Amandla ukuqeqeshwa kwakho kwe-LLM kanye nokuhlola kwe-sagemaker ai generative ai amathuluzi

Namuhla sijabule ukwethula Isikhundla Sombhalo na- Umbuzo Nempendulo Izifanekiso ze-UI kumakhasimende we-sagemaker ai. Le khasi Isikhundla Sombhalo Isifanekiso senza abafaka izichasiselo babantu ukuthi babeke izimpendulo eziningi ezivela kumodeli enkulu yolimi (LLM) ngokususelwa kwinqubo ngokwezifiso, njengokucacisa, ukucaciseleka, noma ukunemba okuyiqiniso. Le mpendulo ebalwa ihlinzeka ngokuqonda okubucayi okusiza ukucwengeka amamodeli ngokuqinisa ukufundwa kusuka ekufundeni komuntu (i-RLHF), ukudala izimpendulo ezihambelana kangcono nokuthanda komuntu. Le khasi Umbuzo Nempendulo Isifanekiso sisiza ukwakhiwa kwekhwalithi ephezulu ye-Q & ngababili ngokususelwa kumavesi wombhalo anikezwe. Lawa mabili asebenza njenge Imininingwane yokubonisa Ngokuqondisa ubuhle be-Fine-tuning (SFT), amamodeli wokufundisa ukuthi ungaphendula kanjani ekufakweni okufanayo ngokunembile.

Kulesi sifundo sebhulogi, sizokuhamba ngendlela yokusetha lezi zifanekiso e-sagemaker ukudala ama-datasets asezingeni eliphakeme ukuqeqesha amamodeli akho amakhulu olimi. Ake sihlole ukuthi ungakuphakamisa kanjani la mathuluzi amasha.

Isikhundla Sombhalo

Isifanekiso esisezingeni lombhalo sivumela abafaka izichasiselo ukuthi babeke izimpendulo zombhalo eziningi ezikhiqizwe imodeli enkulu yolimi ngokususelwa kwinqubo eyenziwe ngokwezifiso enjengokuhambisana, ukucaca, noma ukunemba. Ama-Anvotators avezwa ngezimpendulo ezisheshayo neziningana ezikhiqizwayo, azibeka ngokwezinkombandlela eziqondile kwicala lakho lokusebenzisa. Idatha ebalwa ibanjwa ngefomethi ehlelekile, ichaza ama-indices ahlelwe kabusha kwinqubomgomo ngayinye, njengokuthi “ukucacisa” noma “ukucacisa.” Lolu lwazi lubalulekile kumamodeli we-tuning tuning usebenzisa i-RLHF, ukuhambisana nemiphumela yemodeli eduze kakhulu nokuthandwa ngabantu. Ngaphezu kwalokho, le template isebenza kakhulu ekuhlolweni kwekhwalithi ye-LLM okuphumayo ngokukuvumela ukuthi ubone ukuthi izimpendulo zingakanani zihambisana kanjani nenqubo ehlosiwe.

Ukusetha ku-sagemaker ai console

Okusha I-ai ekhiqizayo Isigaba sengezwe ngaphansi kohlobo lomsebenzi eSagemaker AI Console, okuvumela ukuthi ukhethe lezi zifanekiso. Ukumisa umsebenzi welebula usebenzisa i-AWS Management Console, uqedele lezi zinyathelo ezilandelayo:

1. E-sagemaker ai console, ngaphansi Iqiniso Lomhlabathi Kwiphaneli yokuhambisa, khetha Ukulebula umsebenzi.
2. Qoka Dala umsebenzi welebula.
3. Cacisa indawo yakho ebonakalayo yendawo nendlela yokukhipha. Ukumisa ifayela lokufaka umbhalo, sebenzisa Ukusetha kwedatha yesandla ngaphansi kwa- Dala umsebenzi welebula Futhi faka ifayela le-json nge-Prompt egcinwe ngaphansi kwensimu yomthombo, ngenkathi uhlu lwezimpendulo zemodeli lubekwa ngaphansi kwensimu yezimpendulo. Isimo sombhalo asisekeli Ukusethwa kwedatha okuzenzakalelayo.

