quiz bu lëkkale ak tegtal · Dëgër Niveau

Quiz Ingénieur ci Direction ak Representation

Saytu naka lañuy xaymaa vecteur yiy doxal, jëfandikoo ci inference ak scale, ànd ci risk yi ci generalisation bu aju ci gaaw.

Yooni njiit yu jëm ciIngénieur de direction ak representation
Laaj 1 de 8

Ban poñ ci liggéeyu model bi la activation steering bi di duggu?

Xoolaatal laaj bu nekk8 laaj

Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.

  1. Ban poñ ci liggéeyu model bi la activation steering bi di duggu?
    • Ci jamonoy inference, ci soppali aktivaasioŋ yu nëbbu yi
    • ci diiru tàggat yaram kese
    • Bu ñuy jël ay done
    • Ginaaw ñu imprime mbind mi
  2. naka lañuy faral di xaymaa vecteur de direction?
    • Ni ñuy tàmbali ab matrix ci anam wu mucc
    • Differance moyenne bi am ci digganté aktivisasioŋ yi ci wuutale ay laaj yu baax ak yu baaxul
    • Sooy efaase ay couche ci model bi
    • Ni ñuy lim ay token ci laaj bi
  3. Ban hypothèse moo waral steering activation di dox?
    • Xalaat biy wax ni model yi dañuy denc done yi ci anam wu leer
    • Gëm ni neuron yépp ñooy benn
    • Hypothèse de représentation ligneaire, maanaam màndarga yu bari dañuy nuru yoon yu ligneer
    • Wax jiy wax ni model yi mënuñoo representé konsept yi
  4. Lan mooy koeffisientu escalier bi ci vecteur de direction bi di doxal?
    • Dayo vocabulaire model bi
    • Limu diisaay yi ci model bi
    • Tolluwaayu jàng ci tàggat yaram
    • Ak naka lañuy puus jeffin ji ci yoon wiñ tànn
  5. Lan mooy risk biñ xamee ci dawal model bi ak doole ju ëpp?
    • Koherensi ak kàttan yi ci génne mën nañu yàqu
    • Modèle bi dafay tàggat boppam
    • Dayo fichier bi dafay màgg bu baax
    • Dafay binndaat poid yi ba fàww
  6. Ban pexe lañuy faral di boole ak direction ngir am man-mani yuñ mëna tekki ngir pince?
    • kompresioŋ JPEG
    • Autoencoder yu néew yuy xaaj aktivasioŋ yi ci man-man
    • Bindu tur domen
    • Defragmentaasioŋ disk
  7. How does activation steering differ from fine-tuning the model weights?
    • It retrains the model on a new labeled dataset
    • It changes the tokenizer vocabulary
    • It adds a vector to hidden activations during inference without updating the base weights
    • It replaces the attention layers with a classifier
  8. A steering vector works on the prompts used to compute it but not on a different topic. Which limitation does this illustrate?
    • The vector must be an optimizer state
    • Steering always requires fine-tuning
    • A residual stream contains no features
    • The direction found from one prompt set may not generalize to other contexts