Komawa Labarai
Bidi'aAI Understanding takaitaccen bayani

Taswirorin takarda na Abra suna ƙididdigewa da cinikin bayanai a cikin horar da hoto na watsawa

Wani sabon binciken arXiv yana gabatar da gwaje-gwajen doka don ƙirar rubutu-zuwa-hoto a cikin lissafin kasafin kuɗi daga 10^19 zuwa 10^22 FLOPs. Marubutanta sun ba da rahoton cewa ƙirar hoto na buƙatar ƙarin bayanai a kowane siga fiye da ƙirar harshe don horarwa da inganci.

5 min readRead the primary source
Source-provided image accompanying Abra paper maps compute and data tradeoffs in diffusion image training
Takardun tushe na farkoAn rubuta tushen tushe
Mawallafi
arxiv.org
Tushen hanyar haɗin gwiwa
arxiv.orghttps://arxiv.org/abs/2608.17286
Nau'in tushe
Takardun farko - sanarwar hukuma, takarda, yin rajista, ko shafi na farko da muka karanta kai tsaye.
MaganaFahimtar wannan a cikin daƙiƙa 60

Fara a nan

Mabuɗin sharuddan

Yi lissafi
Abubuwan sarrafawa da ake buƙata don horarwa da gudanar da ƙira, galibi ana auna su cikin sa'o'in FLOPS ko GPU.
Samfurin Yaduwa
Gine-ginen ƙirƙira wanda ke koyon juyar da hayaniya don haɗa hotuna, sauti, ko wani abun ciki.
Asarar horo
Ƙimar kuskuren ƙirar ƙirƙira yayin horo kuma an inganta shi ƙasa akan lokaci.
Gwada kankaAI Model An Bayyana Tambayoyi

Me ya faru

Masu bincike sun gabatar da Abra, dangin da ke sarrafawa na samfuran masu canzawa masu dacewa da kwarara, don nazarin yadda tsarin watsa rubutu-zuwa-hoto sikelin tare da ƙididdigewa da bayanai. Takardar ta ba da rahoton cewa halayyar ƙima ta kasance ana iya tsinkaya a cikin kewayon ƙididdigewa da yawa fiye da binciken da aka yi a baya.

Takardar, wacce aka ƙaddamar ga arXiv a kan Agusta 18, 2026, tana nazarin dokokin ƙima don ƙirar watsa rubutu-zuwa-hoto. Mawallafansa takwas sun gabatar da Abra a matsayin dangin sarrafawa na masu canji masu daidaitawa, suna ba su damar bambanta ma'aunin horo yayin nazarin yadda halayen ƙirar ke canzawa. Ƙididdigar ta ce gwaje-gwajen sun ƙunshi umarni uku na girman ƙididdigewa, daga 10^19 zuwa 10^22 ayyuka masu iyo, kuma sun kai ga babban kasafin kuɗi fiye da aikin da ya gabata. Waɗannan ikirari ne da jaridar ta yi; Tushen da aka kawo ba ya samar da teburin da ke ƙasa, hanyoyin, ko kwatancen gwaji.

Marubutan sun ba da rahoton cewa samfuran watsawa sun yi girma kamar yadda ake iya tsinkaya kamar ƙirar harshe, amma cewa mafi kyawun wurin horon su yana buƙatar ƙarin bayanai. Musamman, maƙasudin yana sanya mafi kyawu a kusan alamomin hoto 200 a kowace siga, wanda aka kwatanta a matsayin sau goma mafi kyawun lissafin Chinchilla don samfuran manyan harshe. A aikace, takarda ta yi jayayya cewa samfurin yaduwa ya kamata a fallasa gabaɗaya ga adadi mai yawa na bayanan hoto dangane da ƙidayar sigar sa lokacin da ƙungiyar ke ƙoƙarin samun mafi daga ƙayyadaddun kasafin horo. Madogararsa baya ayyana madaidaicin wakilcin alamar hoto ko bayyana yadda rabon ya canza a cikin ingancin bayanai, ƙuduri, ko zaɓin taken.