Nasi isibonelo sefayela lethu lokufaka lokufaka:

Faka leli fayela elibonisa ukufakwa endaweni yakho ye-S3 bese unikeza indlela ye-S3 kuleli fayela ngaphansi Indawo yedatha yokufaka:

1. Qoka I-ai ekhiqizayo njengohlobo lomsebenzi bese ukhetha isikhundla se-UI.
   
   
2. Qoka Olandelayo.
3. Faka imiyalo yakho yokubhala. Faka ubukhulu ofuna ukufaka kulo Ubukhulu obuhle kakhulu ingxenye. Isibonelo, esithombeni esingenhla, ubukhulu Ubuzenjezo na- Ukuchazakepha ungangeza, ususe, noma wenze ngokwezifiso lezi ezisuselwa kwizidingo zakho ezithile ngokuchofoza inkinobho ethi “+” ukuze ungeze ubukhulu noma isithonjana sedoti ukuze ubasuse. Ngokwengeziwe, unenketho yokuthi Vumela amazinga okubopha Ngokukhetha ibhokisi lokuhlola. Le nketho inika amandla ama-Annotators ukuthi abeke izimpendulo ezimbili noma ngaphezulu ngokulinganayo uma ekholelwa izimpendulo zekhwalithi efanayo ngobukhulu obuthile.
   
4. Qoka Ukukhombisa umdlalo ngaphambi kokuthi ubonwe ngabantu bonke ukubonisa ithempulethi ye-UI ukuze kubuyekezwe.
   
5. Qoka Dala ukudala umsebenzi welebula.

Lapho ababashoni beletha ukuhlolwa kwabo, izimpendulo zabo zigcinwa ngqo kubhakede lakho elicacisiwe le-S3. Ifayela elibonisa ukuphuma lifaka amasimu wedatha yasekuqaleni kanye nempendulo ye-Worker-Ref elikhomba kufayela lokuphendula isisebenzi ku-S3. Leli fayela lokuphendula labasebenzi liqukethe izimpendulo ezisezingeni eliphakeme ubukhulu obuthile, ezingasetshenziswa ukwenza kahle noma ukuhlola imiphumela yemodeli yakho. Uma ngabetoli abaningi basebenze entweni efanayo yedatha, izichasiselo zabo ngazinye zifakwa ngaphakathi kwaleli fayela ngaphansi kokhiye wezimpendulo, okulula kwezimpendulo. Impendulo ngayinye ifaka okokufaka kwesichasiselo kanye ne-metadata efana nesikhathi sokwamukela, isikhathi sokungifaka, kanye ne-ID yesisebenzi. Nasi isibonelo sefayela lokukhipha i-JSON eliqukethe izichasiselo:

Umbuzo Nempendulo

Umbuzo kanye nethempulethi yezimpendulo ikuvumela ukuthi udale ama-datassets okuqondisa okuhle (SFT) ngokukhiqiza amabhangqa emibuzo nezimpendulo ezifundweni zombhalo. Ama-Annotators afunda umbhalo onikeziwe futhi adale imibuzo efanele nezimpendulo ezihambisanayo. Le nqubo isebenza njengomthombo we Imininingwane yokubonisaUkuqondisa imodeli yokuthi ungayiphatha kanjani imisebenzi efanayo. Isifanekiso sisekela ukufaka okuguquguqukayo, ukuvumela ama-anrotators abhekise wonke amavesi noma izingxenye ezithile zombhalo we-Q & A. Isici esinamakhodi anemibala sixhumanisa imibuzo ezigabeni ezifanele, ukusiza ukuhambisa inqubo yokufaka isichasiselo. Ngokusebenzisa lezi zimbili ze-Q & A, uthuthukisa amandla wemodeli yokulandela imiyalo futhi uphendule ngokunembile kokufakwa kwangempela komhlaba.

Ukusetha ku-sagemaker ai console

Inqubo yokusetha umsebenzi welebula ngombuzo kanye nethempulethi yezimpendulo kulandela izinyathelo ezifanayo njengethempulethi yokwethenjwa. Kodwa-ke, kunomehluko ngendlela olungiselela ngayo ifayela lokufaka bese ukhetha ithempulethi ye-UI efanelekile ukuze ivumelane nomsebenzi we-Q & a.

1. E-sagemaker ai console, ngaphansi Iqiniso Lomhlabathi Kwiphaneli yokuhambisa, khetha Ukulebula umsebenzi.
2. Qoka Dala umsebenzi welebula.
3. Cacisa indawo yakho ebonakalayo yendawo nendlela yokukhipha. Ukumisa umbuzo kanye nefayela lokufaka impendulo, sebenzisa Ukusetha kwedatha yesandla Futhi layisha ifayela le-JSON lapho insimu yomthombo iqukethe khona isigaba sombhalo. Ama-Ancotators azosebenzisa lo mbhalo ukukhiqiza imibuzo nezimpendulo. Qaphela ukuthi ungalayisha umbhalo kusuka kufayela le-.txt noma .csv bese usebenzisa iqiniso lomhlaba Ukusetha kwedatha okuzenzakalelayo ukuyiguqula ibe yifomethi ye-JSON edingekayo.

