Sampling Studio — Peek LLM

Prompt

Start here. The water sentence is nearly decided; the fox cliché is wide open. Compare their wheels.

Strategy

Presets set the three knobs below; every knob stays editable.

Temperature

1.00

Divides every logit before softmax. Low sharpens to a point; high flattens toward a coin toss.

Top-k

all

Keep only the k highest-scoring tokens. k = 1 is greedy decoding.

Top-p · nucleus

1.00

Keep the smallest set of tokens whose probability reaches p.

Throw the dart

Each press spins the wheel once with the current knobs. Auto keeps throwing so you can watch the variance.

Computed from GPT-2 117M

the wheel we spin · kept tokens only

Sampled

This distribution

–top candidate

–entropy · bits

–tokens in pool

–tail bucket

Selected candidate

Hover or tap a bar — each bar is one candidate for the next token, with its real GPT-2 score.

How the knobs work

Temperature. Logits are divided by T before softmax — T < 1 sharpens the leader, T > 1 lifts the whole field.
Top-k. Everything below rank k is deleted before the wheel is built.
Top-p. The wheel keeps only the shortest head of the ranking that covers p of the mass.

p_{i} = \frac{e^{z_{i} / T}}{\sum_{j} e^{z_{j} / T}}

Data & sources

Real next-token logits from GPT-2 117M, computed by a verified numpy forward pass on four handcrafted prompts: up to 16,384 top candidates per prompt plus the remaining tail as a 64-bin logit histogram. The three continuations are precomputed greedy and seeded samples. Runs fully offline.

The Curious Case of Neural Text Degeneration — Holtzman et al. (top-p)
How to generate — Hugging Face
openai/gpt-2 — model and tokenizer