course/pr_1213/en/chapter7/9.md

End-of-chapter quiz[[end-of-chapter-quiz]]

Let's test what you learned in this chapter!

1. Which of the following tasks can be framed as a token classification problem?

2. What part of the preprocessing for token classification differs from the other preprocessing pipelines?

-100 to label the special tokens.", explain: "That's not specific to token classification -- we always use -100 as the label for tokens we want to ignore in the loss." }, { text: "We need to make sure to truncate or pad the labels to the same size as the inputs, when applying truncation/padding.", explain: "Indeed! That's not the only difference, though.", correct: true } ]} />

3. What problem arises when we tokenize the words in a token classification problem and want to label the tokens?

-100 so they are ignored in the loss." }, { text: "Each word can produce several tokens, so we end up with more tokens than we have labels.", explain: "That is the main problem, and we need to align the original labels with the tokens.", correct: true }, { text: "The added tokens have no labels, so there is no problem.", explain: "That's incorrect; we need as many labels as we have tokens or our models will error out." } ]} />

4. What does "domain adaptation" mean?

5. What are the labels in a masked language modeling problem?

6. Which of these tasks can be seen as a sequence-to-sequence problem?

7. What is the proper way to preprocess the data for a sequence-to-sequence problem?

inputs=... and targets=....", explain: "This might be an API we add in the future, but that's not possible right now." }, { text: "The inputs and the targets both have to be preprocessed, in two separate calls to the tokenizer.", explain: "That is true, but incomplete. There is something you need to do to make sure the tokenizer processes both properly." }, { text: "As usual, we just have to tokenize the inputs.", explain: "Not in a sequence classification problem; the targets are also texts we need to convert into numbers!" }, { text: "The inputs have to be sent to the tokenizer, and the targets too, but under a special context manager.", explain: "That's correct, the tokenizer needs to be put into target mode by that context manager.", correct: true } ]} />

{#if fw === 'pt'}

8. Why is there a specific subclass of Trainer for sequence-to-sequence problems?

-100", explain: "That's not a custom loss at all, but the way the loss is always computed." }, { text: "Because sequence-to-sequence problems require a special evaluation loop", explain: "That's correct. Sequence-to-sequence models' predictions are often run using the generate() method.", correct: true }, { text: "Because the targets are texts in sequence-to-sequence problems", explain: "The Trainer doesn't really care about that since they have been preprocessed before." }, { text: "Because we use two models in sequence-to-sequence problems", explain: "We do use two models in a way, an encoder and a decoder, but they are grouped together in one model." } ]} />

{:else}

9. Why is it often unnecessary to specify a loss when calling compile() on a Transformer model?

{/if}

10. When should you pretrain a new model?

11. Why is it easy to pretrain a language model on lots and lots of texts?

12. What are the main challenges when preprocessing data for a question answering task?

13. How is post-processing usually done in question answering?