Enable JavaScript to grade the quiz and save your result. You can still read the explanations below.
Why does word count fail to determine context size?
Expected answer. The tokenizer can split different words into different numbers of tokens
Units depend on the tokenizer’s vocabulary and rules. Spaces and punctuation also count.
Review section 3.1In a causal grid, which keys can a query at position 3 read?
Expected answer. Positions 1, 2 and 3
The mask allows the diagonal and earlier positions. Future positions remain masked during training.
Review section 3.2Why keep tokenization identical when comparing perplexities?
Expected answer. Because loss is averaged per tokenized unit
The denominator is token count. A different segmentation changes what that average represents.
Review section 3.3With probabilities 0.60, 0.25, 0.10 and 0.05 and top-p = 0.80, which options should be kept?
Expected answer. The leading two, then renormalize
The first prefix to reach 0.80 has mass 0.85. Renormalization brings the sum back to one.
Review section 3.4Which test is useful for multimodal extraction of a scanned table?
Expected answer. Verify the value, header, row and page
A correct value in the wrong cell changes its meaning. The evaluation needs to preserve structure.
Review section 3.5