Kediska Jadwalka diirimaadka iyo Cosine Annealing

Jawaabaha iyo sharraxaadaha waxay soo baxayaan markaad qaadanayso kediska.

1. Waa maxay ujeedada ugu weyn ee wejiga kululaynta ee bilowga tababarka?
   • Si loo dejiyo tababarka hore marka miisaanku yahay random iyo gradients ay buuq badan yihiin
   • Si loo kordhiyo xajmiga dufcada si tartiib tartiib ah
   • Si aad u qaboojiso lakabyada hoose ee moodeelka
   • Si aad u cadaadiso moodalka si aad u hesho fikrad degdeg ah
2. Annealing cosine waxa ay kharribasaa heerka waxbarashada ee raacaya qaabkee?
   • Xariiq toosan oo toosan
   • Qalooca cosine ee ku bilowda sare, fududeeya, oo ku fidsan meel eber u dhow
   • Shaqo tillaabo ah oo hoos u dhacda xilliyo go'an
   • Kor u kaca jibbaarada
3. Kee baa ka faa'iidaysanaya kululaynta sababtoo ah qiyaasaha kala duwanaanshaheeda ayaa ah kuwa aan la isku halayn karin goor hore?
   • SGD cad oo aan dardar lahayn
   • Raadinta qulqulka
   • Aadam iyo noocyadiisa
   • K-macneheedu waa
4. Qaabka cosine-ka, muxuu T ka dhigan yahay?
   • Heerkulka moodeelka ee shaybaarka
   • Tirada GPU-yada la isticmaalay
   • Cabbirka ereyada tokenizer
   • Tirada guud ee tillaabooyinka tababarka ee jadwalka
5. In the guide’s warmup-plus-cosine schedule, what begins after warmup reaches its target rate?
   • Cosine decay begins from the peak learning rate
   • All model weights are reset
   • The schedule switches to a constant zero rate
   • Momentum is removed from every optimizer
6. A linear warmup rises from zero to its target rate over 200 steps. About what rate is used halfway through?
   • The full target rate
   • About half of the target rate
   • The learning rate is exactly zero
   • Twice the target rate
7. At the cosine horizon t = T, what multiplier does the guide’s cosine formula produce?
   • Mid
   • One-half
   • Eber
   • It becomes negative one
8. Why use a gradual cosine decay after warmup rather than abruptly setting the rate to zero?
   • It guarantees a global minimum
   • It increases the model’s parameter count
   • It removes the need to check training loss
   • It lets the rate ease down so optimization can settle gradually