Skip to main content

Simulation run on this device. Scenario values are not sent anywhere — the engine is fully local.

ai simulation · free

Sampling Playground

Shape a mock next-token distribution and sample from it live.

Greedy, top-k, top-p, temperature — what changes first?

Engine 1.0.0 · simulation
0.8

Sharpens or flattens the distribution.

0.9
864

Mock next-token candidates.

Full entropy
3.55 bits
Nucleus entropy
3.06 bits
Nucleus size
14 / 24
Top token mass
35.5%tok#0

Warmer + wide nucleus spreads mass — good for ideation, riskier for facts.

#1 tok0 p=35.5%
#2 tok1 p=16.6%
#3 tok2 p=10.4%
#4 tok3 p=7.4%
#5 tok4 p=5.7%
#6 tok5 p=4.6%
#7 tok6 p=3.8%
#8 tok7 p=3.2%

Method

  • Toy Zipf logits over a vocab slice; temperature rescales logits, top-p keeps a nucleus, then entropy and top-mass update live.
  • Teaching approximation of sampling dynamics — not a real tokenizer or model forward pass.