Kordhinta dhererka YaRN iyo macnaha guud
YaRN (Haddana xadhig kale oo dheeri ah) waa farsamo hufan oo lagu fidiyo moodeelka macnaha guud ee la isticmaali karo oo ka fog wixii lagu baray.
Dulmar
It cleverly rescales rotary position embeddings so a model trained on, say, 4K tokens can handle 32K or more with minimal fine-tuning.
quusid qoto dheer
Inta badan LLM-yada casriga ahi waxay calaamadeeyaan boosaska calaamadaynta ee RoPE (Rotary Position Embeddings), kaas oo u rogaaya weydiinta iyo qaybaha muhiimka ah xagal ku xidhan booska. Markaad quudiso taxanaha in ka badan dhererka tababarka, wareegyadani waxay galaan kala duwanaansho aan la arkin oo qaabku wuu burburaa. YaRN, oo lagu soo bandhigay 2023 by Bowen Peng iyo la-hawlgalayaasheeda, waxay tan ku hagaajisaa isdhexgalka NTK-aware ee lagu dabaqo inta jeer: waxay ka tagtaa cabbirrada soo noqnoqda (kaas oo qabta xiriirka maxalliga ah, kuwa gaaban) inta badan lama taaban iyada oo la dhexgelinayo cabbirrada soo noqnoqda hooseeya (taas oo la socota booska fogaanta). YaRN waxay sidoo kale ku darsataa hagaajinta heerkulka dareenka si looga hortago isbeddelada entropy ee ka yimaada xaaladaha dhaadheer. Natiijadu waa wax qabad xoog leh oo muddada dheer ah ka dib markii si fiican loo habeeyey qayb yar oo ka mid ah xogta iyo tillaabooyinka ay u baahan yihiin habab aan caqli-gal ahayn.
Aragtida Farsamada
Xadhiggu wuxuu u qoondeeyaa cabbir kasta oo ku-xidhan inta jeer oo wareeg ah. Dhexdhexaadinta tooska ah ee tooska ah ayaa isku si isku mid ah u cadaadisa dhammaan soo noqnoqoshada, waxyeelaysa cabbirrada soo noqnoqda sare ee qeexaya tafaasiisha maxalliga ah ee wanaagsan. YaRN waxay isticmaashaa hawl qabad si ay u dhexgaliso kaliya cabbirada hoose ee soo noqnoqonaya (mawjadaha dhaadheer) iyadoo la ilaalinayo kuwa soo noqnoqonaya, oo lagu daray 1/sqrt (t) heerkulka dareenka heerkulka kaas oo ilaalinaya jilicsanaanta jilicsan ee xasilloon sida dhererka isku xigxiga u korayo. Habkan NTK-qayb-qayb-qayb-qaybeed ayaa ku fidinaya macnaha guud oo hoos u dhac aad u yar.
Saamaynta Istiraatijiyadeed
Qiimaha iyo miisaaniyada
Go'aamada qaab-dhismeedku waxay horseedaan waxqabadka iyo kharashka hawlgalka sannadaha.
Go'aamo cad
Waxbarashada farsamada waxay ka caawisaa kooxaha inay doortaan xidhmo sax ah, ma aha oo kaliya kan ugu cusub.
Xakamaynta tayada
Doorashooyinka injineernimada ee wanaagsan waxay yareeyaan shilalka la isku halleyn karo ee wax soo saarka.
Mustaqbalka YaRN iyo Kordhinta Dhererka macnaha
Kordhinta macnaha guud hadda waa dhaqanka caadiga ah: moodooyinka furan waxay si joogto ah u raraan noocyada dheeraynta YaRN oo gaaraya 128K calaamado ama ka badan. Cilmi baaristu waxay u socotaa habab kordhinaysa macnaha eber ama meel u dhow hagaajinta ganaaxa, isku darka RoPE dib u habeynta iyo khiyaamada qaabaynta, oo ilaali tayada daaqada buuxda halkii ay ka ahaan lahayd dhamaadka kaliya. Filo is dhex galka adag ee farsamooyinkan si loo sii tababaro si markaas macnaha guud uu yahay hooyo halkii dib loo habeyn lahaa.
