T5 and Text-to-Text Transfer Quiz

Answers and explanations appear as you take the quiz.

1. What is the defining idea behind T5?
   • It only does machine translation
   • It generates images from captions
   • Every NLP task is framed as text input to text output
   • It replaces all training with reinforcement learning
2. What Transformer architecture does T5 use?
   • A full encoder-decoder
   • Encoder-only, like BERT
   • Decoder-only, like GPT
   • No attention at all
3. What pretraining objective does T5 primarily use?
   • Next-sentence prediction only
   • Predicting stock prices
   • Image inpainting
   • Span corruption, masking contiguous spans replaced by sentinel tokens
4. What corpus was T5 pretrained on?
   • Only Wikipedia
   • The C4 Colossal Clean Crawled Corpus
   • A dataset of source code
   • Audio transcripts
5. How does T5 represent a classification task like sentiment analysis?
   • It outputs a probability vector only
   • It returns a numeric ID
   • It generates the literal label word, such as 'positive' or 'negative'
   • It cannot do classification
6. What is FLAN-T5 relative to the original T5?
   • A smaller image model
   • An instruction-tuned descendant that generalizes to unseen tasks
   • A speech recognition system
   • An unrelated database tool
7. When T5 generates a summary, what does decoder cross-attention let it consult?
   • The encoder's representations of the input article
   • Only previously generated summary tokens
   • Only the unprocessed task label
   • The training optimizer's stored gradients
8. In a closed-book T5 question-answering setup, where does the answer's factual information come from at inference?
   • A document retrieval step required for every question
   • A newly supplied answer passage required in the prompt
   • A live database lookup built into span corruption
   • Knowledge encoded in the model's learned parameters