En anglais

One piece of text after another

Computing · Artificial Intelligence · ages 12-13

Name ______________________ Date ____________

How does a language model acquire its language patterns?
- A person writes a separate answer rule for every prompt
- It learns patterns during training on text
- It receives every possible reply from a hidden live writer
Which process describes a language model generating a reply?
- Select a token, add it, and repeat
- Copy an entire stored answer without generating text
- Wait for a person to type every token of the reply
After generating a token, a chatbot leaves it out of the context for choosing the next token.
Circle one: True False
Two runs use the same prompt and context with random token selection turned on. Why might their wording differ?
- Random selection can choose different tokens
- A changed reply always means all training was erased
- The second run must delete the words in the first reply
In a text-generation example, a chatbot continues 'The cat sat on the' with 'mat'. What does this show?
- It proved every sentence with that start must end in mat
- It measured a real cat and identified the surface below it
- It selected a possible next word while generating text
A chatbot confidently gives a date for your history essay. What should you do before using it?
- Treat confident wording as proof that the date is right
- Check the date in a reliable source
- Ask for more confident wording instead of checking it
You use a chatbot to draft wording and suggest a historical fact. How should you review the two outputs?
- Review the wording's meaning and check the fact in a reliable source
- Accept both without review whenever the sentences sound fluent and confident
- Check only spelling and treat every suggested historical fact as established
A friend says every chatbot reply is copied from a fact book, so every reply must be right. Which response fits?
- All replies are checked facts because training removes every possible error
- Every reply is handwritten by a person who checks it before it appears
- Replies can be generated; claims and retrieved sources need checking

Answer key

For grown-ups. Fold this page away before handing over the rest.

One piece of text after another W1-mt_1Zbhamhzai-s1

It learns patterns during training on text · Training on text fits language patterns. People still write the learning software.
Select a token, add it, and repeat · The model generates a sequence by selecting and adding tokens. A token may be a word, a word part or punctuation.
False · The generated token is added to the text available for the next selection.
Random selection can choose different tokens · Random selection can choose different tokens. Those tokens also change the context for later selections.
It selected a possible next word while generating text · Mat is a possible continuation. The output does not prove a unique ending or any fact about a real cat.
Check the date in a reliable source · Fluent, confident wording does not establish a date. Check a reliable source before using it.
Review the wording's meaning and check the fact in a reliable source · Review whether the wording expresses your meaning. Check the historical claim against a reliable source.
Replies can be generated; claims and retrieved sources need checking · A chatbot can generate text rather than copy a checked answer. Even when it retrieves a source, you need to check that source and its support for the claim.