quiz bu lëkkale ak tegtal · Diggante Niveau

DPO buy baamtu ak laaj ak tontu ci tànneef ci net bi

Jàngaleel ni DPO iteratif di yeesal done yiñ taamu, jëfandikoo àttekat yi, ak jënd ak jaay ay rond yu bari ci wàllu njëg ak njubadi.

Yooni njiit yu jëm ciDpo iteratif ak tuning tànneef ci net bi
Laaj 1 de 8

Lan la DPO moytu lu RLHF cosaan (PPO) sàkku?

Xoolaatal laaj bu nekk8 laaj

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

  1. Lan la DPO moytu lu RLHF cosaan (PPO) sàkku?
    • Taggat ab xeetu neexal bu leer
    • Bépp done tàggat yaram
    • Jëfandikoo reso neuronal
    • Taamu nit
  2. Lan moo waral DPO iteratif (ci net bi) di gëna baax def DPO benn yoon ci kaw benn done bu takku?
    • Soxlawul etiketu tànneef
    • Dafay wéy di tàggat ay done ci wàllu politik ak yu bees ginaaw bi xeetu coppite yi di soppeeku
    • Dafay dindi xeetu royuwaay bi
    • Nataal kese lay dox
  3. Ban cër la modelu royuwaay bi di def ci ñàkkum DPO?
    • Dafay defar jëfandikukat bu mujj bi
    • Dafay màndargaal tànneef yi
    • Dafay tënk politik bi jaaraleko ci ratioo log-probabilite, di saytu jaar-jaar bi
    • Dafay wecci tokeniseer bi
  4. Ci benn iteration ci DPO ci net bi, lan mooy xeetu toppalante bi?
    • efaase model bi, nga tàggataat ci noonu
    • Etiketuwaat done yu yàgg yi te doo jël misaal
    • Taggat CNN ci tontu yi
    • Misaalu tontu yu bees, àtte tànneef yi, ba noppi def yeesal DPO
  5. Lan la hiperparametre beta bi ci DPO di saytu?
    • Ak naka la politik bi di wuutee ak xeetu royuwaay bi
    • Dayo vokabileer bi
    • limu GPU yi
    • Dayo nataal bi
  6. Lan mooy xalaat bu am solo bi ci ginaaw DPO iteratif 'di neexal boppam'?
    • Royuwaay bi musul jëfandikoo tànneef
    • Royuwaay bi dafay defar ba noppi àtte tontu boppam ngir sos done tàggat
    • Nit ñi dañu wara etikete peer bu nekk
    • Dafay jëfandikoo ay neexal yu bari
  7. Who can label response preferences in an online DPO iteration?
    • Only the tokenizer
    • Only the reference policy’s embedding layer
    • Human raters or an AI judge such as a reward model
    • A discriminator that never sees responses
  8. What additional work does iterative DPO require compared with using a fixed preference dataset once?
    • Training no model parameters
    • Removing all evaluation from the pipeline
    • Storing every possible response in advance
    • Repeated response generation and preference labeling between updates