Idanwo ti o sopọ mọ itọsọna · Lile Ipele

Adalu ti LoRA Amoye adanwo

Checks LoRA’s low-rank updates, expert gating, frozen base weights, composable skills and per-layer routing for mixed adapters.

Jẹmọ itọsọna awọn ọnaAdalu Of Lora Amoye
Ibeere 1 ti 8

Kini apakan 'LoRA' ti Adalu ti Awọn amoye LoRA tọka si?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Kini apakan 'LoRA' ti Adalu ti Awọn amoye LoRA tọka si?
    • Iṣatunṣe ipo-kekere, eyiti o ṣe ikẹkọ awọn matiri kekere ipo-kekere dipo awọn iwuwo kikun
    • Awoṣe ère ẹkọ imuduro
    • A gun-ibiti o akiyesi siseto
    • Opo gigun ti n ṣajọpọ data fun awọn oluyipada
  2. Kini iṣẹ ti gating/nẹtiwọọki olulana ni MoLE?
    • Lati compress awọn iwọn awoṣe mimọ
    • Lati pinnu iru awọn amoye lati mu ṣiṣẹ ati pẹlu awọn iwọn wo fun titẹ sii ti a fun
    • Lati ṣe ipilẹṣẹ data ikẹkọ fun awọn oluyipada
    • Lati yi ọrọ pada sinu awọn ifibọ
  3. Kini idi ti MoLE nigbagbogbo dara ju ikẹkọ ohun ti nmu badọgba kan lori ọpọlọpọ awọn iṣẹ ṣiṣe ti o dapọ?
    • O nlo iranti GPU diẹ sii nipasẹ apẹrẹ
    • O ṣe imukuro iwulo fun awoṣe ipilẹ eyikeyi
    • O dinku kikọlu ati igbagbe nipa titọju awọn ọgbọn ni lọtọ awọn amoye composable
    • O ṣiṣẹ nikan lori ede kan
  4. Ni LoRA, imudojuiwọn iwuwo ni igbagbogbo kosile bi ewo ni atẹle naa?
    • Matrix ipon ni kikun ti o rọpo W
    • Ilọpo scalar kan lori W
    • A ID ariwo matrix kun kọọkan igbese
    • Ọja ti awọn matirisi ipo kekere meji, awọn akoko B A
  5. Lakoko ikẹkọ MoLE, kini gbogbogbo yoo ṣẹlẹ si awọn iwuwo ti a ti kọkọ tẹlẹ ti awoṣe?
    • Wọn duro ni aotoju lakoko ti awọn oluyipada nikan ati olulana ti ni ikẹkọ
    • Wọn ti wa ni kikun retrained lati ibere
    • Wọn ti paarẹ ati tun ṣe
    • Wọn ti wa ni iwọn si ọkan bit
  6. Ninu awọn awoṣe aworan itanka, iṣoro wo ni akosoagbasomode / fun-Layer gating ni MoLE ṣe iranlọwọ lati yanju?
    • Awọn igbasilẹ intanẹẹti o lọra
    • Ọpọ Erongba LoRAs bori kọọkan miiran nigba ti kq
    • Aini ti ikẹkọ data
    • Awọn ọna kika faili ti ko tọ
  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