quiz bu lëkkale ak tegtal · Dëgër Niveau

Sampling biñ nara def ak Exposition Bias Quiz

Saytu ni échantillonnage biñ waajal di soppi contexte decodeur bi, lu tax probabilité token wurus bi di yàqu, ak njaay gradient yi ci echantillonnage bi.

Yooni njiit yu jëm ciÉchantillonnage buñ waajal ak exposition bias
Laaj 1 de 8

Exposition bias mingi am ndax model yiñ tàggat ak jàngalekat yi dañuy forse rek ci:

Xoolaatal laaj bu nekk8 laaj

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

  1. Exposition bias mingi am ndax model yiñ tàggat ak jàngalekat yi dañuy forse rek ci:
    • Ground-truth tokens, waaye dañu wara jëfandikoo seen bopp ci inference
    • Seen jeton yuñ séentu
    • Jetons bruit yu bari
    • Jetons yu ëlëg
  2. Lan la sampling biñ nara def ci diiru tàggat yaram?
    • Saa yu nekk dafay dundal token dëgg
    • Tann ci anam wu mucc diggante jeton dëgg ak xalaatu model bi boppam
    • Desactivé fonction perte saa yu nekk
    • Yokk ñaari yoon limu jàng ci diir bi
  3. Ak naka la probabilite jëfandikoo token ground-truth di soppikoo ci tàggat ci échantillonnage biñ nara def?
    • Dafay wéy di nekk
    • Dafay tàmbali tuuti ba noppi yokk
    • Dafay tàmbali yéeg ba noppi wàcci dem ba amul dara
    • Dafay yëngu ci anam wu mucc
  4. Kan moo dugal sampling biñ nara def, ak kañ lañu koy def?
    • Vaswani ak ñeneen, 2017
    • Hinton ak ñeneen, 2006
    • LeCun ak ñeneen, 1998
    • Bengio ak ñeneen, 2015
  5. Benn ci jafe-jafe yi ñu xamee ci teknik bi mooy joxewaat ay token yuñ jël ci diiru tàggat-yaram mooy:
    • Jéego jël misaal bu wuute bi mënul wuutale, dafay jafeel gradient yi
    • Dafay tax tàggat yaram nekk deterministik bu mat sëkk
    • Soxlawul benn done buy saytu
    • Dafay dindi soxla dugal ay mbir
  6. Ban jëmmu oraaire lañuy faral di wax ngir yàq probabilite forse jàngalekat bi?
    • Konstant
    • Decay sigmoïde inverse
    • Yokkateg ligneer
    • Dox ci anam wu mujjee
  7. At a decoding step where scheduled sampling chooses the model-output branch, what is fed as the next context token?
    • The next ground-truth token every time
    • A token from a separate validation sequence
    • A token sampled from the model’s predicted distribution
    • The average of all vocabulary embeddings
  8. Why does scheduled sampling usually begin with a high probability of using ground truth?
    • It guarantees unbiased gradients
    • It removes the need for a training schedule
    • It makes inference use ground-truth tokens
    • Early predictions are poor, so a correct prefix can stabilize initial learning