Har ila yau binciken ya ba da rahoton cewa samfuran watsa shirye-shiryen suna da ƙarfi kwatankwacin ƙarfin horo. A kan haka, marubutan sun ba da shawarar cewa masu yin aiki su yi kuskure wajen yin amfani da ƙarin bayanai maimakon gina babban samfuri. Sun ce tsinkaya iri ɗaya ya wuce asarar horo zuwa ma'auni masu inganci, saitunan jagora marasa kyauta, ingancin wakilci, da kuma sifar horo, waɗanda suka ce rugujewa cikin tsari na duniya. Abubuwan da aka kafa masu zaman kansu da ke akwai a nan an iyakance su ne ga rikodin littafin littafin da zayyana. Majiyar ba ta tabbatar da cewa an sake maimaita binciken ba, an sake duba takwarorinsu, an fitar da su azaman abin ƙira, ko kuma ɗauka a samarwa.

A hade, wannan sashe yana gabatar da binciken a matsayin rahoton gwaje-gwaje da fassarar marubutan, ba a matsayin sakamako mai zaman kansa ba. Yana bayyana kewayon lissafin da aka bayar da rahoton, shawarwarin bayanai-zuwa-misali, da kuma halayen da aka bayar da rahoton a ƙarƙashin overtraining, yayin da ake kiyaye iyakokin rikodin da aka kawo. Ƙididdigar ta ba da ƙaddamarwar takarda a matsayi mai girma, amma ba ya samar da tebur, hanyoyi, kwatancen, kwafi, ko shaidar ɗauka. Waɗancan bambance-bambancen suna daga cikin abin da tushe ya kafa da abin da ya bari ba a warware shi ba.

Bayanan tushe: arxiv.org

Me ya sa yake da mahimmanci

Idan binciken ya wuce gwaje-gwajen marubutan, za su iya canza yadda ƙungiyoyi ke keɓe kasafin horo don ƙirar ƙirƙira hoto. Babban ma'anar ita ce ƙara bayanai na iya zama mafi inganci fiye da ci gaba da haɓaka girman ƙirar, mai yuwuwar tasiri farashi, tsara kayan aikin, da haɓaka ƙirar ƙira.

Takardar ta yi magana game da ainihin matsalar tsarawa a cikin haɓaka ƙirar hoto: yadda za a raba ƙayyadaddun albarkatu tsakanin girman samfurin, bayanan horo, da ƙididdigewa. Amintaccen ma'auni mai mahimmanci zai iya taimaka wa masu bincike kimanta yawan ci gaban da za su yi tsammani daga mafi girma samfurin ko horo mai tsawo kafin yin gwaji mai tsada. Zaɓin zaɓin da aka ruwaito don ƙarin bayanai na iya sake karkatar da saka hannun jari zuwa tarin tarin bayanai, tacewa, adanawa, aiwatarwa, da lasisi maimakon zuwa sigogi kaɗai.

Kwatanta tare da aikin ƙirar harshe yana da yuwuwar mahimmanci saboda ƙa'idodin ƙirar harshe-samfurin ya zama maƙasudin gama gari don hasashen sakamakon horo. Ƙididdigar rahoton marubutan na kusan alamomin hoto 200 a kowace ma'auni ya nuna cewa canja wurin rubutun-samfurin harshe kai tsaye zuwa tsarar gani na iya haifar da ƙungiyoyi zuwa ƙirar hoto na ƙasa akan bayanai. Wannan ƙaddamarwa na iya zama mahimmanci ga manyan dakunan gwaje-gwaje da ƙananan ƙungiyoyi waɗanda ke ƙoƙarin yin ingantaccen amfani da ƙayyadaddun kayan aiki. Ba ya, da kanta, ya nuna cewa ƙarin bayanai za su inganta kowane samfurin hoto ko ƙarin bayanan za su kasance masu araha, masu amfani da doka, bambanta, ko kyauta daga kwafi da ƙananan misalai.

Ƙarfin da aka bayar da rahoton zuwa overtraining zai iya sanya yanke shawara na horo ya zama mai gafartawa, musamman ma lokacin da ƙungiya ta sami damar yin amfani da manyan bayanai amma ba za ta iya gane ainihin ma'anar raguwar dawowa ba. Mafi girman da'awar-cewa tsarin yau da kullun kuma yana hasashen ingantattun ma'auni, saitunan jagora, wakilci, da sifofin horo-zai fi dacewa idan ya tsira gwaji a wajen saitin sarrafa Abra. Koyaya, ƙayyadaddun bayanan yana ba da ƙimar awo, kwatancen samfurin, nazarin kuskure, bayanin saitin bayanai, lissafin kuzari, ko shaida game da amfanin ƙasa. Don haka yana goyan bayan ɗaukar wani sanannen sakamakon bincike, yayin da yake barin girman da ingantaccen tasirin tasirinsa mara tabbas.

