Horon Hasashen Multi-Token
Maimakon yin tsinkaya kawai alamar na gaba, an horar da samfurin don yin hasashen alamu da yawa na gaba lokaci guda.
Dubawa
This sharpens learning signals and unlocks faster inference through self-speculative decoding.
Zurfafa nutsewa
An horar da daidaitattun ƙirar harshe tare da tsinkayar alama ta gaba: idan aka ba da mahallin, tsinkaya alamar gaba guda ɗaya. Hasashen Alamar Multi-Token (MTP), wanda takarda ta 2024 Meta ta shahara kuma aka karɓa a cikin DeepSeek-V3, yana ƙara ƙarin shugabannin fitarwa masu nauyi don haka samfurin lokaci guda yana annabta alama ta gaba tare da alamu na 2nd, 3rd, da 4th gaba daga ɓoye ɗaya. Wannan yana tilasta cibiyar sadarwa don yin shirin gaba zuwa gaba kuma yana ƙaddamar da siginar horo - kowane matsayi yanzu yana ba da gudummawar sharuɗɗan asarar da yawa. Meta ya ruwaito musamman manyan ribar da aka samu akan ƙididdigewa da tunani mai ƙima, tare da manyan samfura suna cin gajiyar ƙarin. Mahimmanci, ana iya jefar da ƙarin kawunan bayan horo, don haka girman samfurin a lokacin turawa baya buƙatar girma.
Fahimtar Fasaha
MTP yana liƙa n kawunan tsinkaya masu zaman kansu a saman akwati da aka raba; kai k yana tsinkayar alamar a matsayi t +k daga wakilci a matsayi t. Ana tattara asarar da aka yi yayin horo. Dangane da ra'ayi, shugabannin mataimakan suna ba da damar ƙididdige ƙididdiga na kai-tsaye: ƙirar ta ba da shawarar alamu da yawa a cikin fasfo ɗaya, sannan ta tabbatar da su, ta cimma kusan tsararru 3x cikin sauri ba tare da canza rarraba kayan sarrafawa ba.
Dabarun Tasiri
Gudu da sikelin
Gudun aikin harshe na iya tafiya da sauri ba tare da sadaukar da daidaito ba.
Shiga ku isa
Yana faɗaɗa damar shiga cikin harsuna da salon sadarwa.
Shawarwari masu haske
Ƙungiyoyi za su iya ciyar da ƙarin lokaci akan hukunci yayin da aiki da kai ke sarrafa maimaitawa.
Future of Multi-Token Hasashen Horo
MTP yana zama sinadari na asali a cikin girke-girke na horar da kan iyaka saboda yana inganta duka inganci da saurin ƙididdigewa a ƙaramin farashi. Yi tsammanin haɗin kai tare da ƙididdige ƙididdiga, zurfin hangen nesa, da amfani da matsayin maƙasudin taimako wanda ke inganta shirin dogon sararin sama. Haɗe da ƙirar tunani, tsinkayar matakai da yawa a gaba na iya taimakawa ƙila su kwaikwayi sakamako a ciki kafin yin amsa.
Aiwatar da Gaskiyar Duniya
DeepSeek-V3 ta amfani da makasudin MTP yayin horarwa don haɓaka ingantaccen bayanai da ba da damar ƙididdige ƙima.
Meta's model-generation code nuna daidaito nasarori akan HumanEval da MBPP daga tsinkayar alamu da yawa.
Ƙididdigar ƙididdiga ta kai-da-kai: zayyana alamun 3-4 a kowane wucewar gaba sannan tabbatarwa da sauri, fitarwa-kiyayyar fitarwa.
Mai saurin cikawa da sauri a cikin mataimakan coding inda aka gabatar da alamu da yawa da aka yi la'akari da su a mataki ɗaya.
Hatsari & Tsare-tsare
Abubuwan da aka ruɗe suna iya shigar da rahotanni cikin nutsuwa, kwararar tallafi, ko abubuwan bincike.
Hankali na gaggawa na iya ƙirƙirar sakamako mara daidaituwa a cikin buƙatun iri ɗaya.
Za a iya fallasa bayanan rubutu mai ma'ana idan ikon samun dama yana da rauni.
Taswirar Hanya
Ƙayyade tsarin fitarwa, sautin, da ma'auni masu inganci kafin fitowa.
Amsa a ƙasa tare da amintattun tushe a duk lokacin da daidaito ya shafi mahimmanci.
Ajiye wurin binciken ɗan adam don abubuwan da ake samu masu girma.
Bibiyar tsarin gazawar kuma sake horar da tsokaci ko tafiyar aiki akai-akai.
Ci gaba da Bincike
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Multi-Token Prediction Training quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Jagora na gaba
Hasashen Yawo da Hasashen Multi-Token
Tambayoyin da ake yawan yi
What is Multi-Token Prediction Training?
Maimakon yin tsinkaya kawai alamar na gaba, an horar da samfurin don yin hasashen alamu da yawa na gaba lokaci guda. Wannan yana kaifafa siginonin koyo kuma yana buɗe ra'ayi cikin sauri ta hanyar ƙididdige ƙididdiga na kai.
Menene tsinkayar alamar alama ta ƙara idan aka kwatanta da daidaitaccen horo na gaba?
MTP yana ƙara ƙarin shugabannin fitarwa don haka ƙirar ta yi hasashen alamar gaba tare da alamu da yawa gaba gaba daga ɓoye ɗaya.
Wanne samfuri ne ya yi amfani da manufar Hasashen Alamu da yawa wajen horarwa?
DeepSeek-V3 ya karɓi manufar horon MTP don inganta ingantaccen bayanai da ba da damar yanke ƙira.
Me yasa MTP ke ƙirƙira siginar horo?
Hasashen alamu da yawa na gaba yana nufin kowane matsayi na alama yana haifar da asarar hasashe da yawa, yana ba da ƙarin siginar koyo kowace alama.
Wane fa'ida ne ƙarin shugabannin hasashen ke ba da damar?
Shugabannin mataimakan za su iya tsara alamomi da yawa a kowane fasfo waɗanda aka tabbatar da su, suna hanzarta tsarawa ba tare da canza rarraba kayan sarrafawa ba.
Menene zai faru da shugabannin taimako bayan horo idan ba ku buƙatar saurin-sauri?
Ƙarin shugabannin na zaɓi ne a lokacin turawa; Yin watsi da su yana kiyaye ƙirar girman daidai da daidaitaccen samfurin gaba-gaba.