Falcon Models Quiz

Answers and explanations appear as you take the quiz.

1. Which organization develops the Falcon models?
   • The Technology Innovation Institute (TII) in Abu Dhabi
   • A Silicon Valley startup
   • The Chinese Academy of Sciences
   • A European university consortium
2. What was notable about the RefinedWeb dataset used to train early Falcon models?
   • It was tiny and hand-written
   • It was built by aggressively filtering and deduplicating web data instead of relying on curated sources
   • It contained only images
   • It was entirely synthetic
3. What attention technique do Falcon transformer models use to save memory at inference?
   • No attention at all
   • Full quadratic attention only
   • Multi-query attention
   • Random attention
4. How does Falcon Mamba differ architecturally from standard Falcon models?
   • It is a transformer with more layers
   • It has no neural network
   • It only does image classification
   • It uses a state-space model instead of a transformer
5. Under what kind of terms are Falcon models generally released?
   • Permissive terms encouraging commercial and research use
   • Strictly closed and private
   • Available only to UAE citizens
   • Only as a paid API
6. Falcon is often cited as an example of which broader trend?
   • The end of open-source AI
   • Countries building sovereign, homegrown foundation models
   • Models that only work offline
   • Replacing all GPUs with CPUs
7. In multi-query attention, what do multiple query heads share?
   • One shared query and separate keys for every token
   • The same final vocabulary logits
   • A single set of key and value projections
   • A single attention score for all layers
8. What does Falcon’s RefinedWeb training data illustrate?
   • A state-space model needs no training text
   • All web pages are equally reliable without filtering
   • Synthetic data must always replace real text
   • A large web corpus can be filtered and deduplicated rather than relying only on curated sources