Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele

Iyatọ Iyipada adanwo

Tests expert–amateur scoring, shared-token suppression, plausibility filtering, decoding-time use, and evaluation of contrastive generation.

Jẹmọ itọsọna awọn ọnaIyipada itansan
Ibeere 1 ti 8

Ni iyipada iyatọ, ipa wo ni awoṣe 'magbowo' kekere ṣe?

Awotẹlẹ gbogbo ibeere8 ibeere

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

  1. Ni iyipada iyatọ, ipa wo ni awoṣe 'magbowo' kekere ṣe?
    • Awọn iṣeeṣe log rẹ ti yọkuro lati dinku awọn ipo ikuna pinpin
    • O ṣe agbekalẹ ọrọ igbejade ikẹhin taara
    • O ti wa ni itanran-aifwy lori awọn aṣiṣe iwé
    • O rọpo alamọja nigbati alamọja ko ni idaniloju
  2. Kini idi ti iyokuro awọn iṣeeṣe magbowo naa dinku atunwi ati iṣẹjade alaiwu?
    • Awọn magbowo ti wa ni oṣiṣẹ ko lati tun ọrọ
    • Jeneriki ati awọn ami atunwi jẹ ojurere nipasẹ awọn awoṣe mejeeji, nitorinaa itansan fagile wọn
    • Ope ma npa awọn ami ẹda ẹda lẹhin iran
    • Atunwi ko ṣee ṣe ni kete ti awọn awoṣe meji dibo
  3. What does the adaptive plausibility constraint do in contrastive decoding?
    • It guarantees every remaining continuation is coherent and true
    • It replaces the expert distribution with the amateur distribution
    • It filters out candidates whose expert probability is too low relative to the expert’s top token
    • It forces every token to have equal probability
  4. Ni aijọju kini agbekalẹ igbelewọn ni iyipada iyatọ?
    • log p_amateur iyokuro log p_expert
    • p_expert igba p_amateur
    • Apapọ awọn iṣeeṣe meji
    • log p_expert iyokuro a olùsọdipúpọ igba log p_amateur
  5. Anfani iwulo bọtini kan ti iyipada iyatọ ni pe:
    • Ṣiṣẹ ni akoko itọkasi laisi ikẹkọ afikun
    • Nilo atunṣeto awọn awoṣe mejeeji ni apapọ
    • Nikan ṣiṣẹ lori awọn iṣẹ-ṣiṣe iyasọtọ
    • Imukuro iwulo fun awoṣe nla kan
  6. Eyi ti nigbamii ọna fa awọn itansan agutan to kan nikan awoṣe ile ti ara fẹlẹfẹlẹ?
    • Iwadi ina
    • DoLa, eyiti o ṣe iyatọ ni kutukutu dipo awọn fẹlẹfẹlẹ pẹ
    • Iyipada ojukokoro
    • Iwọn iwọn otutu
  7. The expert assigns probability 0.4 to its top token and the plausibility threshold is 0.1 of that value. What is the minimum probability for a candidate to remain eligible?
    • 0.4
    • 0.1
    • 0.04
    • 0.004
  8. Why can contrastive decoding require more inference work than using the expert model alone?
    • It must retrain both models for every prompt
    • It replaces the vocabulary with image patches
    • It runs only the amateur model
    • It scores candidates with both an expert and an amateur model at inference time