Dhaqangelinta Adduunka-dhabta ah
Kordhinta qaabka furan ee 4K-context ilaa 32K ama 128K si loogu helo su'aal qoraal ah oo dheer oo ka jawaabaya hagaajin gaaban
Awood-siinta nidaamyada dib-u-soo-celinta-kordhinta si ay u galaan marino badan oo isku xiran iyada oo aan la jarin
Awood siinta caawiyayaasha kood u baahan fayl weyn oo kayd ah ama faylal badan hal degdeg ah
La qabsiga qaabka saldhiga ee wada sheekaysi-wareegyo badan oo dhaadheer oo urursan taariikho sheeko oo waaweyn
Khatarta & Dariiqyada Ilaalada
Hagaajinta hal bartilmaameed waxay qarin kartaa daciifnimada nidaamka ballaaran.
Kaabayaasha dhaqaalaha iyo dayactirka inta badan waa la dhayalsadaa.
Nabadgelyada iyo daldaloolada u fiirsashada ayaa kori kara marka nidaamyadu noqdaan kuwo aad u adag.
Qorshe Hawleedka Dhaqangelinta
Qeex daahida, tayada, iyo bartilmaameedyada qiimaha ka hor inta aan la hirgelin.
Benchmark marka la eego culeyska dhabta ah iyo xaaladaha xogta.
La socodka qalabka khaladaadka, leexashada, iyo saamaynta isticmaalaha.
U diyaari dib-u-noqoshada iyo dariiqyada jawaab-celinta dhacdada ka hor inta aanad miisaan.
Sii wad Sahaminta
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 YaRN and Context Length Extension 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
Hagaha xiga
Dhexdhexaadinta Booska ee Kordhinta Macnaha guud
Su'aalaha soo noqnoqda
What is YaRN and Context Length Extension?
YaRN (Haddana xadhig kale oo dheeri ah) waa farsamo hufan oo lagu fidiyo moodeelka macnaha guud ee la isticmaali karo oo ka fog wixii lagu baray. Waxay si xariifnimo leh dib u cabbiraysaa boosaska wareegtada ah si ay moodal lagu tababaro, dheh, 4K token waxay xamili karaan 32K ama ka badan iyadoo la hagaajinayo ugu yar.
Habka codaynta booska ayay YaRN wax ka beddeshaa si ay u kordhiso macnaha guud?
YaRN, oo magaciisa macnihiisu yahay Weli xadhig kale oo dheeri ah, ayaa dib u miisaamaya boosaska wareega ee loo isticmaalo inta badan LLM-yada casriga ah.
Maxaa khaldama marka nooca RoPE uu arko taxane ka dheer dhererkiisa tababarka?
Jagooyinka ka baxsan tababbarka waxay soo saaraan xaglo wareeg ah oo qaabku waligiis arkin, markaa habdhaqanka feejignaanta ayaa si aad ah hoos ugu dhacaya.
Sidee ayuu YaRN u daaweeyaa cabbirrada soo noqnoqda ee xad-dhaafka ah ee kala duwan?
YaRN waxay isticmaashaa NTK-qayb-qaybo ramp kaas oo isku xira mawjada-dheer ee dhumucyada hoose ee soo noqnoqda oo ka taga mugdiga-soo noqnoqda-gaaban inta badan.
Marka laga reebo dib-u-cusboonaynta soo noqnoqoshada, waa maxay hagaajinta dheeraadka ah ee YaRN ay khusayso?
YaRN waxay ku darsataa heerkulbeeg heerkulbeegyo si ay uga hortagto isbeddellada entropy ee dhaca marka macnaha guud uu kordho.
Waa maxay faa'iidada dhabta ah ee muhiimka ah ee YaRN marka loo eego isdhexgalka naive?
YaRN waxay ku guulaysataa tayada nuxurka dheer ee xooggan ka dib markii si fiican loo habeeyey qayb yar oo ka mid ah hababka caqli-galnimada ah ee xogta loo baahan yahay.