Manyan nau'ikan harshe (LLMs) gabaɗaya ana horar da su akan manyan bayanan da ake samu a bainar jama'a waɗanda ke agnostic na yanki. Misali, Meta's Llama model an horar da a kan datasets kamar Rarraba Jama'a, C4, Wikipedia, da ArXiv. Waɗannan rukunonin bayanai sun ƙunshi batutuwa da yankuna da dama. Kodayake samfuran da aka samo suna ba da kyakkyawan sakamako mai ban mamaki ga ayyuka na gabaɗaya, kamar tsarar rubutu da tantance mahalli, akwai shaidar cewa ƙirar da aka horar da takamaiman bayanai na yanki na iya ƙara haɓaka aikin LLM. Misali, bayanan horon da aka yi amfani da su BloombergGPT takaddun takamaiman yanki ne na 51%, gami da labaran kuɗi, filaye, da sauran kayan kuɗi. Sakamakon LLM ya fi LLMs horo akan ƙayyadaddun bayanai marasa iyaka lokacin da aka gwada su akan takamaiman ayyuka na kuɗi. Marubuta na BloombergGPT sun kammala da cewa samfurin su ya zarce duk sauran samfuran da aka gwada na hudu daga cikin ayyukan kudi biyar. Samfurin ya ba da mafi kyawun aiki lokacin da aka gwada ayyukan kuɗi na cikin gida na Bloomberg ta tazara mai faɗi - kusan maki 60 mafi kyau (cikin 100). Ko da yake kuna iya ƙarin koyo game da cikakken sakamakon kimantawa a cikin takarda, samfurin da aka kama daga BloombergGPT takarda na iya ba ku hangen nesa game da fa'idar horar da LLMs ta amfani da takamaiman bayanan yanki na kuɗi. Kamar yadda aka nuna a cikin misalin, ƙirar BloombergGPT ta ba da amsoshi daidai yayin da sauran samfuran da ba takamaiman yanki ba sun yi gwagwarmaya:

Input: CEO of Assicurazioni Generali?
BloombergGPT: Philippe Donnet
GPT-NeoX: Antonio De Lorenzo, Simone Gambarini, Enrico Zanetti
FLAN-T5-XXL: John M Forsyth, Christopher K Peters, {empty string}

Input: CEO of Silicon Valley Bank?
BloombergGPT: Greg Becker
GPT-NeoX: John Hodor, Mike Fridgen, Peter R Briger
FLAN-T5-XXL: Christian Klein, John M Forsyth, John M Forsyth

Input: CEO of IHS Markit Ltd?
BloombergGPT: Lance Uggla
GPT-NeoX: John F Lechleiter, David Steward, Niall O Broin
FLAN-T5-XXL: William M Hawkins, John M Forsyth, John M Forsyth
Input: CEO of Walgreens Boots Alliance?

BloombergGPT: Rosalind Brewer
GPT-NeoX: Mark V Dei, Stefano Pessina, Stefano Pessina
FLAN-T5-XXL: Christian Klein, John M Forsyth, John M Forsyth
Input: CEO of Citigroup Inc?

BloombergGPT: Jane Fraser
GPT-NeoX: Michael L Corbat, Michael L Corbat, Michael L Corbat
FLAN-T5-XXL: Christian Sewing, John M Forsyth, John M Forsyth

Wannan sakon yana ba da jagora don horar da LLMs musamman don yankin kuɗi. Muna rufe mahimman fannoni masu zuwa:

• Tarin bayanai da shiri - Jagora kan samowa da daidaita bayanan kuɗi masu dacewa don ingantaccen horon ƙira
• Ci gaba da horarwa tare da daidaitawa mai kyau - Lokacin amfani da kowace dabara don haɓaka aikin LLM ɗin ku
• Ingantaccen ci gaba da horo kafin horo - Dabaru don daidaita tsarin ci gaba da horarwa, adana lokaci da albarkatu

Wannan sakon ya haɗu da ƙwarewar ƙungiyar binciken kimiyya da aka yi amfani da su a cikin Fasahar Kuɗi na Amazon da kuma Ƙungiyar Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Duniya na AWS. Wasu daga cikin abubuwan sun dogara ne akan takarda Ingantacciyar Ci gaba da Horowa don Gina Takamaiman Manyan Samfuran Harshe.

