Skip to content
Home: How language models work

03Book 04 · Chapter 03 of 07· 5 min ·

Show it the answer you want

A is a finished reply a person wrote as the target. For example, the request is “What is the capital of Australia?” and the written target is “Canberra.” Training then practices next-word prediction on that reply, by , so later outputs copy its wording and structure. Name Canberra, and later replies lean Canberra. Name Sydney in polished prose, and later replies lean that instead.

The next stage shows the model pairs: a request, then a reply someone wrote as the target. The model practices producing replies that look like those targets. That is a — on instructions, if you like the long name.

The tiny demo below is the same idea at toy scale. One extra sentence moves “sofa” up the list. A thousand real move tone, structure, and whether the answer comes first.

The target is the text as written. A gorgeous paragraph that says Sydney is the capital teaches gorgeous-and-wrong. The wheel does not fact-check the example. It copies it.

A tiny model has only practiced “The cat sat on the mat.” Add one example and watch the next-word list move — sofa rises, mat stays. One is not a factory reset.

The cat sat on the....

  • mat58%
  • floor22%
  • hat12%
  • sofa8%

Before the new example, mat is the usual next word. That is what the early practice taught.

Quiz

6 questions

Before moving forward, test your knowledge of what you’ve just learned!

  1. 1.A demonstration is:

  2. 2.If examples bury the answer:

  3. 3.The tiny “sofa” demo shows:

  4. 4.SFT on instructions means:

  5. 5.A beautiful wrong city in the target:

  6. 6.Demonstrations work because:

0 of 6 answered

3 / 7