Byte-Pair Encoding Quiz

Answers and explanations appear as you take the quiz.

1. Which adjacent-symbol operation is repeated in common BPE vocabulary training?
   • Deleting all low-frequency words from the language
   • Merging a most-frequent pair into a larger symbol
   • Sorting complete documents by title
   • Selecting pairs solely by their alphabetical spelling
2. What does BPE treat the text as when it begins training?
   • Whole sentences
   • Fixed 8-token blocks
   • Individual characters or bytes
   • Pre-translated phrases
3. In a byte-level BPE configuration covering all 256 base byte values, how can unfamiliar valid UTF-8 text be represented?
   • Through its constituent bytes when larger pieces are unavailable
   • By storing every possible complete word in the finite vocabulary
   • By requiring the input to be translated into English first
   • By assigning every unseen character one shared unknown ID
4. Where did the BPE algorithm originally come from before NLP adopted it?
   • Image compression
   • Audio encoding
   • Database indexing
   • Data compression
5. When GPT-2's published BPE encoder finds multiple eligible adjacent pairs, what determines its merge choice?
   • A new random vocabulary fit on that one input
   • The text file's creation date
   • The priorities in the learned merge ranks
   • The order of a human-written list of target meanings
6. What can happen when an unfamiliar word lacks larger learned BPE pieces?
   • It must be replaced by an unrelated synonym
   • It remains represented by several smaller available pieces
   • It necessarily becomes one new learned token during inference
   • It is always silently removed from the input
7. An illustrative BPE vocabulary has 256 base byte tokens, two distinct added merge tokens, and one separate special token. What is its total size?
   • 256
   • 258
   • 512
   • 259
8. In GPT-2's published decoder, what is reconstructed before converting the result from UTF-8 into text?
   • The complete byte sequence represented by the token pieces
   • An independent Unicode character from every individual token first
   • A new corpus-frequency table for training merges
   • The model's output-label probabilities