Tattara da shirya bayanan kuɗi

Domain ci gaba da buƙatun horo kafin horo babban ma'auni, inganci, ƙayyadaddun bayanai na yanki. Wadannan su ne manyan matakai don sarrafa saitin bayanan yanki:

• Gano tushen bayanai - Tushen bayanai masu yuwuwa don yankin yanki sun haɗa da buɗaɗɗen gidan yanar gizo, Wikipedia, littattafai, kafofin watsa labarun, da takaddun ciki.
• Tace bayanan yanki - Saboda maƙasudin maƙasudi shi ne ƙaddamar da yanki na yanki, kuna iya buƙatar amfani da ƙarin matakai don tace samfuran da ba su da alaƙa da yankin da aka yi niyya. Wannan yana rage ƙungiyar mara amfani don ci gaba da horarwa kafin horo kuma yana rage farashin horo.
• Gabatarwa - Kuna iya yin la'akari da jerin matakan da aka riga aka tsara don inganta ingancin bayanai da ingancin horo. Misali, wasu hanyoyin bayanai na iya ƙunsar daidaitattun adadin alamun hayaniya; Ana ɗaukar raguwa a matsayin mataki mai amfani don inganta ingancin bayanai da rage farashin horo.

Don haɓaka LLMs na kuɗi, zaku iya amfani da mahimman bayanan bayanai guda biyu: Labarai CommonCrawl da SEC filings. Shigar da SEC bayanin kuɗi ne ko wasu takaddun ƙa'ida da aka ƙaddamar zuwa Hukumar Tsaro da Musanya ta Amurka (SEC). Ana buƙatar kamfanoni da aka jera a bainar jama'a su rubuta takardu daban-daban akai-akai. Wannan yana haifar da adadi mai yawa na takardu a cikin shekaru. News CommonCrawl shine tsarin bayanan da CommonCrawl ya fitar a cikin 2016. Ya ƙunshi labaran labarai daga shafukan labarai a duk faɗin duniya.

News CommonCrawl yana samuwa akan Sabis na Sauƙi na Amazon (Amazon S3) a cikin commoncrawl guga a crawl-data/CC-NEWS/. Kuna iya samun lissafin fayiloli ta amfani da Hanyar Layin Umarnin AWS (AWS CLI) da umarni mai zuwa:

aws s3 ls --recursive s3://commoncrawl/crawl-data/CC-NEWS/

In Ingantacciyar Ci gaba da Horowa don Gina Takamaiman Manyan Samfuran Harshe, mawallafa suna amfani da URL da tsarin tushen kalmomi don tace labaran labarai na kudi daga labarai na yau da kullum. Musamman ma, marubutan suna kula da jerin mahimman kantunan labarai na kuɗi da saitin kalmomi masu alaƙa da labaran kuɗi. Mun gano labarin a matsayin labaran kuɗi idan ko dai ya fito ne daga kantunan labarai na kuɗi ko kuma wasu kalmomin da suka bayyana a cikin URL. Wannan hanya mai sauƙi amma mai tasiri tana ba ku damar gano labaran kuɗi daga kantunan labarai na kuɗi ba kawai ba har ma da sassan kuɗi na kantunan labarai na yau da kullun.

Ana samun fayilolin SEC akan layi ta hanyar SEC's EDGAR (Taron Bayanan Lantarki, Analysis, da Maidowa), wanda ke ba da damar buɗe bayanai. Kuna iya goge fayilolin daga EDGAR kai tsaye, ko amfani da APIs a ciki SageMaker na Amazon tare da ƴan layukan lamba, na kowane lokaci kuma don ɗimbin tickers (watau, mai gano SEC da aka sanya). Don ƙarin koyo, koma zuwa Maido da Shigar SEC.

Tebur mai zuwa yana taƙaita mahimman bayanai na tushen bayanai guda biyu.

Marubutan sun bi ta wasu ƴan matakan da za a iya aiwatarwa kafin a ciyar da bayanai cikin algorithm horo. Na farko, mun lura cewa fayilolin SEC sun ƙunshi rubutu mai hayaniya saboda cire tebur da ƙididdiga, don haka marubutan sun cire gajerun jimloli waɗanda ake zaton tebur ne ko alamun adadi. Abu na biyu, muna amfani da algorithm ɗin hashing mai hankali na yanki don ƙaddamar da sabbin labarai da fage. Don fayilolin SEC, muna ƙaddamarwa a matakin sashe maimakon matakin daftarin aiki. A ƙarshe, muna haɗa takardu zuwa dogon kirtani, mu sanya alama, sannan mu datse alamar zuwa guntun tsayin shigarwar max wanda ke goyan bayan ƙirar don horarwa. Wannan yana inganta abubuwan da ake samu na ci gaba da horarwa kafin horo kuma yana rage farashin horo.