Interactive Mechanism

Ingantacciyar hanyar sadarwa: Yadda A zahiri yake Aiki

Bincika fasahar da ke bayan wannan ci gaban ta hanyar mu'amala.

Model Parameter Size:8B Parameters
VRAM Required5.5 GBGPU memory footprint
Target HardwareMacBook / Single GPUDeployment tier
Privacy100% Air-GappedLocal device capability
Core takeaway: Small, quantized models (3B–8B) now run directly inside smartphones and laptops with complete data privacy, while mammoth 400B+ models remain the domain of datacenter clusters.
Duba ra'ayi na hulɗa+10 Points
AI Models Explained Quiz

In AI, what are a model's "parameters"?

Abin kallo na gaba

Rikodin arXiv da aka kawo bayanai ne, ba tabbataccen sakamako ba. Tambayoyi masu mahimmanci sun haɗa da waɗanne kididdigar bayanai da kimantawa aka yi amfani da su, ko haɓaka alaƙar haɓakawa zuwa wasu gine-gine da ayyukan ƙirƙira hoto, da ko ribar da ake da'awar ta tabbatar da ƙarin bayanai da ƙididdige abin da ake buƙata.

Babban fifiko na farko shine tabbatar da hanya a cikin cikakkiyar takarda. Ya kamata masu karatu su nemi ainihin tsarin gine-gine da jeri na ma'auni, ƙudurin hoto, saitin sanyaya rubutu, tushen bayanai, ƙaddamarwa da hanyoyin tacewa, da yadda ake ƙidayar alamun hoto. Ƙididdigar ta gano kewayon ƙididdigewa amma bai faɗi adadin samfura nawa aka horar a kowane batu, nawa aka gudanar da gudu masu zaman kansu ba, ko kuma yadda aka ƙiyasta rashin tabbas game da ƙaƙƙarfan ƙa'idodin ƙima.

Zane-zanen kimantawa zai ƙayyade yadda za'a iya amfani da ƙarshen ƙarshe. Rikodin yana nufin ma'aunin ingantacciyar ƙira, jagora mara ƙima, da ingancin wakilci amma babu sunan ko ɗaya daga cikin ma'auni kuma ba shi da ƙima. Har yanzu ba a fayyace ko alaƙar da aka ruwaito tana riƙe da hukunce-hukuncen ɗan adam, saurin riko, abun da ke ciki, rubutun rubutu, ra'ayoyi da ba kasafai ba, abun ciki mai aminci, ko gyaran hoto. Har ila yau, ba a sani ba ko ana canja sakamakon sakamakon zuwa gine-ginen ban da dangin Abra da ke sarrafawa, zuwa shawarwari da tsari daban-daban, ko zuwa bayanan kasuwanci tare da rarrabawa daban-daban.

Maimaitawa da samun dama ma suna da mahimmanci. Madogarar ta gano takarda mai shafi 25 mai adadi 19 da hanyoyin haɗin kai zuwa tushen PDF da TeX, amma rubutun da aka kawo bai bayyana cewa akwai lambar horo, ma'aunin ƙira, bayanan bayanai, ko rubutun tantancewa ba. Ya kamata aikin bin diddigin ya gwada rabon bayanai-zuwa-madaidaitan bayanai akan madaidaitan bayanai masu zaman kansu da kayan masarufi, auna farashin kuɗi da makamashi na ƙarin bayanan, da bincika ko overtraining ya kasance mara kyau lokacin da bayanai ke ɗauke da kwafi, rikice-rikicen lasisi, ko son zuciya mai cutarwa. Har sai an amsa waɗannan tambayoyin, an fi fahimtar takarda a matsayin nazarin ƙa'idodi na ƙima da tsarin shawarwarin da aka goyi bayan bincike, ba a matsayin hujja na ƙa'idar duniya ba ga duk tsarin hoto mai yaduwa.

Jagorori masu alaƙa & tambayoyin tambayoyi

AI Model ya bayyanaAI horoMasu canjiMakomar AIGwada abin da kuka sani - gwada gwajin AI kyautaNemo kalmar AI a cikin ƙamus ɗin mu
An sami wannan yana da amfani?