Sampling bench

Change temperature, top-k and top-p. See the probabilities, then draw from them with a repeatable seed.

How this experiment works

Apply temperature to logits, keep top-k, renormalise, retain the smallest nucleus reaching top-p, then renormalise and sample.

What it does not prove

The scores are invented for teaching. These are not predictions from a real language model.

Try this challenge

Can you keep three candidates while making one dominate? Compare the observed counts with their probabilities.

Related Trust Agent courses · Sources and licences

Enable JavaScript to run the experiment on your device.