Ci gaba da horarwa tare da daidaitawa mai kyau

Yawancin LLMs da ake samu gabaɗaya manufa ne kuma basu da takamaiman iyakoki na yanki. LLMs na yanki sun nuna babban aiki a fannin likitanci, kuɗi, ko yanki na kimiyya. Don LLM don samun takamaiman ilimin yanki, akwai hanyoyi guda huɗu: horo daga karce, ci gaba da horarwa kafin horo, ingantaccen koyarwa akan ayyukan yanki, da Maido da Ƙarfafa Ƙarfafawa (RAG).

A cikin ƙirar al'ada, ana amfani da daidaitawa mai kyau don ƙirƙirar ƙayyadaddun samfura don yanki. Wannan yana nufin kiyaye samfura da yawa don ayyuka da yawa kamar hakar mahalli, rarraba niyya, nazarin ji, ko amsa tambaya. Tare da zuwan LLMs, buƙatar kula da ƙira daban-daban ya zama marar amfani ta amfani da dabaru kamar koyo na yanayi ko faɗakarwa. Wannan yana adana ƙoƙarce-ƙoƙarcen da ake buƙata don kula da tarin samfura don alaƙa amma daban-daban ayyuka.

Da fahimta, zaku iya horar da LLMs daga karce tare da takamaiman bayanai na yanki. Kodayake yawancin ayyukan ƙirƙirar LLMs na yanki sun mai da hankali kan horarwa daga karce, yana da tsada sosai. Misali, farashin samfurin GPT-4 fiye da $ 100 don horarwa. An horar da waɗannan samfuran akan haɗakar bayanan yanki da buɗaɗɗen bayanan yanki. Ci gaba da horarwa na iya taimakawa samfura su sami takamaiman ilimi na yanki ba tare da jawo farashin horon farko daga karce ba saboda kun riga kun horar da wani yanki na buɗe LLM akan bayanan yanki kawai.

Tare da ingantaccen koyarwa akan ɗawainiya, ba za ku iya sa ƙirar ta sami ilimin yanki ba saboda LLM kawai ke samun bayanan yanki da ke ƙunshe a cikin tsararrun bayanai masu kyau. Sai dai idan an yi amfani da babban kundin bayanai don daidaitawa mai kyau, bai isa a sami ilimin yanki ba. Samar da manyan bayanai na koyarwa na yawanci ƙalubale ne kuma shine dalilin amfani da LLMs a farkon wuri. Hakanan, daidaitawar koyarwa akan ɗawainiya ɗaya na iya rinjayar aiki akan wasu ayyuka (kamar yadda aka gani a cikin wannan takarda). Duk da haka, gyaran gyare-gyaren koyarwa ya fi amfani da tsada fiye da ɗayan hanyoyin horarwa.

Adadin da ke gaba yana kwatanta ƙayyadaddun aiki na gargajiya. vs in-context koyo yanayin tare da LLMs.

RAG ita ce hanya mafi inganci ta jagorantar LLM don samar da martani da aka kafa a cikin yanki. Ko da yake yana iya jagorantar samfurin don samar da martani ta hanyar samar da bayanai daga yankin a matsayin bayanin taimako, baya samun takamaiman harshe na yanki saboda LLM har yanzu yana dogara ga salon yaren da ba na yanki ba don samar da martani.

Horowa na ci gaba da kasancewa tsaka-tsaki tsakanin horarwa da koyarwa mai kyau game da farashi yayin da yake kasancewa mai ƙarfi madadin samun takamaiman ilimi da salo na yanki. Yana iya samar da samfuri na gaba ɗaya wanda za'a iya aiwatar da ƙarin daidaitawa akan taƙaitaccen bayanin koyarwa. Ci gaba da horarwa na iya zama dabara mai inganci don yankuna na musamman inda saitin ayyuka na ƙasa ke da girma ko ba a sani ba kuma bayanan kunna wa'azin da aka yi wa lakabin yana iyakance. A cikin wasu al'amuran, daidaitawar koyarwa ko RAG na iya zama mafi dacewa.

Don ƙarin koyo game da daidaitawa, RAG, da horon ƙira, koma zuwa Daidaita samfurin tushe, Maidowa Ƙarfafa Ƙarfafa (RAG), Da kuma Horar da Model tare da Amazon SageMaker, bi da bi. Don wannan post ɗin, muna mai da hankali kan ingantaccen ci gaba da horarwa.

