Home▸AI & Machine Learning▸Token Sampling Lab 2D: Temperature, Top-k & Nucleus Sampling

Token Sampling Lab 2D: Temperature, Top-k & Nucleus (Top-p)

Interactive 2D simulator of how a language model picks its next word: a live sorted probability bar chart, a draggable cumulative-nucleus curve and a temperature-response curve, all computed from the same real softmax + top-k + top-p math.

AI & Machine Learning2DModerate60 FPS📱 Mobile-adapted⇄ 3D version

Details

CategoryAI & Machine Learning

RenderingCanvas 2D

DifficultyModerate

Target FPS60

ControlsDrag the temperature, top-k, top-p (nucleus) and peak-sharpness sliders to reshape the live token-probability distribution across the bar chart, cumulative-nucleus curve and temperature-response curve; drag directly on the cumulative curve or the temperature-response curve to scrub top-p or temperature; press Sample next token to draw a real weighted-random token and append it to the generated sentence, and New context to reseed the vocabulary logits.

LLMsamplingtemperaturetop-pdecodingNLP

Related simulations

Grammar-Constrained Decoding: Masking an LLM's Token Distribution — 2DStep through a masked-softmax token distribution one sample at a time as a 2D probability…

Lost in the Middle: Long-Context Recall Curve (2D)Drag a needle fact to any depth inside a long context window and run Monte Carlo recall…

Grammar-Constrained Decoding: Masking an LLM's Token DistributionWatch a language model's per-step token probability distribution get masked down to only… ⊞ All 1161 in AI & Machine Learning →

Every time a large language model writes a word, it first computes a probability for every token in its vocabulary, then samples one. This 2D simulator lays that process out as three linked panels — a sorted probability bar chart, a draggable cumulative-nucleus curve, and a temperature-response curve — all driven by the same real softmax, top-k and top-p math a 3D bar-ring version of this lab uses. Drag the temperature, top-k, top-p and peak-sharpness sliders (or drag directly on the cumulative curve) to reshape the distribution, then press "Sample next token" to perform a real weighted random draw and append the chosen word to a growing sentence.

⚙ Under the hood

Watch a language model's next-token probability distribution reshape across three linked 2D panels — a sorted bar chart, a draggable cumulative-nucleus curve and a temperature-response curve — as you drag temperature, top-k, top-p and peak-sharpness sliders, then sample a real weighted-random token from it.

LLMsamplingtemperaturetop-pdecodingNLP

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install