Tokenization Disparities Across Languages Quiz

Answers and explanations appear as you take the quiz.

1. What is the main cause of large token-count gaps between languages?
   • Some languages are naturally ten times wordier than English
   • Tokenizers were built from training text dominated by English and a few other languages
   • Non-English text is compressed before tokenizing
   • Models charge extra fees for translation
2. Why do scripts like Devanagari or Thai often produce more tokens under byte-based tokenizers?
   • Their characters usually take three bytes each in UTF-8 and may be rarely merged
   • They are always written without spaces
   • They are converted to images first
   • Tokenizers reject them and replace them with English
3. Which of these is NOT one of the three effects of higher token counts named in the guide?
   • Higher cost per message
   • Less text fitting in the context window
   • Slower generation from more steps
   • Automatic refusal to answer
4. What does the term fertility mean in this context?
   • How quickly a model learns a new language
   • The average number of tokens produced per word
   • The number of languages a model supports
   • The size of the training dataset
5. What change did OpenAI describe when it released GPT-4o in 2024?
   • It stopped charging for non-English text
   • It removed its tokenizer entirely
   • It introduced a larger tokenizer that used fewer tokens for many non-English languages
   • It limited GPT-4o to English only
6. What is the best first step to plan costs for a multilingual app?
   • Assume every language costs the same as English
   • Tokenize representative translated documents with your model's exact tokenizer and compare counts
   • Pick the cheapest model regardless of language
   • Translate all user messages into English first without testing
7. What tradeoff comes with making a tokenizer's vocabulary larger?
   • It makes English text impossible to process
   • It increases the size of the embedding and output layers
   • It removes the need for training data
   • It always lowers model quality
8. According to the guide, why is 'some languages are just wordier' a misconception?
   • All translations are exactly the same length
   • Wordiness has no effect on anything
   • Large token gaps come mainly from tokenizer design, not from the languages themselves
   • Tokenizers only count spaces