Hanyar ingantaccen ci gaba da horarwa

Ci gaba da horon horo ya ƙunshi hanyoyi masu zuwa:

• Domain-Adaptive Ci gaba da Horowa (DACP) – A cikin takarda Ingantacciyar Ci gaba da Horowa don Gina Takamaiman Manyan Samfuran Harshe, Mawallafa suna ci gaba da horar da tsarin ƙirar harshen Pythia akan tsarin kuɗi don daidaita shi zuwa yankin kuɗi. Manufar ita ce ƙirƙirar LLMs na kuɗi ta hanyar ciyar da bayanai daga duk yankin kuɗi zuwa ƙirar buɗe ido. Saboda ƙungiyar horarwa ta ƙunshi duk bayanan da aka keɓe a cikin yanki, samfurin da zai haifar ya kamata ya sami takamaiman ilimin kuɗi, ta haka ya zama ƙirar ƙira don ayyuka daban-daban na kuɗi. Wannan yana haifar da ƙirar FinPythia.
• Task-Daɗawa Ci gaba da Horowa (TACP) - Marubuta sun riga sun horar da ƙirar gaba akan bayanan ɗawainiya masu lakabi da marasa laka don daidaita su don takamaiman ayyuka. A wasu yanayi, masu haɓakawa na iya fifita ƙira waɗanda ke isar da ingantacciyar aiki akan gungun ayyuka na cikin gida maimakon ƙirar yanki-janar. An ƙirƙira TACP azaman ci gaba da horon horo da nufin haɓaka aiki akan ayyukan da aka yi niyya, ba tare da buƙatun bayanan da aka yiwa lakabi ba. Musamman, marubutan suna ci gaba da horar da samfuran buɗaɗɗen tushe akan alamun aiki (ba tare da tambari ba). Iyakan farko na TACP ya ta'allaka ne a gina takamaiman LLMs maimakon kafuwar LLMs, saboda kawai amfani da bayanan aikin da ba a lakafta ba don horo. Kodayake DACP yana amfani da corpus mafi girma, yana da tsada sosai. Don daidaita waɗannan iyakoki, marubutan sun ba da shawarar hanyoyi guda biyu waɗanda ke da nufin gina ƙayyadaddun tushe na LLMs yayin da suke kiyaye ingantaccen aiki akan ayyuka masu niyya:

• Ingantacciyar Aiki-Kamar DACP (ETS-DACP) - Mawallafa suna ba da shawarar zaɓar wani yanki na ƙungiyar kuɗi wanda yayi kama da bayanan ɗawainiya ta amfani da haɗa kamanni. Ana amfani da wannan juzu'in don ci gaba da horarwa don ci gaba da inganta shi. Musamman ma, marubutan suna ci gaba da horar da buɗaɗɗen tushen LLM akan ƙaramin rukunin da aka samo daga ƙungiyar kuɗi wanda ke kusa da ayyukan da aka yi niyya a cikin rarrabawa. Wannan zai iya taimakawa inganta aikin aiki saboda mun ɗauki samfurin zuwa rarraba alamun aiki duk da bayanan da ba a buƙata ba.
• Ingantacciyar Aiki-Agnostic DACP (ETA-DACP) - Mawallafa sun ba da shawarar yin amfani da ma'auni kamar ruɗani da nau'in alamar entropy waɗanda ba sa buƙatar bayanan ɗawainiya don zaɓar samfura daga ƙungiyar kuɗi don ingantaccen ci gaba da horarwa. An ƙera wannan hanyar don magance al'amuran da ba a samu bayanan ɗawainiya ba ko kuma an fi son ƙirar yanki mai fa'ida don babban yanki. Marubutan sun ɗauki nau'i biyu don zaɓar samfuran bayanai waɗanda ke da mahimmanci don samun bayanan yanki daga ɓangaren bayanan yanki na farko na horo: sabon abu da bambancin. Sabon abu, wanda aka auna ta hanyar ruɗani da aka rubuta ta tsarin manufa, yana nufin bayanin da LLM bai gani a baya ba. Bayanai tare da babban sabon abu suna nuna sabon ilimin ga LLM, kuma ana kallon irin waɗannan bayanan a matsayin mafi wahalar koyo. Wannan yana sabunta LLMs na gabaɗaya tare da zurfin ilimin yanki yayin ci gaba da horarwa. Bambance-bambance, a gefe guda, yana ɗaukar nau'ikan rarraba nau'ikan token a cikin yanki na yanki, wanda aka rubuta a matsayin sifa mai fa'ida a cikin binciken koyon karatu akan ƙirar harshe.

