Isku darka Kediska Khubarada LoRA

Jawaabaha iyo sharraxaadaha waxay soo baxayaan markaad qaadanayso kediska.

1. Waa maxay qaybta 'LoRA' ee isku dhafka khubarada LoRA?
   • La qabsiga Darajada-hoose, kaas oo tababara maaddooyinka darajada hoose ee yaryar halkii uu ka ahaan lahaa miisaan buuxa
   • Qaabka abaalmarinta barashada xoojinta
   • Hab fiiro gaar ah oo fog
   • Dhuumaha xogta ku shubaya adapters
2. Waa maxay shaqada shabakada gating/router ee MoLE?
   • Si loo cadaadiyo miisaanka qaabka saldhigga
   • Si loo go'aamiyo khubarada la hawlgelinayo iyo culeyska loogu talagalay gelinta la bixiyay
   • Si loo soo saaro xogta tababarka ee adaabiyeyaasha
   • Si aad qoraalka ugu beddelo gundhigyo
3. Waa maxay sababta MoLE ay inta badan uga fiican tahay tababbarka hal adabiyeyaasha hawlo badan oo isku dhafan?
   • Waxay isticmaashaa xusuusta GPU badan oo naqshadaysan
   • Waxay meesha ka saaraysaa baahida loo qabo nooc kasta oo saldhig ah
   • Waxay yaraynaysaa faragelinta iyo illoobidda iyada oo xirfadaha lagu hayo khubaro kala duwan oo isku dhafan
   • Waxa ay ku shaqeysaa hal luqad kaliya
4. LoRA gudaheeda, cusboonaysiinta miisaanka ayaa sida caadiga ah loo muujiyaa midkee kuwan soo socda?
   • Matrix cufan oo buuxa oo beddelaya W
   • Hal-dhufane scalar ah oo ku yaal W
   • Matrix qaylo ah oo random ah ayaa ku daray tallaabo kasta
   • Wax soo saarka laba shaybaar oo darajo hoose, B times A
5. Inta lagu jiro tababarka MoLE, maxaa guud ahaan ku dhacaya miisaanka la sii tababbaray ee moodeelka?
   • Waxay ku jiraan baraf halka adaabiyeyaasha iyo router-ka la tababaray
   • Si buuxda ayaa loo tabobaray meel eber ah
   • Waa la tirtiray oo dib loo dhisay
   • Waxaa lagu qiyaasaa hal xoogaa
6. Qaababka sawirada faafinta, waa maxay dhibka uu ka caawiyaa xalinta kala sareynta/lakabka gating ee MoLE?
   • Soodejin internet oo gaabis ah
   • Fikrado badan oo LoRAs ah oo midba midka kale ka awood badan marka la sameeyo
   • La'aanta xogta tababarka
   • Qaababka faylka khaldan
7. How can a gated LoRA mixture differ from merging all adapter weights into one fixed update?
   • It must retrain the pretrained backbone for every prompt
   • It can use only one expert for all inputs
   • It can select or weight experts conditionally for a task or layer
   • It removes every low-rank matrix before inference
8. In a LoRA update with rank r, what does r control most directly?
   • The number of decoder tokens generated
   • The number of original model layers removed
   • The size of the training vocabulary
   • The dimension of the low-rank factors and thus the capacity and parameter cost of the update