Nasi isibonelo sefayela eliveziwe lokufaka:

Faka leli fayela elibonisa ukufakwa endaweni yakho ye-S3 bese unikeza indlela ye-S3 kuleli fayela ngaphansi Indawo yedatha yokufaka

1. Qoka I-ai ekhiqizayo njengohlobo lomsebenzi bese ukhetha Umbuzo Nempendulo Ui
2. Qoka Olandelayo.
3. Faka imiyalo yakho yokubhala. Ungahlela ezinye izilungiselelo ukulawula umsebenzi. Ungacacisa inani eliphansi neliphezulu le-Q & A ngababili abasebenzi okufanele bakhiqize kusuka endimeni yombhalo enikeziwe. Ngokwengeziwe, ungachaza inani eliphansi neliphezulu lamagama wombuzo nezinkundla zokuphendula, ukuze izimpendulo zivumelane nezidingo zakho. Ungangeza futhi omaki wemibuzo ongayikhetha ukuze uhlukanise umbuzo bese uphendula ngababili. Isibonelo, ungafaka amathegi anjengokuthi “Yini,” “Kanjani,” noma “Kungani” ukuqondisa abahloli emsebenzini wabo. Uma amathegi achazwe ngaphambilini anele, unenketho yokuvumela abasebenzi ukuthi bafake amathegi abo ngokwezifiso ngokwawo ngokunika amandla i- Vumela abasebenzi ukuthi bacacisele amathegi wangokwezifiso . Lokhu kuvumelana nezimo kusiza izichasiselo ezihlangabezana nezidingo ezithile zecala lakho lokusebenzisa.
   
4. Lapho lezi zilungiselelo zilungiselelwe, ungakhetha Ukukhombisa umdlalo ngaphambi kokuthi ubonwe ngabantu bonke I-UI ukuqinisekisa ukuthi ihlangabezana nezidingo zakho ngaphambi kokuqhubeka.
   
5. Qoka Dala ukudala umsebenzi welebula.

Lapho ama-Adotitators eletha umsebenzi wawo, izimpendulo zawo zigcinwa ngqo ibhakede lakho elibekiwe le-S3. Le khasi okubonakalayo okubonakalayo Ifayela liqukethe izinkambu zedatha zoqobo kanye ne Impendulo-Ref-Ref ukuthi kukhomba kufayela lokuphendula labasebenzi ku-S3. Leli fayela lokuphendula labasebenzi lifaka phakathi izichasiselo ezinemininingwane ezinikezwe abasebenzi, njengezimpendulo ezisezingeni noma ngazimbili zemibuzo nezimpendulo ezikhiqizwe ngomsebenzi ngamunye.

Nasi isibonelo salokho okukhona okukhona kungabukeka ngakulokhu:

I-CrealeLabeleLob API

Ngaphezu kokudala le misebenzi yelebula nge-Amazon SageMaker Ai Console, amakhasimende angasebenzisa futhi Dala ilebula umsebenzi we-API Ukusetha isikhundla sombhalo kanye nombuzo kanye nokuphendula imisebenzi ngokuhlelekile. Le ndlela ihlinzeka ngezifiso ezengeziwe ze-automation nokuhlanganiswa kokuhamba komsebenzi okukhona. Usebenzisa i-API, ungachaza ukucushwa komsebenzi, ukuvezwa kokufaka, kanye nezifanekiso zomsebenzi wesisebenzi, nokubheka inqubekela phambili yomsebenzi ngokuqondile kusuka kuhlelo lwakho lokusebenza noma uhlelo.

Ngomhlahlandlela wesinyathelo ngesinyathelo sokuthi ungakusebenzisa kanjani lokhu, ungabheka ezincwadini ezilandelayo, ezihamba kuyo yonke inqubo yokuhlelwa kokufunda komuntu (i-hitl) yokuqinisa ukuqiniswa kwempendulo yomuntu (i-RLHF) esebenzisa izifundo zombhalo kanye nemibuzo nezibonisi. Lezi zincwajana zizokuqondisa ngokusetha iqiniso elidingekayo leqiniso elidingekayo, ukulanda amafayela wesampula json ngezimpendulo nezimpendulo, ukuwaguqula ukuze zibonise ukufakwa kweqiniso kweqiniso, futhi kubhekwe imisebenzi yokubhala. Zibuye zimboze kabusha imiphumela yeposi ukudala idatha ehlanganisiwe ngezimpendulo ezisezingeni.