Wannan adadi yana kwatanta misalin ETS-DACP (hagu) vs. ETA-DACP (dama).

Mun yi amfani da tsare-tsaren samfuri guda biyu don zaɓar wuraren bayanai da gaske daga ƙungiyar kuɗi da aka ware: samfur mai wuya da samfur mai laushi. Ana yin na farko ne ta farko ta ƙirƙira ƙungiyar kuɗi ta hanyar ma'auni masu dacewa sannan zaɓi samfuran saman-k, inda aka ƙaddara k bisa ga kasafin horo. Don na ƙarshe, marubutan suna ba da ma'aunin samfuri ga kowane maki bayanai bisa ga ma'auni, sannan ba da izini samfurin abubuwan bayanan k don saduwa da kasafin horo.

Sakamako da bincike

Marubutan suna kimanta sakamakon LLMs na kuɗi akan ɗimbin ayyuka na kuɗi don bincika ingancin ci gaba da horarwa:

• Bankin Jumlar kudi - Aikin rarraba ra'ayi akan labaran kuɗi.
• FIQA SA - Ayyukan rarraba ra'ayi na tushen al'amari dangane da labaran kuɗi da kanun labarai.
• kanun labarai - Ayyukan rarrabuwa na binary akan ko kanun labarai akan mahaɗin kuɗi ya ƙunshi wasu bayanai.
• Ner - Aiki mai suna na kuɗi mai suna wanda ya dogara da sashin kimanta haɗarin bashi na rahoton SEC. Kalmomi a cikin wannan ɗawainiyar an tsara su da PER, LOC, ORG, da MISC.

Saboda LLMs na kuɗi an daidaita koyarwar da kyau, marubutan suna kimanta samfura a cikin saitin harbi 5 don kowane ɗawainiya saboda ƙarfi. A matsakaita, FinPythia 6.9B ya zarce Pythia 6.9B da 10% a cikin ayyuka huɗu, wanda ke nuna ingancin takamaiman horo na ci gaba na yanki. Don samfurin 1B, haɓakawa ba shi da zurfi, amma aikin har yanzu yana inganta 2% a matsakaici.

Hoto mai zuwa yana kwatanta bambancin aiki kafin da bayan DACP akan samfuran biyu.

Hoto na gaba yana nuna misalai biyu masu inganci waɗanda Pythia 6.9B da FinPythia 6.9B suka samar. Don tambayoyin da suka shafi kuɗi guda biyu game da manajan mai saka jari da kuma lokacin kuɗi, Pythia 6.9B ba ta fahimtar kalmar ko gane sunan, yayin da FinPythia 6.9B ke samar da cikakkun amsoshi daidai. Misalai masu inganci sun nuna cewa ci gaba da horarwa kafin horo yana bawa LLMs damar samun ilimin yanki yayin aiwatarwa.

Teburin da ke gaba yana kwatanta ingantattun hanyoyin ci gaba na gaba da horo. ETA-DACP-ppl shine ETA-DACP dangane da ruɗani (sabon abu), kuma ETA-DACP-ent ya dogara ne akan entropy (bambance-bambancen). ETS-DACP-com yayi kama da DACP tare da zaɓin bayanai ta hanyar matsakaicin duk ma'auni uku. Abubuwan da ke biyowa kaɗan ne daga sakamakon:

• Hanyoyin zaɓin bayanai suna da inganci - Sun zarce daidaitattun horo na gaba da gaba tare da kawai 10% na bayanan horo. Ingantacciyar horo na ci gaba da ci gaba da ya haɗa da Task-Similar DACP (ETS-DACP), Task-Agnostic DACP dangane da entropy (ESA-DACP-ent) da Aiki-Kamar DACP dangane da duk ma'auni uku (ETS-DACP-com) ya fi daidaitattun DACP a matsakaita duk da cewa an horar da su a kan kawai 10% na hada-hadar kudi.
• Zaɓin bayanan aiki-sane da aiki yana aiki mafi kyau daidai da ƙananan binciken ƙirar harshe - ETS-DACP yana rikodin mafi kyawun matsakaicin aiki tsakanin duk hanyoyin kuma, dangane da duk ma'auni guda uku, yana yin rikodin aikin aiki na biyu mafi kyau. Wannan yana nuna cewa yin amfani da bayanan ɗawainiyar da ba a lakafta shi ba har yanzu hanya ce mai inganci don haɓaka aikin aiki a cikin yanayin LLMs.
• Zaɓin bayanan aiki-agnostic yana kusa da na biyu - ESA-DACP-ent yana biye da aikin tsarin zaɓin bayanai na ɗawainiya, yana nuna cewa har yanzu za mu iya haɓaka aikin aiki ta hanyar zabar samfurori masu inganci waɗanda ba a haɗa su da takamaiman ayyuka ba. Wannan yana buɗe hanya don gina LLMs na kuɗi don ɗaukacin yanki yayin samun ingantaccen aikin aiki.

