Kediska Isku-darka Wakiilada

Jawaabaha iyo sharraxaadaha waxay soo baxayaan markaad qaadanayso kediska.

1. Isku-dhafka wakiilada, waa maxay doorka tusaalaha ururiyaha?
   • Si loo habeeyo jawaabaha moodooyinka soo jeedinta lagana dhigo hal jawaab oo la hagaajiyay
   • Si loo tirtiro dhammaan marka laga reebo hal qabyo si aan kala sooc lahayn
   • Si degdega loogu turjumo kood
   • Si loo tababaro moodooyinka soo jeedinta
2. Waa maxay hantida muhiimka ah ee LLMs ay warqadda MOA ku tiirsan tahay?
   • In LLMs aysan akhrin karin wax soo saarka moodooyinka kale
   • Taasi LLM-yadu waxay soo saaraan jawaabo ka fiican marka la muujiyo jawaabaha asxaabta ('wada-shaqeynta')
   • In LLMs ay mar walba ogolaadaan
   • Moodooyinka waaweyni waa kuwo aan loo baahnayn
3. Sidee ayay MOA kaga duwan tahay codaynta aqlabiyadda fudud?
   • Waxay ku soo qaadanaysaa jawaabta ugu badan oo hadal ah
   • Waxay isticmaashaa hal nooc oo keliya
   • Isku-dariyuhu wuxuu akhriyaa dhammaan qoraallada sida macnaha guud wuxuuna abuuraa iskudar cusub
   • Waxay iska indha tirtay soo-jeediyeyaasha gabi ahaanba
4. Waa maxay sababta kala duwanaanshiyaha moodooyinka soo jeedinta ay qiimo ugu leeyihiin MOA?
   • Waxay dammaanad qaadaysaa wax soo saar isku mid ah
   • Waxay hoos u dhigtaa tirada lakabyada eber
   • Waxay ka dhigaysaa dhammaan jawaabaha khaldan
   • Qoysaska moodada ah ee kala duwan waxay ku biiriyaan awoodo is kaaba ah oo isku-darku isku dari karo
5. Waa maxay kharashka ugu muhiimsan ee habka MOA?
   • Latency sare iyo xisaabinta, mar haddii lakab kasta ku dhufto wicitaanada model
   • Uma baahnid wax ka-fiirsi haba yaraatee
   • Waxa ay soo saari kartaa hal kelmad
   • Waxay tirtirtaa soo jeedintii asalka ahayd
6. Muxuu warqadda MoA ka sheegay moodooyinka il-furan ee AlpacaEval 2.0?
   • Waa ay wadi kari waayeen bar-tilmaameedka
   • MoA ee moodooyinka furan ayaa dhaafay dhibcaha GPT-4 Omni
   • Moodooyinka furan had iyo jeer waxay dhaliyeen eber
   • Kaliya moodooyinka xiran ayaa la tijaabiyay
7. What information is given to an agent in a later Mixture-of-Agents layer?
   • Only the original prompt with all earlier drafts removed
   • Only a majority-vote label
   • The outputs produced by agents in the preceding layer
   • A gradient update from each previous model
8. What is the purpose of placing several proposer models in the first MoA layer?
   • Train a shared base model from scratch
   • Choose one response by majority vote and discard the rest
   • Compress all outputs into a single token
   • Create a set of candidate responses for later agents and the aggregator to use