LLM Sampling Visualizer
Hard-coded next-token logits for one step — tune temperature, top_p, top_k and see how the distribution reshapes before sampling.
Prompt & completion
The weather today is ▌
High temperature → more varied samples over time
Sample history appears here…
Sampling parameters
P(token) = exp(logiti / T) / Σ exp(logitj / T)
Effective next-token distribution
14 / 14 tokens active
In sampling pool
Filtered out (top_k / top_p)