quiz bu lëkkale ak tegtal · Dëgër Niveau

Groupe Relatif Politik Optimisation Quiz

Jëfandikool GRPO grupu neexal normalisasioŋ, dindi ŋàññikat, ak royuwaay-politigu yamale, boole ci li njariñ relatif yi tekki ci benn laaj.

Yooni njiit yu jëm ciGroupe bi gëna xéewale ci wàllu politik
Laaj 1 de 8

Which separately learned estimator does the original GRPO approach avoid training?

Xoolaatal laaj bu nekk8 laaj

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

  1. Which separately learned estimator does the original GRPO approach avoid training?
    • The policy that generates completions
    • A value or critic model
    • Every possible reward function
    • The reference policy’s vocabulary
  2. naka la GRPO di xaymaa njariñu benn matal buñ jox?
    • Sooy laaj benn boppu valeur buñu jàng ci token bu nekk
    • Jëfandikoo neexal bu ñor te amul benn normalisasioŋ
    • Soo koy méngale ak li nga def ci beneen laaj
    • Ni ñuy méngale neexal bi ak moyenne bi ak standard deviation bi ci grupu mbokkam yi
  3. Ban xeetu xam-xam moo tax GRPO siiw?
    • DeepSeekMath ak DeepSeek-R1 xeetu xalaat
    • OpenAI's GPT-2 series
    • Google's original BERT
    • __AAR_13__'s Llama 2 models de base
  4. Ban cër la terme KL-divergence def ci GRPO?
    • Dafay defar mbooloo mu mat
    • Dafay wecci xeetu neexal
    • Dafay tere politik biñ tàggat dem fu sori lool ci xeetu royuwaay bu liw
    • Dafay xayma valeur token bu nekk
  5. Lan moo waral GRPO di gëna am solo ngir tàggat ay xeetu xalaat ci math wala kode?
    • Soxlawul benn siñaal neexal
    • Du dox ludul ci etiketu tànneefi nit
    • Mënul jëfandikoo misaali matt
    • Dafay dox ci anam wu natureel ak ay neexal yuñ mëna saytu, yu lalu ci sàrt (lu melni, ndax tontu bi méngoo na wala ndax test yi jàll nañu) ñu jox ko poñ ci kuréel guñ jël misaal
  6. Lan mooy benn ci njuumte yiñ xam ni liggéey bi ñuy topp (lu melni Dr. GRPO) dafay jéema saafara ci vanille GRPO?
    • Tendens ngir baña bàyyi xel ci neexal bi
    • Guddaay ak jafe-jafe bi ñu dugal ci ni ñuy normalisee neexal yi
    • Ñàkka mëna sample lu ëpp benn matt
    • Li ñuy laaj ci reso kritik bu mag
  7. Three GRPO completions for one prompt receive rewards 2, 4, and 6. Which completion has zero centered advantage?
    • The completion rewarded 2
    • The completion rewarded 6
    • The completion rewarded 4
    • All completions because rewards differ
  8. Why compare GRPO reward statistics among completions of the same prompt?
    • They provide a relative baseline for alternative answers to that prompt
    • They require unrelated prompts to have equal difficulty
    • They replace checking whether an answer is correct
    • They make the highest reward identical across all tasks