How do LLMs calculate users’ input? They use tokens. A word, a sentence, or even punctuation are within it. For example:
I love you. (4 tokens. I/love/you/.)
This is unbelievable. (5 tokens. This/is/un/believable/.)
We can use natural language, keep things more precise and clear, and avoid unnecessary formatting.
LLM “reads” tokens only, which means, they don’t read words or phrases. It depends on people’s preferences. However, what you feed them affects directly how they predict what they should respond next. LLMs do not write, they predict. So if brackets and symbols are in between your words, it’ll make LLMs confused.
The less tokens we use, the better? It depends. But the less you give, the more room it has for LLM to remember, and to use it in chat/RPs.
Please don’t use W++ structures like this:
[{Character(“Mitsu Sato”)
{Age(“25” + “Twenty-Five”)
Gender(“Female”)
Apperance(“Long-Black Wolf Cut” + “Bow-Shaped Lips” + “Defined Jaw” + “Dead Gray Eyes” + “Slightly bulk biceps” + “Plump Glutes” + “Slightly toned-abs”)}]
This doesn’t make your bot smarter. Here are some reason.