Learn LLM with your own model

Choose text

Fruit (Good) Animals (Good) King and queen (Bad) Own text

Citrus fruit: lemon, orange, lime.
The lemon is a citrus fruit. It is sour and yellow, and grows on the tree in the citrus grove.
The orange is a citrus fruit. It is sweet and orange, and grows on the tree in the citrus grove.
The lime is a citrus fruit. It is sour and green, and grows on the tree in the citrus grove.
Berries: strawberry, raspberry, blueberry.
The strawberry is a berry. It is red and sweet, and grows on the bush in the berry patch.
The raspberry is a berry. It is red and tart, and grows on the bush in the berry patch.
The blueberry is a berry. It is blue and sweet, and grows on the bush in the berry patch.
Stone fruit: plum, peach, apricot.
The plum is a stone fruit. It is dark and sweet, and grows on the branches in the orchard.
The peach is a stone fruit. It is pink and juicy, and grows on the branches in the orchard.
The apricot is a stone fruit. It is orange and sweet, and grows on the branches in the orchard.
Tropical fruit: banana, pineapple, mango.
The banana is a tropical fruit. It is yellow and sweet, and grows high up in the palm grove.
The pineapple is a tropical fruit. It is brown and sweet, and grows high up in the palm grove.
The mango is a tropical fruit. It is orange and sweet, and grows high up in the palm grove.

Choosing or editing text rebuilds the model automatically — every word gets a fresh, random vector.

Vocabulary growth

Text with repeating words is good training text, and that's what makes the LLM model good.

Dashed line = how many unique words there would be if every word were new. The real line falls behind as words start repeating.

A flatter curve means more repetition, which helps the model learn reliably, and it's part of why AI companies want more, original text: rare words need repetition too.

Groups:

Show numbers / words

Rounds: 1

Training

Training (gradient descent)

Loss = how wrong the model's guesses are on average. Each "Train" click nudges the vectors to make correct guesses — like the example below — more likely.

Press "Train" to see one real example worked through.

Most similar

Click a word in the model above to see what it's most similar to — and watch this list change as you train.

Ready.

Ask the model using ONLY words from the training text

This model is far too small for human-like answers — expect it to surface related words, not coherent sentences. Try continuing one of these:

Stop threshold (confidence) 3%

Gray words are your question; highlighted words are what the model added. Word choice has some randomness, so answers vary.

Training is random, so quality varies between runs. If the answer looks weak, try "Reset model" and train again.