বহুভাষিক এলএলএম কুইজ

আপনি কুইজ নেওয়ার সাথে সাথে উত্তর এবং ব্যাখ্যাগুলি উপস্থিত হয়।

1. A model’s interface lists ten languages. What does that alone establish about answer quality?
   • Equal accuracy in all ten
   • Equal token counts in every script
   • Only that the interface offers them, not equal task performance
   • Identical cultural context for every user
2. Why can two translations of similar meaning consume different context capacity?
   • Tokenization can assign different token counts across languages
   • All scripts use one character per token
   • Meaning fixes token count exactly
   • The model ignores non-English text
3. What does cross-lingual transfer not guarantee?
   • That shared patterns can help a second language
   • That a multilingual model learns representations
   • That transfer is measurable
   • Correct handling of local idioms and specialized terminology
4. Why use natural local examples rather than only machine-translated English benchmark items?
   • Local examples make every model perfect
   • Translated prompts can be unnatural or erroneous for the target language
   • English benchmarks contain no questions
   • Translation always fixes cultural mismatch
5. A support request mixes languages and a product code. What should evaluation check?
   • Whether the answer preserves names and codes while handling both languages
   • Only the length of the English portion
   • Only whether the model emits one language label
   • Only the number of fonts installed
6. Which user-experience details need review beyond the model’s words?
   • Only total model parameters
   • Only the English prompt template
   • Fonts, text direction, locale formats, citations and fallback behavior
   • Only server memory use
7. In the guide’s invented example, why is 86% combined accuracy misleading?
   • It proves both languages reached 86%
   • It hides 50% accuracy on the smaller language-B group
   • It uses no reference answers
   • It means every B answer was wrong
8. A requested-language task is unreliable. What should the product do?
   • Silently switch to an unsupported answer
   • Report only a global average
   • Hide the source so errors are less visible
   • Communicate the limit and preserve access to the source