Ukugcina

Ngokwethulwa kwesikhundla sombhalo kanye nombuzo kanye nezibonisi zokuphendula, ama-Amazon Sagemaker Ai Empolers amakhasimende ukukhiqiza ama-datasets asezingeni eliphakeme wokuqeqesha amamodeli amakhulu olimi ngempumelelo. Lawa makhono owakhelwe ngaphakathi enza lula inqubo yamamodeli alungiselelwe kahle emisebenzini ethile futhi avumelane nemiphumela yawo ngokuncamelayo komuntu, kungaba ngokufunda okuhle okufundwayo noma ngokuqiniswa kokufunda okuvela kwimpendulo yabantu. Ngokufaka izifanekiso, ungahlola kangcono futhi uhlanze amamodeli akho ukuze uhlangabezane nezidingo zesicelo sakho esithile, usize ukufeza imiphumela enembile, ethembekile, futhi yomsebenzisi. Noma ngabe ukudala ama-datassets okuqeqeshwa noma okuhlola imiphumela yemodeli yakho, i-sagemaker AI ihlinzeka ngamathuluzi owadingayo ukuze uphumelele ekwakheni ama-state-of-art art Solutions.Ukuqala ukudala imininingwane elungiselelwe kahle ngethempulethi emisha:

Mayelana nababhali

ISundar Raghavan Ingabe ukwakhiwa kwezixazululo ze-AI akhiqizayo ku-AWS, ukusiza amakhasimende ukuthi asebenzise ama-Amazon Bedrock kanye nezinsizakalo ze-AWS ezizayo zokuqamba, ukwakha kanye nokuthumela ama-AI Agents kanye nezicelo ze-AI ezi-Scablerative ai. Esikhathini sakhe samahhala, uSundar uthanda ukuhlola izindawo ezintsha, edlulisa amasampula ama-eateries endawo futhi amukele ngaphandle.

Ama-Jesse Manti Imenenja yomkhiqizo ephezulu e-Amazon Bedrock, insiza kanjiniyela we-AWS DEENTION AI. Usebenza ekuxhumaneni kwe-AI nokuxhumana komuntu nenhloso yokwakha nokwenza ngcono imikhiqizo nezinsizakalo ze-AI akhiqizayo ukufeza izidingo zethu. Phambilini, uJesse wabamba ubunjiniyela bokuhola kweqembu leqembu e-apula naku-lumeds, futhi wayengusosayensi omkhulu ekuqalisweni kweSilicon Valley. Une-MS ne-PH.D. Ukusuka e-University of Florida, kanye ne-MBA evela eNyuvesi yaseCalifornia, eBerkeley, isikole sebhizinisi.

Niharika jayanti Ingabe unjiniyela ongaphambili e-Amazon, lapho aklama khona futhi athuthukise ukuhlangana komsebenzisi ukujabulisa amakhasimende. Ufake isandla ekuvulekeni ngempumelelo kwamathuluzi we-LLM wokuhlola e-Amazon Bedrock nase-Amazon SageMaker Unified Studio. Ngaphandle komsebenzi, uNiharika uyakujabulela ukubhukuda, ukushaya indawo yokuzivocavoca kanye ne-crocheting.

Muyun yan ngunjiniyela wesoftware ophezulu e-Amazon Web Services (AWS) iSageMaker AI Team. Ngaphezulu kweminyaka engu-6 ku-AWS, ubheka kakhulu ekuthuthukiseni amapulatifomu asuselwa kumshini wokulebula. Umsebenzi wakhe ugxile ekwakheni nasekuphaseni izinhlelo zokusebenza zesoftware ezintsha zokulebula izixazululo, ukunika amandla amakhasimende ukufinyelela amakhono okulebula kokusika. UMuyin uphethe ama-MS ku-Computer Engineering evela eBoston University.

Kavya Kotra Ingabe unjiniyela wesoftware eqenjini le-Amazon SageMaker Ground Thround, usiza ukukhelwa kwezinhlelo zesoftware. UKavya wadlala indima ebalulekile ekuthuthukisweni nasekwethulweni kwamathuluzi we-AI akhiqizayo ku-sagemaker. Phambilini, uKavya wabamba izindima zobunjiniyela ngaphakathi kwe-AWS EC2 Networking, ne-Amazon ezwakalayo. Esikhathini sakhe samahhala, uyakujabulela ukupenda, futhi ehlola isimo semvelo sikaSeattle.

U-Alan Ismaiel unjiniyela wesoftware at aw esuselwa eNew York City. Ugxile ekwakheni nasekulondolozeni imikhiqizo ye-AI / ML ehlelekile, njengeqiniso le-Amazon SageMaker ne-Amazon Bedrock. Ngaphandle komsebenzi, u-Alan ufunda indlela yokudlala i-pickleball, ngemiphumela exubile.