Kuramba Sampling Fine-Tuning Quiz

Mhinduro uye tsananguro zvinoonekwa paunenge uchitora bvunzo.

1. Which loop best describes how Rejection Sampling Fine-Tuning improves a model?
   • Gadzira mhinduro dzakawanda dzemumiriri, chengeta dzakanyanya-zvibodzwa, uye dzokorora pazviri
   • Dzima yakaipisisa-inoita modhi layers
   • Ramba data rese rekudzidziswa uye tanga patsva
   • Sample kamwe uye gara uchigamuchira zvakabuda
2. Munzvimbo dzinotariswa semasvomhu kana kodhi, chii chinogona kushanda semubairo?
   • A akadzidza mubairo modhi chete
   • Kuti mhinduro ndeyechokwadi here kana kuti bvunzo dzakapasa
   • Hurefu hwemhinduro
   • Nhamba yezviratidzo zvakagadzirwa
3. Which model family does the cited paper describe using rejection sampling in its alignment pipeline?
   • GPT-4o
   • BERT Base
   • Meta’s Llama 2-Chat
   • Whisper Large
4. Why might a team choose rejection-sampling fine-tuning over a policy-gradient approach for some tasks?
   • It never needs a candidate generator or scoring rule
   • It guarantees no reward hacking
   • It always requires less compute than every RL method
   • It can use supervised learning on selected generations instead of updating through policy-gradient RL
5. Kudzidzira pane yakanyanya-mubayiro samples zvinobudirira kunoita chii?
   • Distills yakanakisa-ye-N mhando mune imwe modhi-yemuenzaniso zvinobuda
   • Inowedzera saizi yemazwi
   • Inobvisa kugona kwemuenzaniso kugadzira zvinyorwa
   • Inopeta kaviri parameter kuverenga
6. What should teams monitor when repeatedly fine-tuning on filtered model-generated outputs?
   • Only the number of model parameters
   • Output diversity alongside held-out correctness
   • Whether every answer uses the same wording
   • Only the training loss on selected samples
7. What is a limitation of rejection-sampling fine-tuning’s candidate selection?
   • It can only be used with image labels
   • It always changes the tokenizer
   • It cannot train on a high-quality answer that was never generated or retained
   • It makes the scoring function unnecessary
8. What risk arises if the scoring model used to filter candidates is flawed?
   • The policy becomes unable to generate any text
   • The tokenizer automatically changes vocabulary
   • The reference labels become unnecessary
   • The policy may learn to exploit the scorer rather than improve the intended quality