoc kusuka esikhungweni somdlavuza kaMoffitt.
I-Sungmin Hong Ingabe i-Senior isebenzisa usosayensi e-Amazon Generative Ai Innovation Center lapho asiza khona ukusheshisa amacala ahlukahlukene okusebenzisa amakhasimende ama-AWS. Ngaphambi kokujoyina ama-Amazon, iSungmin kwakungumuntu ocwaningisi lwangaphambili ucwaningo eHarvard Medical School. Uphethe Ph.D. Kusayensi yekhompyutha evela eNew York University. Ngaphandle komsebenzi, uzikhandla ngokugcina izitshalo zakhe zasendlini ziphila iminyaka engu-3 +.
Mengdie (flora) wang Ungusosayensi wedatha e-AWS Generative AI Innovation Center, lapho asebenza khona namakhasimende azokwakha kanye nokusebenzisa izixazululo ze-AI ezi-Scablerable AI ezibhekele izinselelo zazo ezihlukile. Ubheka amasu we-Model ngokwezifiso kanye nezinhlelo ze-Agent-based AI, asiza izinhlangano zokuhlanganisa ama-harness amandla aphelele wobuchwepheshe be-AI afe. Ngaphambi kwama-AWS, uFlora wathola iziqu zenkosi yakhe kwisayensi yamakhompiyutha evela e-University of Minnesota, lapho athuthukisa khona ubuchwepheshe bakhe ekufundeni komshini kanye nobuhlakani bokufakelwa.
U-Anila Joshi Ineminyaka engaphezu kweshumi yokwakha izixazululo ze-AI. Njengomholi we-AWSI GEO ku-AWS DECAMENT AI Innovation Center, anila Peoneers application annove app application ai ecindezela imingcele yezentengiso futhi asheshise ukwamukelwa kwezinsizakalo ze-AWS namakhasimende ngokusiza amakhasimende ai Solutions.
