Free local tool
Temperature Intuition Simulator
Visualize how temperature and top-p sampling change a toy output distribution, with an honest explanation that the model is illustrative.
Key facts
Cite factsFact
Toy model only
The simulator uses a small synthetic token distribution. It teaches intuition, not Claude internals.
Fact
Temperature changes sharpness
Lower temperature concentrates probability on high-score options; higher temperature spreads probability across more options.
Fact
Top-p trims the tail
The top-p control keeps the smallest set of options whose cumulative probability reaches the selected threshold.
Interactive visualizer
Sampling intuition, not model introspection
The options below are a toy distribution for the next word in a sentence. Move the controls to see how the candidate set and observed samples change.
Interpretation
Probability after controls
Observed sample counts
How to Use This Tool
- Adjust temperature. Move the temperature slider to see probability mass concentrate or spread across the toy options.
- Adjust top-p. Move the top-p slider to watch lower-probability options enter or leave the candidate set.
- Regenerate samples. Generate a sample batch and compare the observed counts with the probability bars.
Tool FAQ
Is this a Claude model simulator?
No. It is a small client-side teaching model. It does not reveal or approximate Claude token probabilities.
Why do sample counts differ from the bars?
The bars show expected probabilities. The sample chips show one random draw, so small batches naturally vary.
Should I tune prompts by temperature alone?
No. Sampling settings are one variable. A reproducible experiment should also control model, prompt, fixture, tools, output format, and scoring.
Internal links