Book a free call
TrainingProductsTeamSupportContactLabBlog Book a free call

Lab build · October 4, 2026

Token Lab

A real language model, trained by SquidTrain on public-domain books and running entirely in your browser. Type a sentence and watch how it picks the next word: what it reads, where it looks, and why the same prompt can end differently every time.

1 · Your text Edit it or pick a starter

2 · What the model reads

Context window

Solid pieces are your text, dashed ones the model wrote. It never sees letters or words, only these numbered pieces (hover one for its number). Click any piece to see where the model looked when it was there.

5 · Where it looked Attention

Layer
Head

Each layer has several heads, and each head learned on its own to track something different. Some follow the previous word, some find the start of the sentence, some link names and quotes.

3 · What comes next

4 · How it chooses

0 always takes the top choice and tends to repeat itself. Higher values flatten the odds and let unlikely words through.
Only the k most likely pieces are allowed.
Keep the smallest set of pieces whose odds add up to p.

6 · Watch it learn

Error on unseen text:

Real snapshots saved while we trained it. At the start it knows nothing; within a few hundred steps it has words; later it has grammar, names and style. Lower error means better guesses about text it never saw in training.

The books it read

Built with Claude Opus 5.5 by SquidTrain. AI training for companies, teams and individuals. See our training

Loading the model…

How to use Token Lab

  1. Pick a starter or type your own sentence in panel 1. Panel 2 shows how the model splits it into numbered tokens. It never sees letters or words.
  2. Press Next token to watch one pick at a time. Panel 3 shows the odds it gave each option before it chose. Press Write 40 tokens to let it run.
  3. Move the Temperature slider in panel 4 and run the same sentence again. At zero it always takes the top pick and soon repeats itself. Higher values let unlikely words through, which is why the same question can get a different answer each time.
  4. Click any token to see in panel 5 which earlier tokens the model looked at from there. Switch layers and heads to compare what each one learned to track.
  5. Drag the snapshot slider in panel 6 back to step 0. That is the same model before it read anything. Drag forward to watch words, then grammar, appear. Use this snapshot above loads it into the panels so you can try it yourself.
  6. Switch model size in panel 7 to compare three models trained on the same books.

About this build

Built with Claude Opus 5.5 on October 4, 2026. We trained the model ourselves on 144 public-domain books from Project Gutenberg, using one office PC with an NVIDIA RTX 3090, then wrote an engine that runs it in your browser. Nothing you type here is sent anywhere. Read how we built it.

It is small on purpose, so it is fluent and often wrong. That is the point: every model you use at work picks words the same way, one likely piece at a time. Knowing that changes how you check its answers, and it is part of every training engagement we run, for companies, teams and individuals, at work and in daily life. More builds in the Lab.

Sunlight shining down through deep blue water

Find light in the deep waters of AI

Let's talk! A short call will help you learn more about our training and consulting solutions. We can also discuss whether our existing licensed tools can help, or whether you need a custom build.

Book a free call