From-scratch Foundation Model CLEVI AI(ivy)

“Ta yaya ƙaramar kamfani ta iya ƙirƙirar babban samfurin harshe na AI mai matuƙar girma, wanda yake a matakin mafi girma a duniya?”

Ta yaya ƙaramar farawa ta ƙirƙiri babban samfurin harshe mai matuƙar girma?

Wannan ita ce tambayar da aka fi yi tun bayan da CLEVI ta bayyana samfurin Foundation mai ma'aunin tiriliyan 1.4 na sigogi.

Sabanin kamfanoni da yawa waɗanda kawai suke yin fine-tuning na samfuran buɗaɗɗen tushe, CLEVI ta gudanar da dukkan matakai da kanta, daga tattara bayanai, tsara samfurin, babban horo mai rarraba aiki, har zuwa tantance inganci.

A ƙasa muna gabatar da sirrin nasararta da ƙarfin gasa dalla-dalla.

1. Ƙaramar kamfani za ta ƙirƙiri babban samfurin harshe mai matuƙar girma?

A watan Yuli na 2025, lokacin da CLEVI ta bayyana babban samfurin harshe mai matuƙar girma da ta ƙirƙira da kanta a kasuwa, abokan ciniki (masu amfani) da yawa sun shiga cikin mamaki da shakku. “Da gaske an ƙirƙire shi da kansu?” “Ba kawai sun ɗan sauya samfurin buɗaɗɗen tushe ba ne?”

A zahiri, kamfanoni da yawa a cikin gida da ƙasashen waje suna yin transfer learning (fine-tuning) kawai ga samfuran buɗaɗɗen tushe, sannan su tallata su a matsayin samfuran da suka ƙirƙira da kansu.

Amma CLEVI ta gabatar da **‘From-scratch Foundation Model’**.

Wato, ta tsara kuma ta aiwatar da dukkan matakai da kanta, daga tattara bayanai, tsara samfurin, babban horo mai rarraba aiki, har zuwa tantance inganci.

2. Dalilin da ya sa haɓaka babban samfurin harshe mai matuƙar girma yake da wahala

Don horar da From-Scratch Foundation Model, dole ne a horar da dukkan abubuwan da ke ƙasa.

Ana buƙatar manyan tarin bayanai masu inganci

• Bambance-bambance: Tabbatar da bambance-bambancen harsuna, fannoni da tsare-tsare (rubutu, hotuna da sauransu)
• Tsabtacewa: Cire hayaniya, sarrafa inganci, da tace bayanan da aka maimaita ko masu cutarwa
• Lasisi: Bayyana haƙƙin mallaka da izinin amfani da bayanai a sarari

Manyan ababen more rayuwa na kwamfuta

• Rukunin GPU/TPU masu babban aiki: Haɗa dubban zuwa dubun-dubatar nodes don tallafawa fasahar babban horo mai rarraba aiki
• Tsarin aiki na ingantaccen horo mai rarraba aiki: DeepSpeed, Megatron-LM, Ray da sauransu
• Ma'ajiyar bayanai/cibiyar sadarwa: Yanayin da ke tallafawa shigarwa da fitar da manyan bayanai da kuma hanyar sadarwa mai sauri

Tsara gine-ginen samfurin

• Nuna tsarin zamani: Transformer, MoE(Mixture of Experts), multimodal da sauransu
• Scalability: la’akari da faɗaɗawa kamar adadin parameters, adadin layers, tsawon input da sauransu
• Inganci: inganta saurin horarwa/hasashe da amfani da memory

Dabarun horarwa da algorithms

• Manufar pretraining: language model (misali, next token prediction), multimodal (misali, contrastive learning) da sauransu
• Dabarun optimization: mixed precision, gradient accumulation, learning rate schedule da sauransu
• Normalization/stabilization: dropout, layer norm, weight decay da sauransu

Tsarin kimantawa da tabbatarwa

• Saitin benchmarks: amfani da daidaitattun datasets (Benchmarks)
• Kimantawar cikin gida: kimantawa bisa ainihin yanayin amfani
• Ci gaba da sa ido: duba inganci, son zuciya da tsaro yayin da bayan horarwa

Ƙwarewar ma’aikata da ƙungiya

• Masu binciken AI/ML: tsara model da haɓaka algorithms
• Injiniyoyin bayanai: tattarawa/tsarkakewa/ gudanar da bayanai
• Injiniyoyin infrastructure: tsarin rarrabawa da gudanar da cloud/on-premises
• Manajan ayyuka: sarrafa jadawali, kasafin kuɗi da inganci

