Log inSign up

Learn / [Tutorial 2]

Shorter instructions are followed better

Qwen3 0.6BrawP03
What you will learn
A small model is reached by examples, not explanations

This time the problem turns lowercase English words into uppercase. Given hello, the answer is HELLO. Ask a person and it needs no explanation at all. Which is exactly why people write a long prompt for it.

The easy thing to try first

Most people start by writing something like this.

The usual shape
다음에 주어지는 영어 단어를 모두 대문자로 바꾸어 출력해 주세요.
hello

A person reading it knows what it means. But to a 0.6B model this text is not an instruction, just text. The model carries on with whatever is likely to follow it. So it writes an explanation like 다음에 주어지는 단어는... (the word given next is...), or produces something like 네, 알겠습니다 (yes, understood).

Why a longer instruction is worse

Three things overlap. First, the smaller the model, the weaker its instruction following. Second, a long text contains several paths the model could follow. Third, the models in these problems learned from far more English than Korean. Writing at length in Korean blurs it that much more.

If a long instruction does not work, writing a longer one is almost no help. Change direction.

Turning it into examples

Take the method from the previous lesson exactly as it was. Do not explain. Show examples.

Put this in
cat -> CAT
dog -> DOG
hello -> 

Two lines and the rule is visible: left of the arrow is lowercase, right of it is the same word in uppercase. The third line is written only as far as the arrow, so the model fills in HELLO. There is not one character of instruction in it, and it works better than instructing did.

When you pick examples

  • Pick them in the same shape as the real input. If the input is one word, the examples should be one word.

  • Avoid examples that can be answered from memory. If a word you used as an example turns up in a case too, you cannot tell whether the rule was learned.

  • Two or three lines is usually enough. If adding more does not help, length is not the problem.

Try it yourself

Put the sample answer in to pass it, then delete the two example lines and leave only 대문자로 바꿔라. behind, one line meaning make it uppercase. Seeing how differently the same model behaves makes the point of this lesson land at once.

Example answer

Copy this as is into the box with the same name on the problem page. ␣ is a space, ↵ marks a line break. What gets copied is the original text.

pre_prompt
cat␣->␣CAT↵
dog␣->␣DOG↵
post_prompt
␣->␣

Once it passes, change it on purpose. Seeing what breaks it sticks with you longer than knowing the answer.