Tambaya ɗaya mai mahimmanci game da ci gaba da horon horo shine ko yana yin mummunan tasiri akan ayyukan da ba na yanki ba. Har ila yau, marubutan sun ƙididdige samfurin da aka riga aka horar da su akan ayyuka guda huɗu da ake amfani da su sosai: ARC, MMLU, TruthQA, da HellaSwag, waɗanda ke auna ƙarfin amsa tambaya, tunani, da kuma kammalawa. Marubutan sun gano cewa ci gaba da horarwa ba ya yin illa ga ayyukan da ba na yanki ba. Don ƙarin bayani, koma zuwa Ingantacciyar Ci gaba da Horowa don Gina Takamaiman Manyan Samfuran Harshe.

Kammalawa

Wannan sakon ya ba da haske game da tattara bayanai da ci gaba da dabarun horarwa don horar da LLMs don yankin kuɗi. Kuna iya fara horar da naku LLMs don ayyukan kuɗi ta amfani da su Koyarwar SageMaker ta Amazon or Amazon Bedrock a yau.

Game da Authors

Yong Xie ƙwararren masanin kimiyya ne a Amazon FinTech. Yana mai da hankali kan haɓaka manyan samfuran harshe da aikace-aikacen AI na Generative don kuɗi.

Karan Aggarwal Babban Masanin Kimiyya ne wanda aka Aiwatar da Amazon FinTech tare da mai da hankali kan Generative AI don amfani da kuɗi. Karan yana da gogewa mai yawa a cikin bincike-jerin lokaci da kuma NLP, tare da sha'awar koyo daga iyakanceccen bayanai masu lakabi

Ayzaz Ahmad Manajan Kimiyya ne da aka Aiwatar a Amazon inda yake jagorantar ƙungiyar masana kimiyya da ke gina aikace-aikace daban-daban na Koyon Injin da Generative AI a cikin Kuɗi. Abubuwan bincikensa suna cikin NLP, Generative AI, da Agents LLM. Ya samu digirin digirgir (PhD) a fannin injiniyan lantarki daga jami'ar Texas A&M.

Qingwei Li ƙwararren ƙwararren Koyon Inji a Sabis na Yanar Gizo na Amazon. Ya samu Ph.D. a Operations Research bayan da ya karya asusun tallafin bincike na mai ba shi shawara kuma ya kasa ba da kyautar Nobel da ya yi alkawari. A halin yanzu yana taimaka wa abokan ciniki a cikin sabis na kuɗi don gina hanyoyin koyon injin akan AWS.

Raghvender Arni yana jagorantar Ƙungiyar Haɓakawa Abokin Ciniki (CAT) a cikin AWS Masana'antu. CAT ƙungiyar haɗin gwiwar duniya ce ta abokin ciniki da ke fuskantar injiniyoyin girgije, injiniyoyin software, masana kimiyyar bayanai, da ƙwararrun AI / ML da masu zanen kaya waɗanda ke tafiyar da ƙididdigewa ta hanyar ƙirar ƙira, kuma suna fitar da ingantaccen aikin girgije ta hanyar ƙwarewar fasaha ta musamman.

• Ƙunshin Ƙarfafa SEO & Rarraba PR. Samun Girma a Yau.
• PlatoData.Network Tsaye Generative Ai. Karfafa Kanka. Shiga Nan.
• PlatoAiStream. Web3 Mai hankali. Ilmi Ya Faru. Shiga Nan.
• PlatoESG. Carbon, CleanTech, Kuzari, Muhalli, Solar, Gudanar da Sharar gida. Shiga Nan.
• PlatoHealth. Biotech da Clinical Trials Intelligence. Shiga Nan.
• Source: https://aws.amazon.com/blogs/machine-learning/efficient-continual-pre-training-llms-for-financial-domains/