La’akari da ɗabi’a, doka da tsaro

• Ɗabiyar AI: son zuciya, tsaro, gaskiya da ɗaukar alhaki
• Bin doka: bayanan sirri, haƙƙin mallaka da ikon mallakar bayanai
• Tsaro: hana fitar bayanai/model da sarrafa damar shiga

A halin yanzu, adadin masu binciken AI a duk duniya kusan 2,000 ne, kuma galibinsu sun tattaru ne a manyan kamfanonin fasaha kamar META, Google da XAI. A cikin gida, ƙungiyoyin da za su iya tsara komai da kansu daga tattara bayanai har zuwa manyan hanyoyin sadarwar horarwa domin ƙirƙirar Foundation model ba su da yawa matuƙa.

3. Ƙirƙirar babban language model da ƙaramin adadin ma’aikata.

Bisa gogewarsa ta fiye da shekaru 10 wajen binciken deep learning da machine learning, Shugaba Lee Hwan-ho ya kafa kamfanin (Clevi) shekaru 3 da suka wuce da manufar “ƙirƙirar AI mafi kyau a duniya”.

Ta hanyar tsara da kansa dukkan matakai, ciki har da gina bututun bayanan kamfanin, tsara manyan hanyoyin sadarwar deep learning da aka rarraba, haɓaka tsarin model da tabbatar da inganci, ya samu ƙwarewar da ake buƙata wajen haɓaka Foundation Model.

Sakamakon horar da model sau da dama, yanzu ya mallaki Clevi-5-x model wanda zai iya yin gasa da model na matakin mafi girma a duniya.

Don samun horar da manyan harsunan samfurin yadda ya kamata, ana buƙatar ababen more rayuwa da ke haɗa GPU ɗari zuwa dubbai a matsayin gungu, wanda zai iya aiwatar da ƙididdiga har sau tiriliyan da yawa a cikin daƙiƙa guda a lokaci ɗaya. A cikin wannan tsari, fasahar da ke rage jinkirin daidaita ƙididdiga da sadarwa a babban yanayin rarrabawa tana da muhimmanci. Har zuwa yanzu, saboda babu “algorithm ɗin ingantaccen sarrafawa” a cikin ƙasar, yawancin kamfanoni ba su iya horar da samfurai yadda ya kamata ba. Shugaban Clevi, Lee Hwan-ho, ya yi nasarar horar da babban samfurin harshen da ke da sigogin tiriliyan 1.4 (1.4 Trillion) a karon farko a cikin ƙasar, ta hanyar haɓaka wannan keɓaɓɓen algorithm ɗin sarrafa ƙididdigar rarrabawa da kansa.

4. Dalilin da ya sa aka iya haɓaka samfura daban-daban da ƙananan ma’aikatan haɓakawa

Yawancin manyan kamfanonin fasaha ma suna da fasahohin sarrafa manyan ababen more rayuwa da tsarkake bayanai.

Don sarrafa waɗannan, ana buƙatar ɗaruruwa zuwa dubbai na ma’aikatan bincike.

Amma Clevi tana haɓakawa da adana samfura daban-daban tare da ƙananan ma’aikatan bincike.

Ta yaya hakan ya yiwu?

Clevi tana haɓakawa da adana samfura daban-daban da ƙananan ma’aikata ta hanyar fasahar Teacher-Student Model(Knowledge Distilation).

Yanzu bari mu duba fasahar Teacher-Student.

Teacher-Student(ɗanɗanar ilimi)

A sauƙaƙe, fasahar Teacher-Student tana amfani da babban samfurin harshe (Teacher Model) don tantancewa da ba da ra’ayi kan samfurin ɗa (Student Model), ta yadda za a iya maye gurbin ɗaruruwan ma’aikatan bincike da ƙananan ma’aikata tare da haɓaka samfura daban-daban cikin sauri.

• Mother Model (babban samfurin) yana tantancewa da ba da ra’ayi kan samfuran fannoni daban-daban, yana maye gurbin ɗaruruwan ma’aikatan bincike.
• Idan aka gudanar da horo ta wannan hanya ta Clevi-x-platform, ana iya horar da samfura masu babban aiki kamar Reasoning, VLM, Physical AI, Coding Agent da sauransu, sannan a tura su cikin kwanaki 10 zuwa 15.

Clevi Coding Agent

Clevi Phsycal AI Learning Platform

VLM-Vision Language Model

Wakilin tsaro

5. Bambancin fasaha da inganci na Clevi-x-platform

Ayyukan samfurin da faɗaɗawarsa

• Fifiko wajen tunani (CoT/Reasoning): cip-5-x ya haɗa da ci gaban tunani na matakai da gabatar da hujjoji a cikin falsafar ƙirarsa, yana nuna ingantaccen ikon lissafi da fassara wajen warware matsalolin kimiyya, lissafi da injiniyanci. Ana ƙirarsa da sarrafa shi da nufin cimma matakin aiki da ya kai ko ya wuce GPT-5 bisa ma’aunin gwaje-gwajen cikin gida, yayin da ake sarrafa maimaitawa da iya tabbatarwa a ƙarƙashin yanayin prompt iri ɗaya a matsayin muhimman ma’aunan inganci.

• Cikakken bakan multimodal: Ana iya samarwa da haɗa VLM mai mai da hankali kan manufa (cip-5-vision), babban samfurin sarrafa bayanan multimodal (ivy-4-mm), da ivy-3-text mai sauƙi don aiki a kan na’ura, a kan dandali guda, domin zaɓar haɗin da ya fi dacewa da halayen aiki (bincike/rarrabawa/takaitawa da tunani/bayyanawa/aiwatarwa).

Dacewa da aikace-aikacen kamfanoni

• Tsaro da samuwa a kan tsarin cikin gida: Ana sarrafa dukkan matakai a cikin rufaffiyar hanyar sadarwa (shigarwa-horarwa-sabis-ajiyar bayanai), tare da ƙirar HA, faɗaɗawa ta hanyar kayayyaki, da SVCE (keɓaɓɓen amintaccen yanayin aiwatarwa) domin kare bayanai da sigogin samfurin daban-daban.
• Ɗaukarwa da faɗaɗawa cikin sauri: Ivy Chat (tattaunawa/agents na multimodal don ayyukan kasuwanci) da API Platform (daidaitaccen SaaS na duniya) suna tallafawa aiwatarwa da sarrafawa cikin sauri.

Bambancin AI na zahiri (Physical AI)

• Haɗa kwaikwayon da ya dogara da NVIDIA Isaac da koyon ƙarfafawa bisa juyin halitta (Evolutionary RL), tare da canjin Sim2Real da horarwa iri ɗaya ta amfani da bayanan roba, ya inganta ingancin horarwa da ikon daidaitawa a wurin aiki. Ƙarancin madadin da ke rufe komai daga koyon dijital da na’ura-robot zuwa turawa ya sa wannan ya bambanta.

Ingancin sabis da shugabanci

• Ana sarrafa sigar samfurai/prompt/kayan aiki, tsarin tantancewa (ilimin Koriya da Ingilishi, ayyuka na masana’antu daban-daban, ma’aunan ruɗu da tsaro), sarrafa canje-canje (SLO/SLA), da sa ido kan aiki (jinkiri, ƙimar nasara, TCO) ta hanyar tsari guda na dandali.

A halin yanzu, akwai kamfanonin sabis na samfurorin AI sama da 300 a cikin Koriya ta Kudu, amma

CLEVI ce kaɗai ke iya samar da sabis na cikin gida da na gajimare a lokaci guda ta amfani da nata babban samfurin tushe mai babban aiki.

6. Hanyoyin amfani da samfurin harshe a masana’antu

Tun da gabatar da AI na buƙatar babban saka jari da cikakken tabbatarwa, yana da matuƙar muhimmanci a kwatanta da gwada mafita kai tsaye domin tabbatar da cewa tana taimaka wa kamfani a zahiri.

• CLEVI ba ta tsaya kawai wajen ƙirƙirar samfura ba, tana kuma samar da hanyoyin AI da za a iya amfani da su a ainihin muhallin masana’antu.
• Misali, tana da cikakkun tashoshin kasuwanci na kamfanoni, kamar Ivy Chat Platform, Clevi API Platform, keɓancewa bisa masana’antu, da biyan buƙatun tsaro.
• Tana da ƙwarewar fasaha da ba kasafai ake samu a cikin gida ko ƙasashen waje ba, a fannoni kamar AI na zahiri, robotics, da sarrafa bayanan multimodal.

CLEVI tana samar da ingantattun hanyoyin AI waɗanda ke kallon AI a matsayin ‘hanya’ ba ‘manufa’ ba, kuma suke ba da gudummawa ga ingancin aiki na zahiri da ƙirƙire-ƙirƙiren masana’antu. Ku dandana da kanku bambancin ainihin From-scratch Foundation model.

Tambaya da neman demo: [email protected]

Copyright© 2025 Clevi Inc. All rights reserved.